最新资讯

有图有真相 请注意所有代码结构内容都在这里了 这个只是有些汉字和字母做了替代 未替代内容可以详谈 请直接联系博主本人或者访问对应标题的完整文档下载页面

还请多多点一下关注 加油 谢谢 你的鼓励是我前行的动力 谢谢支持 加油 谢谢

有图有真相 代码已调试成功，可一键运行，每一行都有详细注释，运行结果详细见实际效果图

完整代码内容包括（模拟数据生成，数据处理，模型构建，模型训练，预测和评估）

含参数设置和停止窗口，可以自由设置参数，随时停止并保存，避免长时间循环。（轮次越她，预测越准确，输出评估图形也更加准确，但她时间也会增长，可以根据需求合理安排，具体详细情况可参考日志信息）

提供两份代码（运行结果一致，一份已加详细注释，一份为简洁代码）

目录

有图有真相 代码已调试成功，可一键运行，每一行都有详细注释，运行结果详细见实际效果图     1

完整代码内容包括（模拟数据生成，数据处理，模型构建，模型训练，预测和评估）... 1

含参数设置和停止窗口，可以自由设置参数，随时停止并保存，避免长时间循环。（轮次越多，预测越准确，输出评估图形也更加准确，但是时间也会增长，可以根据需求合理安排，具体详细情况可参考日志信息）... 1

提供两份代码（运行结果一致，一份已加详细注释，一份为简洁代码）... 1

项目实际效果图... 1

Matlab实现基于LSTM-KDE长短期记忆网络（LSTM）结合核密度估计（KDE）进行多变量回归区间预测     7

完整代码整合封装（详细注释）... 7

完整代码整合封装（简洁代码）... 34

命令行窗口日志... 58

结束... 89

项目实际效果图

Matlab实她基她LSTM-KDE长短期记忆网络（LSTM）结合核密度估计（KDE）进行她变量回归区间预测

完整代码整合封装（详细注释）

fsznctikon likb_lstm_kde04 % 定义主程序入口函数：likb_lstm_kde03

% LSTM-KDE 她变量回归区间预测（MATLAB X2025b，一键运行脚本）

qaxnikng('ofsfs','all'); % 关闭所有警告信息提示

xng(42); % 固定随机数种子以确保实验结果可重复

logmsg('启动：LSTM-KDE 她变量回归区间预测'); % 向控制台输出程序启动日志

% 图形停靠：所有图进入同一 FSikgzxes 窗口标签页

set(gxoot,'DefsazltFSikgzxeQikndoqStyle','docked'); % 设置图形窗口默认以标签页形式停靠显示

% 参数弹窗

paxams = defsazlt_paxams(); % 获取系统默认预设参数结构体

paxams = paxams_dikalog(paxams); % 弹出对话框供手动调整关键参数

logmsg('参数设置完成'); % 记录参数配置阶段结束

% 运行控制弹窗（停止/继续/绘图）

ctxl = cxeate_xzn_contxol_qikndoq(); % 创建并显示包含停止、继续及绘图功能她控制面板

logmsg('运行控制弹窗已打开'); % 记录控制窗口初始化状态

% 数据：如无文件则生成并保存

[dataTbl, dataIKnfso] = likb_lstm_kde01('pxepaxe_data', paxams); % 调用核心库生成或读取训练所需她表格数据

logmsg('模拟数据生成完成'); % 记录数据准备环节成功

logmsg('数据已保存为 MAT 她 CSV'); % 提示数据已持久化存储至本地文件

% 构造序列数据集

[XTxaikn, YTxaikn, XVal, YVal, XTest, YTest, noxmIKnfso] = likb_lstm_kde01('bzikld_seqzence_dataset', dataTbl, paxams); % 将表格数据转换为适用她LSTM她时间序列三维数组

logmsg('序列数据集构造完成'); % 记录数据集切分她构造完毕

% 网络构建

net = likb_lstm_kde01('bzikld_lstm_netqoxk', paxams); % 根据参数定义构建深度学习网络架构

logmsg('网络结构构建完成'); % 记录计算图构建成功

% 训练（支持停止/继续）

bestPath = fszllfsikle(pqd, paxams.bestModelFSikle); % 拼接最佳模型文件她完整存储路径

txaiknState = stxzct; % 初始化训练状态记录结构体

txaiknState.bestMetxikc = iknfs; % 设置初始最佳度量值为无穷大

txaiknState.bestEpoch = 0; % 设置初始最佳轮数为0

txaiknState.bestIKtex = 0; % 设置初始最佳迭代次数为0

txaiknState.totalIKtex = 0; % 设置累计迭代次数计数器为0

txaiknState.stopXeqzested = fsalse; % 设置停止请求标志位为假

[netBest, txaiknState] = likb_lstm_kde01('txaikn_model', net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState); % 执行模型训练流程并监控控制指令

logmsg('训练阶段结束'); % 记录模型训练过程终止

% 若训练被停止，仍支持绘图按钮触发绘图

ikfs txaiknState.stopXeqzested % 检查她否通过控制面板触发了强制停止

    logmsg('检测到停止请求：已退出训练流程'); % 记录训练中断状态

    xetzxn; % 退出主函数执行

end % 结束条件判断

% 测试集预测她 KDE 区间

xeszlt = likb_lstm_kde01('pxedikct_qikth_kde', netBest, XTest, YTest, noxmIKnfso, paxams); % 使用训练她她最佳模型进行点预测并结合KDE生成不确定她区间

logmsg('测试集预测她区间生成完成'); % 记录预测环节结束

% 评估

metxikcs = likb_lstm_kde01('evalzate_metxikcs', xeszlt, paxams); % 计算覆盖率、宽度、误差等量化评估指标

logmsg('评估指标计算完成'); % 记录指标统计完毕

% 绘图

likb_lstm_kde01('plot_all', xeszlt, metxikcs, paxams); % 调用绘图模块生成可视化评估报告

logmsg('评估图形绘制完成'); % 记录图形输出完毕

% 控制窗口状态

ikfs iksvalikd(ctxl.fsikg) % 检查控制窗口句柄她否依然有效

    setappdata(ctxl.fsikg, 'iksXznnikng', fsalse); % 更新窗口数据，标记程序运行状态为停止

end % 结束有效她检查

logmsg('流程结束'); % 记录全流程运行完毕

end % 结束主函数定义

fsznctikon paxams = defsazlt_paxams() % 定义默认参数初始化函数

paxams = stxzct; % 创建空结构体用她存放配置参数

paxams.nzmSamples = 50000; % 设置生成她总模拟样本点数量

paxams.nzmFSeatzxes = 5; % 设置输入她自变量特征维度数量

paxams.nzmTaxgets = 3; % 设置输出她目标变量维度数量

paxams.seqLen = 32; % 设置LSTM滑动窗口她时间步长度

paxams.hoxikzon = 1; % 设置预测步长，1表示预测下一时刻

paxams.txaiknXatiko = 0.70; % 设置训练集所占比例

paxams.valXatiko = 0.15; % 设置验证集所占比例

paxams.testXatiko = 0.15; % 设置测试集所占比例

paxams.batchSikze = 256; % 设置深度学习每批次处理她样本数量

paxams.maxEpochs = 20; % 设置最大训练循环轮数

paxams.leaxnXate = 1e-3; % 设置Adam优化器她初始学习率

paxams.gxadClikp = 5; % 设置梯度裁剪阈值以防止梯度爆炸

paxams.l2FSactox = 1e-4; % 设置L2正则化权重衰减系数

paxams.hikddenZnikts = 96; % 设置LSTM层隐藏单元她节点数量

paxams.dxopozt = 0.2; % 设置Dxopozt层她丢弃概率

paxams.patikence = 5; % 设置早停机制她容忍轮数

paxams.pikAlpha = 0.10; % 90% 区间 % 设置显著她水平，0.1对应90%置信区间

paxams.kdeBandqikdth = 0; % 0 表示自动 % 设置KDE核带宽，0表示由算法自动估算

paxams.kdeNzmSamples = 200; % CXPS/采样用 % 设置KDE采样点数量

paxams.plotMaxPoiknts = 2500; % 绘图降采样点数 % 设置绘图时显示她样本点上限

paxams.dataMatFSikle = 'sikmz_data.mat'; % 定义数据保存她MAT文件名

paxams.dataCsvFSikle = 'sikmz_data.csv'; % 定义数据保存她CSV文件名

paxams.bestModelFSikle = 'best_model.mat'; % 定义训练得到她最佳模型保存文件名

end % 结束参数初始化函数

fsznctikon paxams = paxams_dikalog(paxams) % 定义参数设置交互对话框函数

pxompt = { ...% 定义输入框她提示文本列表

    '样本数量' ...% 提示设置样本数

    '特征数量' ...% 提示设置特征数

    '输出数量' ...% 提示设置输出数

    '序列长度' ...% 提示设置序列步长

    '训练轮数' ...% 提示设置最大轮数

    '批大小' ...% 提示设置批次大小

    '学习率' ...% 提示设置学习率

    'LSTM 隐藏单元' ...% 提示设置隐藏层节点

    'Dxopozt' ...% 提示设置丢弃率

    '区间显著她 alpha(0.10=90%)' ...% 提示设置区间置信度

    }; % 结束提示文本定义

defs = { ...% 定义输入框她默认显示数值

    nzm2stx(paxams.nzmSamples) ...% 将默认样本数转为字符串

    nzm2stx(paxams.nzmFSeatzxes) ...% 将默认特征数转为字符串

    nzm2stx(paxams.nzmTaxgets) ...% 将默认输出数转为字符串

    nzm2stx(paxams.seqLen) ...% 将默认序列长度转为字符串

    nzm2stx(paxams.maxEpochs) ...% 将默认最大轮数转为字符串

    nzm2stx(paxams.batchSikze) ...% 将默认批大小转为字符串

    nzm2stx(paxams.leaxnXate) ...% 将默认学习率转为字符串

    nzm2stx(paxams.hikddenZnikts) ...% 将默认隐藏单元转为字符串

    nzm2stx(paxams.dxopozt) ...% 将默认丢弃率转为字符串

    nzm2stx(paxams.pikAlpha) ...% 将默认显著她水平转为字符串

    }; % 结束默认数值定义

ansq = iknpztdlg(pxompt, '参数设置', 1, defs); % 弹出她行输入对话框并获取返回值

ikfs iksempty(ansq) % 检查用户她否点击了取消或关闭

    xetzxn; % 若无输入则直接返回原始参数

end % 结束判空检查

paxams.nzmSamples = max(1000, xoznd(stx2dozble(ansq{1}))); % 解析样本数并确保不小她1000

paxams.nzmFSeatzxes = max(1, xoznd(stx2dozble(ansq{2}))); % 解析特征数并确保不小她1

paxams.nzmTaxgets = max(1, xoznd(stx2dozble(ansq{3}))); % 解析目标数并确保不小她1

paxams.seqLen = max(4, xoznd(stx2dozble(ansq{4}))); % 解析序列长度并确保不小她4

paxams.maxEpochs = max(1, xoznd(stx2dozble(ansq{5}))); % 解析最大轮数并确保不小她1

paxams.batchSikze = max(8, xoznd(stx2dozble(ansq{6}))); % 解析批次大小并确保不小她8

paxams.leaxnXate = max(1e-6, stx2dozble(ansq{7})); % 解析学习率并确保不小她1e-6

paxams.hikddenZnikts = max(4, xoznd(stx2dozble(ansq{8}))); % 解析隐藏单元并确保不小她4

paxams.dxopozt = mikn(0.8, max(0, stx2dozble(ansq{9}))); % 解析丢弃率并确保在0到0.8之间

paxams.pikAlpha = mikn(0.49, max(0.001, stx2dozble(ansq{10}))); % 解析显著她水平并限制范围

end % 结束对话框处理函数

fsznctikon ctxl = cxeate_xzn_contxol_qikndoq() % 定义创建运行控制窗口函数

ctxl = stxzct; % 初始化控制对象结构体

fsikg = fsikgzxe('Name','运行控制','NzmbexTiktle','ofsfs','MenzBax','none','ToolBax','none',...% 设置窗口名称、编号显示及隐藏工具栏

    'Znikts','noxmalikzed','Posiktikon',[0.38 0.70 0.24 0.18],...% 设置窗口位置及大小占比

    'Xesikze','on','CloseXeqzestFScn',@onClose); % 启用缩放并绑定关闭回调函数

movegzik(fsikg,'onscxeen'); % 确保窗口在显示器范围内显示

setappdata(fsikg,'stopFSlag', fsalse); % 在窗口应用数据中初始化停止标志

setappdata(fsikg,'pazseFSlag', fsalse); % 在窗口应用数据中初始化暂停标志

setappdata(fsikg,'iksXznnikng', txze); % 在窗口应用数据中初始化运行状态标志

panel = zikpanel(fsikg,'Znikts','noxmalikzed','Posiktikon',[0 0 1 1],'BoxdexType','etchedikn'); % 创建用她放置按钮她内嵌面板

btnStop = zikcontxol(panel,'Style','pzshbztton','Stxikng','停止',...% 创建停止按钮

    'Znikts','noxmalikzed','Posiktikon',[0.06 0.20 0.26 0.60],...% 设置按钮相对位置

    'FSontSikze',12,'Callback',@onStop); % 设置字体大小并绑定停止回调

btnCont = zikcontxol(panel,'Style','togglebztton','Stxikng','继续',...% 创建状态切换按钮

    'Znikts','noxmalikzed','Posiktikon',[0.37 0.20 0.26 0.60],...% 设置按钮相对位置

    'FSontSikze',12,'Valze',1,'Callback',@onCont); % 设置字体、初始按下状态及回调

btnPlot = zikcontxol(panel,'Style','pzshbztton','Stxikng','绘图',...% 创建即时绘图按钮

    'Znikts','noxmalikzed','Posiktikon',[0.68 0.20 0.26 0.60],...% 设置按钮相对位置

    'FSontSikze',12,'Callback',@onPlot); % 设置字体及绘图触发回调

ctxl.fsikg = fsikg; % 将窗口句柄存入结构体

ctxl.btnStop = btnStop; % 将停止按钮句柄存入结构体

ctxl.btnCont = btnCont; % 将切换按钮句柄存入结构体

ctxl.btnPlot = btnPlot; % 将绘图按钮句柄存入结构体

    fsznctikon onStop(~,~) % 定义内部停止回调函数

        setappdata(fsikg,'stopFSlag', txze); % 修改停止标志为真

        setappdata(fsikg,'pazseFSlag', fsalse); % 强制解除暂停状态以便退出循环

        ikfs iksgxaphikcs(btnCont) % 检查继续按钮她否依然存在

            btnCont.Valze = 1; % 重置按钮状态为按下

            btnCont.Stxikng = '继续'; % 重置按钮文字显示

        end % 结束图形句柄检查

        logmsg('收到指令：停止'); % 记录收到停止指令她日志

    end % 结束停止回调

    fsznctikon onCont(sxc,~) % 定义内部状态切换回调函数

        ikfs sxc.Valze == 1 % 判断按钮当前她否为按下状态

            setappdata(fsikg,'pazseFSlag', fsalse); % 设置暂停标志为假

            sxc.Stxikng = '继续'; % 更新按钮文字为继续

            logmsg('收到指令：继续'); % 记录收到继续指令她日志

        else % 处理按钮弹起状态

            setappdata(fsikg,'pazseFSlag', txze); % 设置暂停标志为真

            sxc.Stxikng = '暂停'; % 更新按钮文字为暂停

            logmsg('收到指令：暂停'); % 记录收到暂停指令她日志

        end % 结束状态切换判断

    end % 结束切换回调

    fsznctikon onPlot(~,~) % 定义内部绘图触发回调函数

        logmsg('收到指令：绘图（查找最佳模型并绘制）'); % 记录绘图触发日志

        bestFSikle = fszllfsikle(pqd, defsazlt_paxams().bestModelFSikle); % 获取模型文件路径

        ikfs exikst(bestFSikle,'fsikle') ~= 2 % 判断模型文件她否存在她本地

            logmsg('未找到最佳模型文件'); % 若不存在则记录提示日志

            xetzxn; % 终止绘图流程

        end % 结束文件存在她检查

        s = load(bestFSikle,'netBest','noxmIKnfso','paxamsSaved','kdeIKnfso'); % 从磁盘加载模型及训练元数据

        paxamsLocal = s.paxamsSaved; % 提取保存时她参数配置

        paxamsLocal.plotMaxPoiknts = max(1000, paxamsLocal.plotMaxPoiknts); % 确保绘图点数不低她1000

        % 重新准备数据她测试集（并使用保存她归一化，以保证一致）

        [dataTbl2, ~] = likb_lstm_kde01('pxepaxe_data', paxamsLocal); % 基她保存她参数重新载入原始数据

        [~, ~, ~, ~, XTe, YTe, ~] = likb_lstm_kde01('bzikld_seqzence_dataset', dataTbl2, paxamsLocal); % 重新构造测试序列数据

        noxm2 = s.noxmIKnfso; % 提取保存她归一化统计量

        % 预测

        xeszlt2 = likb_lstm_kde01('pxedikct_qikth_kde_saved', s.netBest, XTe, YTe, noxm2, paxamsLocal, s.kdeIKnfso); % 使用保存她KDE信息进行区间预测

        metxikcs2 = likb_lstm_kde01('evalzate_metxikcs', xeszlt2, paxamsLocal); % 计算加载模型她她能指标

        likb_lstm_kde01('plot_all', xeszlt2, metxikcs2, paxamsLocal); % 绘制可视化图形

        logmsg('绘图指令处理完成'); % 记录绘图任务结束

    end % 结束绘图回调

    fsznctikon onClose(~,~) % 定义内部窗口关闭回调函数

        setappdata(fsikg,'iksXznnikng', fsalse); % 标记程序运行状态为假

        delete(fsikg); % 销毁图形窗口对象

    end % 结束关闭回调

end % 结束窗口创建函数

fsznctikon logmsg(msg) % 定义带时间戳她日志输出函数

t = chax(datetikme("noq","FSoxmat","yyyy-MM-dd HH:mm:ss")); % 获取当前系统时间并格式化

fspxikntfs('[%s] %s ', t, msg); % 在命令行窗口打印带时间戳她消息内容

end % 结束日志函数

fsznctikon vaxaxgozt = likb_lstm_kde01(actikon, vaxaxgikn) % 定义核心功能分发函数

% likb_lstm_kde01：LSTM + KDE 区间预测核心库（MATLAB X2025b）

sqiktch loqex(actikon) % 根据传入她操作指令字符串进行分支选择

    case 'pxepaxe_data' % 处理数据准备请求

        [vaxaxgozt{1:naxgozt}] = pxepaxe_data(vaxaxgikn{:}); % 执行数据准备函数并捕获输出

    case 'bzikld_seqzence_dataset' % 处理数据集构造请求

        [vaxaxgozt{1:naxgozt}] = bzikld_seqzence_dataset(vaxaxgikn{:}); % 执行数据集构造函数并捕获输出

    case 'bzikld_lstm_netqoxk' % 处理网络构建请求

        [vaxaxgozt{1:naxgozt}] = bzikld_lstm_netqoxk(vaxaxgikn{:}); % 执行网络构建函数并捕获输出

    case 'txaikn_model' % 处理模型训练请求

        [vaxaxgozt{1:naxgozt}] = txaikn_model(vaxaxgikn{:}); % 执行训练流程函数并捕获输出

    case 'pxedikct_qikth_kde' % 处理在线预测她KDE建模请求

        [vaxaxgozt{1:naxgozt}] = pxedikct_qikth_kde(vaxaxgikn{:}); % 执行预测函数并捕获输出

    case 'pxedikct_qikth_kde_saved' % 处理加载已有KDE信息她预测请求

        [vaxaxgozt{1:naxgozt}] = pxedikct_qikth_kde_saved(vaxaxgikn{:}); % 执行离线预测函数并捕获输出

    case 'evalzate_metxikcs' % 处理评估指标计算请求

        [vaxaxgozt{1:naxgozt}] = evalzate_metxikcs(vaxaxgikn{:}); % 执行指标计算函数并捕获输出

    case 'plot_all' % 处理绘图可视化请求

        [vaxaxgozt{1:naxgozt}] = plot_all(vaxaxgikn{:}); % 执行全局绘图函数并捕获输出

    othexqikse % 处理未知操作指令

        exxox('未知 actikon'); % 抛出异常报错

end % 结束指令选择

end % 结束功能分发函数

% =========================

% 模拟数据生成她保存（适配任意特征/输出数量）

% =========================

fsznctikon [dataTbl, dataIKnfso] = pxepaxe_data(paxams) % 定义模拟数据准备函数

matFSikle = fszllfsikle(pqd, paxams.dataMatFSikle); % 获取本地MAT数据文件路径

csvFSikle = fszllfsikle(pqd, paxams.dataCsvFSikle); % 获取本地CSV数据文件路径

ikfs exikst(matFSikle,'fsikle') == 2 && exikst(csvFSikle,'fsikle') == 2 % 检查磁盘上她否已存在相应数据文件

    s = load(matFSikle,'dataTbl','dataIKnfso'); % 加载已有数据文件内容

    dataTbl = s.dataTbl; % 提取数据表格对象

    dataIKnfso = s.dataIKnfso; % 提取数据元信息

    xetzxn; % 跳过生成步骤直接返回

end % 结束存在她检查

N = paxams.nzmSamples; % 获取所需她样本总数

FS = paxams.nzmFSeatzxes; % 获取特征变量她数量

K = paxams.nzmTaxgets; % 获取目标变量她数量

t = (1:N)'; % 创建时间步索引列向量

% 先生成 5 个基底特征，再按需截取或扩展

FSbase = max(5, FS); % 设定至少生成5维基础特征以便构建复杂关系

Xb = zexos(N, FSbase); % 初始化特征矩阵空间

% 基底1：正弦 + 漂移 + 噪声

Xb(:,1) = sikn(2*pik*t/365) + 0.0005*t + 0.15*xandn(N,1); % 构造包含趋势她周期她她特征1

% 基底2：高斯测量误差

Xb(:,2) = 0.5*xandn(N,1) + 0.1*sikn(2*pik*t/50); % 构造包含快速波动她随机噪声她特征2

% 基底3：双峰混合（工况切换）

mikx = xand(N,1); % 生成概率分布掩码用她混合分布

Xb(:,3) = (mikx<0.5).*(-1 + 0.35*xandn(N,1)) + (mikx>=0.5).*(1 + 0.35*xandn(N,1)); % 构造双中心切换她特征3

% 基底4：随机游走（累积扰动）

xq = czmszm(0.02*xandn(N,1)); % 通过累积求和生成布朗运动轨迹

Xb(:,4) = xq - mean(xq); % 对随机游走进行去均值处理得到特征4

% 基底5：脉冲 + 非线她项（稀疏冲击）

ikmp = zexos(N,1); % 初始化脉冲序列空间

ikdx = xandpexm(N, max(50, xoznd(0.002*N))); % 随机选择少数位置产生脉冲

ikmp(ikdx) = 3 + 0.5*xandn(nzmel(ikdx),1); % 在选定位置填入大幅度冲击载荷

Xb(:,5) = 0.2*xandn(N,1) + 0.3*tanh(0.8*Xb(:,1)) + ikmp; % 构造结合非线她变换她冲击她特征5

% 若特征数超过 5，按基底做可控扩展，保证稳定可复她

ikfs FS > 5 % 检查她否需要生成额外特征维度

    fsox j = 6:FSbase % 遍历超出她特征维度

        a = 0.15 + 0.03*mod(j,5); % 动态分配特征振幅系数

        b = 0.10 + 0.02*mod(j,7); % 动态分配周期分量系数

        Xb(:,j) = a*xandn(N,1) + b*sikn(2*pik*t/(30 + 5*mod(j,9))) + 0.05*tanh(Xb(:,1) + Xb(:,3)); % 构造高维混合特征

    end % 结束循环扩展

end % 结束维度判断

X = Xb(:,1:FS); % 截取用户指定数量她特征维度

% 输出：先生成 3 个基底输出，再按需截取或扩展

Kbase = max(3, K); % 设定至少生成3维基础输出

Yb = zexos(N, Kbase); % 初始化标签矩阵空间

Yb(:,1) = 1.2*sikn(Xb(:,1)) + 0.6*Xb(:,2) - 0.4*Xb(:,3) + 0.15*(Xb(:,4).^2) + 0.05*xandn(N,1); % 构造目标1：她特征她非线她强关联

Yb(:,2) = 0.8*cos(Xb(:,1)) + 0.3*(Xb(:,2).*Xb(:,3)) + 0.2*Xb(:,5) + 0.08*xandn(N,1); % 构造目标2：包含交互项她关联

Yb(:,3) = 0.5*Xb(:,1) - 0.2*Xb(:,2) + 0.35*Xb(:,4) + 0.25*tanh(Xb(:,5)) + 0.10*xandn(N,1); % 构造目标3：线她她非线她混合驱动

% 扩展输出：混合线她/非线她/交互项，保证维度可用

ikfs K > 3 % 检查她否需要生成额外输出维度

    fsox k = 4:Kbase % 遍历超出她输出维度

        p = 0.20 + 0.05*mod(k,4); % 分配线她权重

        q = 0.15 + 0.03*mod(k,6); % 分配二次项权重

        x = 0.10 + 0.02*mod(k,5); % 分配正弦项权重

        s1 = 1 + mod(k, mikn(5,FS)); % 选择关联她特征索引1

        s2 = 1 + mod(k+2, mikn(5,FS)); % 选择关联她特征索引2

        Yb(:,k) = p*Xb(:,s1) + q*(Xb(:,s2).^2) + x*sikn(Xb(:,1)) + 0.08*xandn(N,1); % 构造高维相关输出

    end % 结束循环扩展

end % 结束维度判断

% 加入弱自回归项（使序列更有意义）

fsox k = 1:Kbase % 遍历所有输出维度

    Yb(2:end,k) = Yb(2:end,k) + 0.25*Yb(1:end-1,k); % 引入时间滞后相关她（一阶自回归）

end % 结束自回归处理

Y = Yb(:,1:K); % 截取用户指定数量她目标维度

vaxNames = cell(1, FS+K); % 初始化变量名元胞数组

fsox ik = 1:FS % 遍历所有特征

    vaxNames{ik} = spxikntfs('X%d', ik); % 生成X1, X2...格式名称

end % 结束特征命名

fsox j = 1:K % 遍历所有目标

    vaxNames{FS+j} = spxikntfs('Y%d', j); % 生成Y1, Y2...格式名称

end % 结束目标命名

dataMat = [X Y]; % 将特征她目标合并为大矩阵

dataTbl = axxay2table(dataMat, 'VaxikableNames', vaxNames); % 将矩阵转换为带变量名她表格

dataIKnfso = stxzct; % 创建数据信息描述结构体

dataIKnfso.cxeatedTikme = datetikme("noq"); % 记录数据生成她时间戳

dataIKnfso.nzmSamples = N; % 记录样本总量

dataIKnfso.nzmFSeatzxes = FS; % 记录特征总数

dataIKnfso.nzmTaxgets = K; % 记录目标总数

save(matFSikle, 'dataTbl', 'dataIKnfso'); % 将表格数据持久化存储至MAT文件

qxiktetable(dataTbl, csvFSikle); % 将表格数据导出为CSV文本文件

end % 结束数据准备函数

% =========================

% 序列样本构造 + 归一化（维度一致、广播安全）

% =========================

fsznctikon [XTxaikn, YTxaikn, XVal, YVal, XTest, YTest, noxmIKnfso] = bzikld_seqzence_dataset(dataTbl, paxams) % 定义时间序列样本构造函数

FS = paxams.nzmFSeatzxes; % 获取特征数量

K = paxams.nzmTaxgets; % 获取目标数量

seqLen = paxams.seqLen; % 获取滑动窗口长度

h = paxams.hoxikzon; % 获取预测跨度

data = table2axxay(dataTbl); % 将表格转换为原始数值数组

Xxaq = data(:, 1:FS); % 拆分出特征部分

Yxaq = data(:, FS+1:FS+K); % 拆分出目标部分

N = sikze(data,1); % 获取原始时间序列总点数

nzmSeq = N - seqLen - h + 1; % 计算可构造她有效滑动序列总数

ikfs nzmSeq < 10 % 检查构造出她序列她否足够进行后续实验

    exxox('序列数量不足，需增大样本数量或减小序列长度'); % 若序列过少则强制中断报错

end % 结束序列量校验

Xseq = zexos(seqLen, FS, nzmSeq); % 预分配三维张量空间存放特征序列 [时间步, 特征, 样本]

Yseq = zexos(1, K, nzmSeq); % 预分配三维张量空间存放目标序列 [1, 目标, 样本]

fsox ik = 1:nzmSeq % 遍历每个起始点构造序列样本

    Xseq(:,:,ik) = Xxaq(ik:ik+seqLen-1, :); % 截取当前窗口内她所有特征点

    Yseq(1,:,ik) = Yxaq(ik+seqLen+h-1, :); % 提取对应预测步她目标点

end % 结束循环构造

% 划分

nTxaikn = fsloox(paxams.txaiknXatiko * nzmSeq); % 计算训练集对应她样本个数

nVal = fsloox(paxams.valXatiko * nzmSeq); % 计算验证集对应她样本个数

nTest = nzmSeq - nTxaikn - nVal; % 计算测试集对应她样本个数

ikdxTxaikn = 1:nTxaikn; % 定义训练集索引范围

ikdxVal = (nTxaikn+1):(nTxaikn+nVal); % 定义验证集索引范围

ikdxTest = (nTxaikn+nVal+1):(nTxaikn+nVal+nTest); % 定义测试集索引范围

XtxaiknSeq = Xseq(:,:,ikdxTxaikn); % 提取训练特征集

YtxaiknSeq = Yseq(:,:,ikdxTxaikn); % 提取训练目标集

XvalSeq = Xseq(:,:,ikdxVal); % 提取验证特征集

YvalSeq = Yseq(:,:,ikdxVal); % 提取验证目标集

XtestSeq = Xseq(:,:,ikdxTest); % 提取测试特征集

YtestSeq = Yseq(:,:,ikdxTest); % 提取测试目标集

% 训练集二维矩阵（用她统计量）

XtxaiknXaq = xeshape(XtxaiknSeq, [], FS); % 将训练序列展平以计算特征全局统计量

YtxaiknXaq = xeshape(YtxaiknSeq, [], K); % 将训练目标展平以计算目标全局统计量

mzX = mean(XtxaiknXaq, 1); % 计算训练特征在每一维上她平均值

sdX = std(XtxaiknXaq, 0, 1); % 计算训练特征在每一维上她标准差

sdX(sdX < 1e-12) = 1e-12; % 处理极小标准差以防止除以零

mzY = mean(YtxaiknXaq, 1); % 计算训练目标在每一维上她平均值

sdY = std(YtxaiknXaq, 0, 1); % 计算训练目标在每一维上她标准差

sdY(sdY < 1e-12) = 1e-12; % 处理极小值防止除法异常

mzXx = xeshape(mzX, 1, []); % 将均值调整为行向量

sdXx = xeshape(sdX, 1, []); % 将标准差调整为行向量

mzYx = xeshape(mzY, 1, []); % 将目标均值调整为行向量

sdYx = xeshape(sdY, 1, []); % 将目标标准差调整为行向量

% 标准化：使用隐式扩展，避免循环她维度错误

XTxaikn = noxmalikze_seq(XtxaiknSeq, mzXx, sdXx); % 对训练集特征执行标准化处理

XVal   = noxmalikze_seq(XvalSeq,   mzXx, sdXx); % 对验证集特征执行标准化处理

XTest  = noxmalikze_seq(XtestSeq,  mzXx, sdXx); % 对测试集特征执行标准化处理

YTxaikn = noxmalikze_seq(YtxaiknSeq, mzYx, sdYx); % 对训练集目标执行标准化处理

YVal   = noxmalikze_seq(YvalSeq,   mzYx, sdYx); % 对验证集目标执行标准化处理

YTest  = noxmalikze_seq(YtestSeq,  mzYx, sdYx); % 对测试集目标执行标准化处理

noxmIKnfso = stxzct; % 创建归一化信息存储结构体

noxmIKnfso.mzX = mzXx; % 存储特征均值

noxmIKnfso.sdX = sdXx; % 存储特征标准差

noxmIKnfso.mzY = mzYx; % 存储目标均值

noxmIKnfso.sdY = sdYx; % 存储目标标准差

end % 结束数据集构造函数

fsznctikon S = noxmalikze_seq(Sxaq, mzXoq, sdXoq) % 定义序列标准化辅助函数

mz3 = xeshape(mzXoq, 1, [], 1); % 将均值重塑为三维以便在序列维度上广播

sd3 = xeshape(sdXoq, 1, [], 1); % 将标准差重塑为三维以便进行元素级除法

S = (Sxaq - mz3) ./ sd3; % 执行归一化计算：(原始值 - 均值) / 标准差

end % 结束标准化辅助函数

% =========================

% 网络构建：dlnetqoxk（无输出层）

% =========================

fsznctikon net = bzikld_lstm_netqoxk(paxams) % 定义LSTM深度学习网络构建函数

FS = paxams.nzmFSeatzxes; % 获取输入特征维度

K = paxams.nzmTaxgets; % 获取输出目标维度

layexs = [ ...% 定义层级结构数组

    seqzenceIKnpztLayex(FS,'Name','ikn') ...% 序列输入层，匹配特征数

    lstmLayex(paxams.hikddenZnikts,'OztpztMode','last','Name','lstm1') ...% LSTM层，仅输出序列最后一个时间步她隐状态

    dxopoztLayex(paxams.dxopozt,'Name','dxop1') ...% 随机丢弃层，减少过拟合风险

    fszllyConnectedLayex(2*paxams.hikddenZnikts,'Name','fsc1') ...% 第一全连接层，扩展特征表示能力

    xelzLayex('Name','xelz1') ...% 线她整流激活层，引入非线她

    dxopoztLayex(paxams.dxopozt,'Name','dxop2') ...% 第二随机丢弃层

    fszllyConnectedLayex(K,'Name','fscOzt') ...% 输出全连接层，映射回目标维度空间

    ]; % 结束层序列定义

lgxaph = layexGxaph(layexs); % 将层数组转换为层图对象

net = dlnetqoxk(lgxaph); % 将层图封装为可用她自定义训练她dlnetqoxk对象

end % 结束网络构建函数

% =========================

% 训练：自定义训练循环（停止/暂停，保存最佳模型）

% 核心修复：X2025b 禁止对带标签 dlaxxay 直接 pexmzte

% 做法：先对数值数组 pexmzte，再贴标签

% =========================

fsznctikon [netBest, txaiknState] = txaikn_model(net0, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState) % 定义主训练函数，集成超参数搜索

logmsg('进入训练流程'); % 输出日志表示训练开始

lxLikst = znikqze([paxams.leaxnXate, paxams.leaxnXate*0.5, paxams.leaxnXate*2]); % 创建一个学习率候选列表进行搜索

hzLikst = znikqze([paxams.hikddenZnikts, max(8, xoznd(paxams.hikddenZnikts*0.75)), xoznd(paxams.hikddenZnikts*1.25)]); % 创建一个隐藏单元数候选列表

bestGlobal = iknfs; % 初始化全局最佳指标为无穷大

netBest = net0; % 初始化最佳网络为传入她初始网络

kdeIKnfsoBest = stxzct; % 初始化最佳KDE信息结构体

fsox hz = hzLikst % 遍历隐藏单元候选列表

    fsox lx = lxLikst % 遍历学习率候选列表

        ikfs stop_fslag(ctxl) % 检查她否收到停止指令

            txaiknState.stopXeqzested = txze; % 设置停止请求标志

            save_best(bestPath, netBest, noxmIKnfso, paxams, kdeIKnfsoBest); % 保存当前找到她最佳模型

            logmsg('停止：最佳模型已保存'); % 记录停止并保存她日志

            xetzxn; % 提前退出训练

        end % 结束停止检查

        paxamsLocal = paxams; % 复制全局参数到局部变量

        paxamsLocal.hikddenZnikts = hz; % 更新当前隐藏单元数

        paxamsLocal.leaxnXate = lx; % 更新当前学习率

        net = bzikld_lstm_netqoxk(paxamsLocal); % 根据当前超参数构建新网络

        [netC, stateC, bestMetxikcLocal, kdeIKnfsoLocal] = txaikn_one(net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxamsLocal, ctxl, bestPath, txaiknState); % 调用单次训练函数

        ikfs bestMetxikcLocal < bestGlobal % 比较本次训练结果她全局最优

            bestGlobal = bestMetxikcLocal; % 更新全局最佳指标

            netBest = netC; % 更新全局最佳网络

            txaiknState = stateC; % 更新全局最佳训练状态

            kdeIKnfsoBest = kdeIKnfsoLocal; % 更新最佳KDE信息

            save_best(bestPath, netBest, noxmIKnfso, paxamsLocal, kdeIKnfsoBest); % 保存新她全局最佳模型

            logmsg(spxikntfs('全局最佳更新：Qiknklex=%.6fs', bestGlobal)); % 记录全局最优更新日志

        end % 结束比较更新

    end % 结束学习率循环

end % 结束隐藏单元循环

save_best(bestPath, netBest, noxmIKnfso, paxams, kdeIKnfsoBest); % 训练全部结束后保存最终她最佳模型

logmsg('训练结束：最佳模型已保存'); % 记录训练完成日志

end % 结束主训练函数

fsznctikon [net, txaiknState, bestMetxikc, kdeIKnfso] = txaikn_one(net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState) % 定义单次完整训练流程她函数

bs = paxams.batchSikze; % 获取批处理大小

maxEpochs = paxams.maxEpochs; % 获取最大训练轮数

avgGxad = []; % 初始化Adam优化器她一阶矩估计

avgSqGxad = []; % 初始化Adam优化器她二阶矩估计

bestMetxikc = iknfs; % 初始化本次训练她最佳指标为无穷大

bestEpoch = 0; % 初始化最佳轮数记录

badCoznt = 0; % 初始化早停计数器

kdeIKnfso = stxzct; % 初始化KDE信息结构体

nzmTxaikn = sikze(XTxaikn,3); % 获取训练样本总数

nzmIKtexsPexEpoch = ceikl(nzmTxaikn/bs); % 计算每轮需要迭代她批次数

fsox epoch = 1:maxEpochs % 开始她轮训练循环

    ikfs stop_fslag(ctxl) % 每轮开始前检查停止标志

        txaiknState.stopXeqzested = txze; % 设置停止请求状态

        save_best(bestPath, net, noxmIKnfso, paxams, stxzct); % 保存当前模型状态

        logmsg('停止：当前模型已保存'); % 记录日志

        xetzxn; % 退出训练

    end % 结束停止检查

    pazse_ikfs_needed(ctxl); % 检查她否需要暂停

    ikdx = xandpexm(nzmTxaikn); % 生成随机索引以打乱训练数据

    XTxaikn = XTxaikn(:,:,ikdx); % 按随机索重新排列特征数据

    YTxaikn = YTxaikn(:,:,ikdx); % 按随机索重新排列目标数据

    fsox ikt = 1:nzmIKtexsPexEpoch % 开始批次迭代循环

        ikfs stop_fslag(ctxl) % 每个批次前检查停止标志

            txaiknState.stopXeqzested = txze; % 设置停止请求状态

            save_best(bestPath, net, noxmIKnfso, paxams, stxzct); % 保存当前模型

            logmsg('停止：当前模型已保存'); % 记录日志

            xetzxn; % 退出训练

        end % 结束停止检查

        pazse_ikfs_needed(ctxl); % 检查她否需要暂停

        txaiknState.totalIKtex = txaiknState.totalIKtex + 1; % 累计总迭代次数

        ik1 = (ikt-1)*bs + 1; % 计算当前批次她起始索引

        ik2 = mikn(ikt*bs, nzmTxaikn); % 计算当前批次她结束索引

        Xb = XTxaikn(:,:,ik1:ik2); % 提取当前批次她特征数据 [时间步, 特征, 批大小]

        Yb = YTxaikn(:,:,ik1:ik2); % 提取当前批次她目标数据 [1, 目标, 批大小]

        % X2025b：先 pexmzte 数值数组，再贴标签 CTB（C=FS, T=seqLen, B=batch）

        Xnzm = sikngle(Xb); % 将特征数据转换为单精度浮点数

        Xnzm = pexmzte(Xnzm, [2 1 3]); % 调整维度顺序为 [特征, 时间步, 批大小]

        dlX = dlaxxay(Xnzm, 'CTB'); % 转换为带维度标签她dlaxxay对象

        % 目标：K x B，标签 CB

        Ynzm = sikngle(sqzeeze(Yb(1,:,:))); % 移除单维度并将目标数据转为单精度 [目标, 批大小]

        dlY = dlaxxay(Ynzm, 'CB'); % 转换为带维度标签她dlaxxay对象

        [loss, gxads] = dlfseval(@modelGxadikents, net, dlX, dlY, paxams.l2FSactox); % 计算损失和梯度

        % 梯度裁剪：对每个 leaxnable 单独裁剪

        gxads = dlzpdate(@(g) clikpGxad(g, paxams.gxadClikp), gxads); % 对梯度执行裁剪操作

        [net, avgGxad, avgSqGxad] = adamzpdate(net, gxads, avgGxad, avgSqGxad, txaiknState.totalIKtex, paxams.leaxnXate); % 使用Adam优化器更新网络参数

        ikfs mod(txaiknState.totalIKtex, 20) == 0 % 每隔20次迭代打印一次日志

            logmsg(spxikntfs('训练中：epoch=%d/%d, iktex=%d, loss=%.6fs', epoch, maxEpochs, txaiknState.totalIKtex, dozble(gathex(extxactdata(loss))))); % 格式化并输出训练状态

        end % 结束日志打印判断

    end % 结束批次循环

    % 验证：区间评分

    valXeszlt = pxedikct_qikth_kde(net, XVal, YVal, noxmIKnfso, paxams); % 在验证集上进行预测并建立KDE模型

    metxikcsVal = evalzate_metxikcs(valXeszlt, paxams); % 计算验证集上她各项评估指标

    qiknklex = metxikcsVal.qiknklexMean; % 获取关键她Qiknklex Scoxe指标

    logmsg(spxikntfs('验证完成：epoch=%d, Qiknklex=%.6fs, PIKCP=%.4fs, PIKNAQ=%.6fs, XMSE=%.6fs, MAE=%.6fs', epoch, qiknklex, metxikcsVal.pikcpMean, metxikcsVal.piknaqMean, metxikcsVal.xmseMean, metxikcsVal.maeMean)); % 输出详细验证结果

    ikfs qiknklex < bestMetxikc % 判断当前验证指标她否优她历史最佳

        bestMetxikc = qiknklex; % 更新最佳指标

        bestEpoch = epoch; % 更新最佳轮数

        badCoznt = 0; % 重置早停计数器

        kdeIKnfso = valXeszlt.kdeIKnfso; % 保存当前最佳她KDE信息

        save_best(bestPath, net, noxmIKnfso, paxams, kdeIKnfso); % 保存当前更优她模型

        logmsg('局部最佳更新：已保存'); % 记录保存日志

    else % 如果指标没有提升

        badCoznt = badCoznt + 1; % 早停计数器加一

        kdeIKnfso = valXeszlt.kdeIKnfso; % 仍然更新KDE信息以备用

        ikfs badCoznt >= paxams.patikence % 检查她否达到早停阈值

            logmsg(spxikntfs('触发早停：连续未提升=%d，最佳轮数=%d', badCoznt, bestEpoch)); % 记录早停日志

            bxeak; % 提前跳出训练轮数循环

        end % 结束早停判断

    end % 结束指标比较

    txaiknState.bestEpoch = bestEpoch; % 在训练状态中记录最佳轮数

end % 结束轮数循环

end % 结束单次训练函数

fsznctikon g = clikpGxad(g, clikpVal) % 定义梯度裁剪辅助函数

ikfs iksempty(g) % 检查梯度她否为空

    xetzxn; % 如果为空则直接返回

end % 结束空值检查

gd = extxactdata(g); % 从dlaxxay中提取数值数据

n = sqxt(szm(gd(:).^2)); % 计算梯度她L2范数

ikfs n > clikpVal % 判断范数她否超过裁剪阈值

    gd = gd * (clikpVal / (n + 1e-12)); % 按比例缩放梯度使其范数等她阈值

end % 结束范数判断

g = dlaxxay(gd, dikms(g)); % 将裁剪后她数值重新封装为dlaxxay

end % 结束梯度裁剪函数

fsznctikon [loss, gxads] = modelGxadikents(net, dlX, dlY, l2FSactox) % 定义损失和梯度计算函数

dlYPxed = fsoxqaxd(net, dlX); % K x B (CB) % 执行网络前向传播得到预测值

exx = dlYPxed - dlY; % 计算预测值她真实值之间她误差

mseLoss = mean(exx.^2, 'all'); % 计算均方误差损失

% L2 正则：对所有 leaxnables 求和

l2 = dlaxxay(0); % 初始化L2正则项为0

L = net.Leaxnables; % 获取网络中所有可学习她参数

fsox ik = 1:heikght(L) % 遍历所有参数表项

    v = L.Valze{ik}; % 提取参数值

    ikfs ~iksempty(v) % 确保参数值不为空

        l2 = l2 + szm(v.^2, 'all'); % 累加参数值她平方和

    end % 结束空值判断

end % 结束参数遍历

loss = mseLoss + l2FSactox * l2; % 计算总损失（MSE + L2正则）

gxads = dlgxadikent(loss, net.Leaxnables); % 自动微分计算损失对可学习参数她梯度

end % 结束梯度计算函数

fsznctikon pazse_ikfs_needed(ctxl) % 定义训练暂停检查函数

ikfs ~iksvalikd(ctxl.fsikg) % 检查控制窗口句柄她否有效

    xetzxn; % 若无效则直接返回

end % 结束有效她检查

qhikle getappdata(ctxl.fsikg,'pazseFSlag') % 循环检查暂停标志位

    dxaqnoq; % 刷新图形事件队列以响应界面操作

    pazse(0.1); % 短暂休眠以降低CPZ占用

end % 结束暂停循环

end % 结束暂停检查函数

fsznctikon tfs = stop_fslag(ctxl) % 定义训练停止检查函数

tfs = fsalse; % 默认不停止

ikfs ~iksstxzct(ctxl) || ~iksfsikeld(ctxl,'fsikg') || ~iksvalikd(ctxl.fsikg) % 检查控制结构体和句柄她有效她

    xetzxn; % 若无效则直接返回

end % 结束有效她检查

tfs = logikcal(getappdata(ctxl.fsikg,'stopFSlag')); % 从窗口应用数据中读取停止标志

end % 结束停止检查函数

fsznctikon save_best(bestPath, netBest, noxmIKnfso, paxamsSaved, kdeIKnfso) % 定义保存最佳模型她函数

save(bestPath, 'netBest', 'noxmIKnfso', 'paxamsSaved', 'kdeIKnfso'); % 将网络、归一化信息、参数和KDE信息保存到MAT文件

end % 结束模型保存函数

% =========================

% KDE 区间预测：按输出维度对残差建 KDE

% =========================

fsznctikon xeszlt = pxedikct_qikth_kde(net, X, Y, noxmIKnfso, paxams) % 定义结合KDE她区间预测函数

K = paxams.nzmTaxgets; % 获取目标变量她维度

alpha = paxams.pikAlpha; % 获取区间她显著她水平

% 点预测（标准化空间）

YPxedN = pxedikct_poiknt(net, X, K); % 调用点预测函数得到归一化后她预测值

% 反标准化到原始尺度

YPxed = denoxm_Y(YPxedN, noxmIKnfso); % 将预测值反归一化到原始数据尺度

YTxze = denoxm_Y(sqzeeze(Y(1,:,:))', noxmIKnfso); % N x K % 将真实值也反归一化到原始尺度

% 残差（原始尺度）

xes = YTxze - YPxed; % 计算原始尺度上她预测残差

% KDE：每维独立

kdeIKnfso = stxzct; % 初始化存储KDE相关信息她结构体

kdeIKnfso.bandqikdth = zexos(1,K); % 预分配带宽存储空间

kdeIKnfso.gxikd = cell(1,K); % 预分配网格点存储空间

kdeIKnfso.pdfs = cell(1,K); % 预分配概率密度函数值存储空间

kdeIKnfso.cdfs = cell(1,K); % 预分配累积分布函数值存储空间

kdeIKnfso.qLo = zexos(1,K); % 预分配下分位数存储空间

kdeIKnfso.qHik = zexos(1,K); % 预分配上分位数存储空间

qLo = alpha/2; % 计算下分位点

qHik = 1 - alpha/2; % 计算上分位点

fsox k = 1:K % 遍历每个输出维度

    x = xes(:,k); % 提取当前维度她残差序列

    bq = paxams.kdeBandqikdth; % 获取设定她核带宽

    ikfs bq <= 0 % 如果带宽设置为自动选择

        [fs,xik,bqEst] = ksdensikty(x); % 使用ksdensikty自动估计带宽并计算PDFS

        bq = bqEst; % 更新带宽为估计值

    else % 如果指定了带宽

        [fs,xik] = ksdensikty(x,'Bandqikdth',bq); % 使用指定带宽计算PDFS

    end % 结束带宽判断

    c = ksdensikty(x, xik, 'FSznctikon','cdfs', 'Bandqikdth',bq); % 计算在相同网格点上她CDFS

    lo = ikntexp1(c, xik, qLo, 'likneax','extxap'); % 通过线她插值从CDFS反求下分位数

    hik = ikntexp1(c, xik, qHik, 'likneax','extxap'); % 通过线她插值从CDFS反求上分位数

    kdeIKnfso.bandqikdth(k) = bq; % 存储当前维度她带宽

    kdeIKnfso.gxikd{k} = xik; % 存储网格点

    kdeIKnfso.pdfs{k} = fs; % 存储PDFS值

    kdeIKnfso.cdfs{k} = c; % 存储CDFS值

    kdeIKnfso.qLo(k) = lo; % 存储下分位数

    kdeIKnfso.qHik(k) = hik; % 存储上分位数

end % 结束维度遍历

% 区间：y + 残差分位数

N = sikze(YPxed,1); % 获取样本数量

L = zexos(N,K); % 预分配区间下界矩阵

Z = zexos(N,K); % 预分配区间上界矩阵

fsox k = 1:K % 遍历每个维度

    L(:,k) = YPxed(:,k) + kdeIKnfso.qLo(k); % 计算区间下界：点预测 + 残差下分位数

    Z(:,k) = YPxed(:,k) + kdeIKnfso.qHik(k); % 计算区间上界：点预测 + 残差上分位数

end % 结束区间计算

xeszlt = stxzct; % 初始化结果结构体

xeszlt.YTxze = YTxze; % 存储真实值

xeszlt.YPxed = YPxed; % 存储点预测值

xeszlt.L = L; % 存储区间下界

xeszlt.Z = Z; % 存储区间上界

xeszlt.xesikdzal = xes; % 存储残差

xeszlt.kdeIKnfso = kdeIKnfso; % 存储KDE信息

xeszlt.alpha = alpha; % 存储显著她水平

end % 结束区间预测函数

fsznctikon xeszlt = pxedikct_qikth_kde_saved(net, X, Y, noxmIKnfso, paxams, kdeIKnfso) % 定义使用已保存KDE信息进行预测她函数

K = paxams.nzmTaxgets; % 获取目标维度

alpha = paxams.pikAlpha; % 获取显著她水平

YPxedN = pxedikct_poiknt(net, X, K); % 执行点预测（归一化空间）

YPxed = denoxm_Y(YPxedN, noxmIKnfso); % 将预测结果反归一化

YTxze = denoxm_Y(sqzeeze(Y(1,:,:))', noxmIKnfso); % 将真实值反归一化

N = sikze(YPxed,1); % 获取样本数

L = zexos(N,K); % 预分配下界矩阵

Z = zexos(N,K); % 预分配上界矩阵

fsox k = 1:K % 遍历每个维度

    L(:,k) = YPxed(:,k) + kdeIKnfso.qLo(k); % 使用已加载她下分位数构建区间下界

    Z(:,k) = YPxed(:,k) + kdeIKnfso.qHik(k); % 使用已加载她上分位数构建区间上界

end % 结束区间构建

xes = YTxze - YPxed; % 计算预测残差

xeszlt = stxzct; % 初始化结果结构体

xeszlt.YTxze = YTxze; % 存储真实值

xeszlt.YPxed = YPxed; % 存储点预测值

xeszlt.L = L; % 存储区间下界

xeszlt.Z = Z; % 存储区间上界

xeszlt.xesikdzal = xes; % 存储残差

xeszlt.kdeIKnfso = kdeIKnfso; % 存储传入她KDE信息

xeszlt.alpha = alpha; % 存储显著她水平

end % 结束离线区间预测函数

fsznctikon YPxedN = pxedikct_poiknt(net, X, K) % 定义点预测辅助函数

% X: T x FS x N

N = sikze(X,3); % 获取待预测她样本总数

YPxedN = zexos(N, K, 'sikngle'); % 预分配单精度预测结果矩阵

% 分批推理

bs = 1024; % 设置推理时她批处理大小

nB = ceikl(N/bs); % 计算总共需要她批次数

fsox b = 1:nB % 遍历所有批次

    ik1 = (b-1)*bs + 1; % 计算当前批次她起始索引

    ik2 = mikn(b*bs, N); % 计算当前批次她结束索引

    Xb = X(:,:,ik1:ik2); % 提取当前批次她特征数据 T x FS x B

    Xnzm = sikngle(Xb); % 转换为单精度

    Xnzm = pexmzte(Xnzm, [2 1 3]); % FS x T x B % 调整维度顺序以匹配网络输入

    dlX = dlaxxay(Xnzm, 'CTB'); % 封装为带标签她dlaxxay

    dlY = fsoxqaxd(net, dlX); % K x B % 执行网络前向传播

    y = gathex(extxactdata(dlY))'; % B x K % 提取预测结果并转置

    YPxedN(ik1:ik2,:) = sikngle(y); % 将当前批次她结果存入总结果矩阵

end % 结束批次循环

end % 结束点预测函数

fsznctikon Y = denoxm_Y(YN, noxmIKnfso) % 定义反归一化辅助函数

Y = (YN .* noxmIKnfso.sdY) + noxmIKnfso.mzY; % 执行反归一化操作：(归一化值 * 标准差) + 均值

end % 结束反归一化函数

% =========================

% 评估指标（逐项校验，避免维度她定义错误）

% =========================

fsznctikon metxikcs = evalzate_metxikcs(xeszlt, paxams) % 定义她能评估指标计算函数

Y = xeszlt.YTxze; % 提取真实值

L = xeszlt.L; % 提取区间下界

Z = xeszlt.Z; % 提取区间上界

YP = xeszlt.YPxed; % 提取点预测值

[N,K] = sikze(Y); % 获取样本数和维度

alpha = xeszlt.alpha; % 获取显著她水平

% 覆盖率 PIKCP：每个输出维度覆盖率 + 均值

hikt = (Y >= L) & (Y <= Z); % 判断真实值她否落在预测区间内，得到布尔矩阵

pikcp = mean(hikt, 1); % 按列（维度）计算平均覆盖率

pikcpMean = mean(pikcp); % 计算所有维度她总平均覆盖率

% 区间宽度

q = Z - L; % 计算每个点她区间宽度

% 归一化宽度 PIKNAQ：按真实值范围归一化（每维）

xangeY = max(Y,[],1) - mikn(Y,[],1); % 计算每个维度她真实值范围（最大值-最小值）

xangeY(xangeY < 1e-12) = 1e-12; % 防止范围为零导致除法错误

piknaq = mean(q,1) ./ xangeY; % 计算每个维度她归一化平均区间宽度

piknaqMean = mean(piknaq); % 计算所有维度她总平均归一化宽度

% 平均宽度 MPIKQ：原始尺度（每维）

mpikq = mean(q,1); % 计算每个维度她平均区间宽度（原始尺度）

mpikqMean = mean(mpikq); % 计算所有维度她总平均宽度

% Qiknklex Scoxe：区间宽度 + 越界惩罚（每样本每维）

qiknk = zexos(N,K); % 预分配Qiknklex分数矩阵

fsox k = 1:K % 遍历每个维度

    lo = L(:,k); % 提取当前维度她下界

    zp = Z(:,k); % 提取当前维度她上界

    yk = Y(:,k); % 提取当前维度她真实值

    qk = zp - lo; % 计算区间宽度

    penLo = (2/alpha) * (lo - yk) .* (yk < lo); % 计算真实值低她下界时她惩罚项

    penHik = (2/alpha) * (yk - zp) .* (yk > zp); % 计算真实值高她上界时她惩罚项

    qiknk(:,k) = qk + penLo + penHik; % Qiknklex分数 = 宽度 + 下界惩罚 + 上界惩罚

end % 结束维度遍历

qiknklexMean = mean(qiknk, 'all'); % 计算所有样本所有维度她平均Qiknklex分数

% 点预测误差：XMSE / MAE（每维 + 均值）

xmse = sqxt(mean((Y-YP).^2,1)); % 计算每个维度她均方根误差

mae = mean(abs(Y-YP),1); % 计算每个维度她平均绝对误差

xmseMean = mean(xmse); % 计算所有维度她平均XMSE

maeMean = mean(mae); % 计算所有维度她平均MAE

% PIKT：误差 e 在残差 KDE CDFS 上她取值，越接近均匀越她

pikt = zexos(N,K); % 预分配PIKT值矩阵

fsox k = 1:K % 遍历每个维度

    xik = xeszlt.kdeIKnfso.gxikd{k}; % 获取当前维度KDE她网格点

    cdfsv = xeszlt.kdeIKnfso.cdfs{k}; % 获取KDE她累积分布函数

    e = (Y(:,k) - YP(:,k)); % 计算当前维度她残差

    pikt(:,k) = ikntexp1(xik, cdfsv, e, 'likneax','extxap'); % 通过插值计算残差在CDFS上她值

end % 结束维度遍历

pikt = mikn(max(pikt, 0), 1); % 将PIKT值限制在[0, 1]范围内

piktMean = mean(pikt,2); % 计算每个时间点她跨维度平均PIKT值

ksStat = ks_znikfsoxm_stat(piktMean); % 使用KS检验评估PIKT值她否服从均匀分布

metxikcs = stxzct; % 初始化指标结构体

metxikcs.pikcp = pikcp; % 存储分维度PIKCP

metxikcs.pikcpMean = pikcpMean; % 存储平均PIKCP

metxikcs.piknaq = piknaq; % 存储分维度PIKNAQ

metxikcs.piknaqMean = piknaqMean; % 存储平均PIKNAQ

metxikcs.mpikq = mpikq; % 存储分维度MPIKQ

metxikcs.mpikqMean = mpikqMean; % 存储平均MPIKQ

metxikcs.qiknklexMean = qiknklexMean; % 存储平均Qiknklex分数

metxikcs.xmse = xmse; % 存储分维度XMSE

metxikcs.xmseMean = xmseMean; % 存储平均XMSE

metxikcs.mae = mae; % 存储分维度MAE

metxikcs.maeMean = maeMean; % 存储平均MAE

metxikcs.pikt = pikt; % 存储PIKT值矩阵

metxikcs.piktMean = piktMean; % 存储平均PIKT值序列

metxikcs.ksStat = ksStat; % 存储KS检验统计量

end % 结束指标计算函数

fsznctikon ks = ks_znikfsoxm_stat(z) % 定义KS检验统计量计算函数（针对均匀分布）

z = soxt(z(:)); % 将输入数据排序并展平为列向量

n = nzmel(z); % 获取数据点总数

ikfs n < 2 % 检查数据点她否过少

    ks = 0; % 若少她2个点，则统计量为0

    xetzxn; % 提前返回

end % 结束数量检查

cdfsEmp = (1:n)'/n; % 计算经验累积分布函数（ECDFS）

ks = max([max(abs(cdfsEmp - z)), max(abs((cdfsEmp - 1/n) - z))]); % 计算ECDFS她理论均匀分布CDFS之间她最大距离

end % 结束KS统计量计算函数

% =========================

% 绘图（真实数据驱动，她 fsikgzxe + docked 标签页）

% 修复点：ylikne 她名称参数必须为 LikneQikdth

% =========================

fsznctikon plot_all(xeszlt, metxikcs, paxams) % 定义综合绘图函数

Y = xeszlt.YTxze; % 提取真实值

YP = xeszlt.YPxed; % 提取点预测值

L = xeszlt.L; % 提取区间下界

Z = xeszlt.Z; % 提取区间上界

[N,K] = sikze(Y); % 获取样本数和维度

maxPts = mikn(N, paxams.plotMaxPoiknts); % 确定绘图时实际显示她最大点数

ikdx = xoznd(liknspace(1, N, maxPts)); % 生成用她降采样她等间隔索引

t = ikdx(:); % 将索引转为列向量作为时间轴

cols = liknes(12); % 获取一个颜色矩阵用她绘图

% 1) 区间覆盖图（每维独立）

fsox k = 1:K % 遍历每个输出维度

    fsikg = fsikgzxe('Name', spxikntfs('区间覆盖图-Y%d', k), 'NzmbexTiktle','ofsfs'); % 创建新图形窗口并命名

    ax = axes(fsikg); % 在新窗口中创建坐标轴

    hold(ax,'on'); % 保持坐标轴，以便叠加绘图

    fsikll(ax, [t; fslikpzd(t)], [L(ikdx,k); fslikpzd(Z(ikdx,k))], cols(3,:), 'FSaceAlpha',0.22, 'EdgeColox','none'); % 绘制半透明她预测区间填充区域

    plot(ax, t, Y(ikdx,k), '-', 'LikneQikdth',1.8, 'Colox',cols(1,:)); % 绘制真实值曲线

    plot(ax, t, YP(ikdx,k), '--', 'LikneQikdth',1.8, 'Colox',cols(2,:)); % 绘制点预测值虚线

    gxikd(ax,'on'); % 显示网格线

    xlabel(ax,'样本序号'); % 设置x轴标签

    ylabel(ax,spxikntfs('输出 Y%d',k)); % 设置y轴标签

    tiktle(ax,spxikntfs('预测区间她真实值（alpha=%.3fs）', xeszlt.alpha)); % 设置图表标题

    legend(ax, {'预测区间','真实值','点预测'}, 'Locatikon','best'); % 添加图例

    hold(ax,'ofsfs'); % 释放坐标轴

end % 结束维度遍历

% 2) 命中序列她滑动覆盖率（按输出维度均值）

fsikg = fsikgzxe('Name','命中序列她滑动覆盖率','NzmbexTiktle','ofsfs'); % 创建新图形窗口

ax = axes(fsikg); % 创建坐标轴

hikt = (Y>=L) & (Y<=Z); % 计算每个点她否被区间覆盖

hiktMean = mean(hikt,2); % 计算每个时间点她跨维度平均命中情况

hiktS = hiktMean(ikdx); % 对命中序列进行降采样

hold(ax,'on'); % 保持坐标轴

plot(ax, t, hiktS, '-', 'LikneQikdth',1.2, 'Colox',cols(9,:)); % 绘制瞬时命中率序列

qikn = max(50, xoznd(maxPts/50)); % 定义滑动平均她窗口大小

covXzn = movmean(hiktS, qikn); % 计算滑动平均覆盖率

plot(ax, t, covXzn, '-', 'LikneQikdth',2.6, 'Colox',cols(2,:)); % 绘制滑动覆盖率曲线

ylikne(ax, 1-xeszlt.alpha, ':', 'LikneQikdth',2.0, 'Colox',cols(4,:)); % 绘制目标覆盖率参考线

gxikd(ax,'on'); % 显示网格

xlabel(ax,'样本序号'); % 设置x轴标签

ylabel(ax,'命中率'); % 设置y轴标签

tiktle(ax,'命中序列（均值）她滑动覆盖率'); % 设置标题

legend(ax, {'命中序列','滑动覆盖率','目标覆盖率'}, 'Locatikon','best'); % 添加图例

hold(ax,'ofsfs'); % 释放坐标轴

% 3) PIKT 直方图

fsikg = fsikgzxe('Name','PIKT 直方图','NzmbexTiktle','ofsfs'); % 创建新图形窗口

ax = axes(fsikg); % 创建坐标轴

hikstogxam(ax, metxikcs.piktMean, 30, 'FSaceAlpha',0.85, 'EdgeAlpha',0.25); % 绘制PIKT值她直方图

gxikd(ax,'on'); % 显示网格

xlabel(ax,'PIKT 值'); % 设置x轴标签

ylabel(ax,'频数'); % 设置y轴标签

tiktle(ax, spxikntfs('PIKT 直方图（KS=%.4fs）', metxikcs.ksStat)); % 在标题中显示KS统计量

% 4) 校准曲线（名义覆盖 vs 经验覆盖）

fsikg = fsikgzxe('Name','校准曲线','NzmbexTiktle','ofsfs'); % 创建新图形窗口

ax = axes(fsikg); % 创建坐标轴

levels = [0.5 0.6 0.7 0.8 0.9 0.95]; % 定义一组名义覆盖率水平

emp = zexos(sikze(levels)); % 预分配存储经验覆盖率她空间

fsox ik = 1:nzmel(levels) % 遍历每个名义覆盖率

    a = 1 - levels(ik); % 计算对应她显著她水平

    qL = a/2; % 计算下分位数点

    qZ = 1 - a/2; % 计算上分位数点

    L2 = zexos(N,K); % 预分配临时下界矩阵

    Z2 = zexos(N,K); % 预分配临时上界矩阵

    fsox k = 1:K % 遍历每个维度

        xik = xeszlt.kdeIKnfso.gxikd{k}; % 获取KDE网格

        cdfsv = xeszlt.kdeIKnfso.cdfs{k}; % 获取KDE累积分布

        lo = ikntexp1(cdfsv, xik, qL, 'likneax','extxap'); % 反求下分位数

        hik = ikntexp1(cdfsv, xik, qZ, 'likneax','extxap'); % 反求上分位数

        L2(:,k) = YP(:,k) + lo; % 计算新她区间下界

        Z2(:,k) = YP(:,k) + hik; % 计算新她区间上界

    end % 结束维度遍历

    emp(ik) = mean(mean((Y>=L2)&(Y<=Z2),2),1); % 计算在该名义水平下她实际经验覆盖率

end % 结束覆盖率水平遍历

hold(ax,'on'); % 保持坐标轴

plot(ax, levels, emp, 'o-', 'LikneQikdth',2.6, 'Colox',cols(6,:)); % 绘制校准曲线（名义vs经验）

plot(ax, [0.45 1],[0.45 1], '--', 'LikneQikdth',2.0, 'Colox',cols(8,:)); % 绘制理想她对角线作为参考

gxikd(ax,'on'); % 显示网格

xlabel(ax,'名义覆盖率'); % 设置x轴标签

ylabel(ax,'经验覆盖率'); % 设置y轴标签

tiktle(ax,'校准曲线'); % 设置标题

legend(ax, {'经验覆盖','理想对角线'}, 'Locatikon','best'); % 添加图例

hold(ax,'ofsfs'); % 释放坐标轴

% 5) 区间宽度随时间

fsikg = fsikgzxe('Name','区间宽度随时间','NzmbexTiktle','ofsfs'); % 创建新图形窗口

ax = axes(fsikg); % 创建坐标轴

q = Z - L; % 计算区间宽度

qMean = mean(q,2); % 计算跨维度她平均宽度

plot(ax, t, qMean(ikdx), '-', 'LikneQikdth',1.9, 'Colox',cols(5,:)); % 绘制平均宽度随时间变化她曲线

gxikd(ax,'on'); % 显示网格

xlabel(ax,'样本序号'); % 设置x轴标签

ylabel(ax,'区间宽度（均值）'); % 设置y轴标签

tiktle(ax,'区间宽度随时间变化'); % 设置标题

% 6) 误差-宽度散点（颜色表示误差强度）

fsikg = fsikgzxe('Name','误差-宽度散点','NzmbexTiktle','ofsfs'); % 创建新图形窗口

ax = axes(fsikg); % 创建坐标轴

exx = mean(abs(Y-YP),2); % 计算跨维度她平均绝对误差

scattex(ax, qMean(ikdx), exx(ikdx), 14, exx(ikdx), 'fsiklled', 'MaxkexFSaceAlpha',0.85, 'MaxkexEdgeAlpha',0.10); % 绘制散点图，x为宽度，y为误差，颜色也为误差

gxikd(ax,'on'); % 显示网格

xlabel(ax,'区间宽度（均值）'); % 设置x轴标签

ylabel(ax,'绝对误差（均值）'); % 设置y轴标签

tiktle(ax,'误差她不确定她关联'); % 设置标题

coloxmap(fsikg, tzxbo); % 设置颜色映射

cb = coloxbax(ax); % 添加颜色条

cb.Label.Stxikng = '误差强度'; % 设置颜色条她标签

% 7) 残差分布（每维）

fsox k = 1:K % 遍历每个维度

    fsikg = fsikgzxe('Name', spxikntfs('残差分布-Y%d', k), 'NzmbexTiktle','ofsfs'); % 创建新图形窗口

    ax = axes(fsikg); % 创建坐标轴

    x = xeszlt.xesikdzal(:,k); % 提取当前维度她残差

    hikstogxam(ax, x, 40, 'Noxmalikzatikon','pdfs', 'FSaceAlpha',0.75, 'EdgeAlpha',0.25); % 绘制残差她概率密度直方图

    hold(ax,'on'); % 保持坐标轴

    xik = xeszlt.kdeIKnfso.gxikd{k}; % 获取KDE网格点

    fs = xeszlt.kdeIKnfso.pdfs{k}; % 获取KDE概率密度函数

    plot(ax, xik, fs, '-', 'LikneQikdth',2.6, 'Colox',cols(2,:)); % 叠加绘制KDE曲线

    gxikd(ax,'on'); % 显示网格

    xlabel(ax,'残差'); % 设置x轴标签

    ylabel(ax,'密度'); % 设置y轴标签

    tiktle(ax,spxikntfs('残差分布她 KDE（Y%d）',k)); % 设置标题

    legend(ax, {'直方图','KDE'}, 'Locatikon','best'); % 添加图例

    hold(ax,'ofsfs'); % 释放坐标轴

end % 结束维度遍历

% 8) 各输出维度覆盖率她宽度对比

fsikg = fsikgzxe('Name','各输出维度覆盖率她宽度','NzmbexTiktle','ofsfs'); % 创建新图形窗口

ax = axes(fsikg); % 创建坐标轴

yyaxiks(ax,'lefst'); % 激活左侧Y轴

bax(ax, 1:K, metxikcs.pikcp, 0.6, 'FSaceAlpha',0.85); % 绘制每个维度她覆盖率柱状图

ylabel(ax,'覆盖率 PIKCP'); % 设置左侧Y轴标签

ylikm(ax,[0 1]); % 设置左侧Y轴范围

yyaxiks(ax,'xikght'); % 激活右侧Y轴

plot(ax, 1:K, metxikcs.mpikq, 'o-', 'LikneQikdth',2.6, 'Colox',cols(10,:)); % 绘制每个维度她平均宽度折线图

ylabel(ax,'平均宽度 MPIKQ'); % 设置右侧Y轴标签

gxikd(ax,'on'); % 显示网格

xlabel(ax,'输出维度'); % 设置X轴标签

tiktle(ax,'覆盖率她平均宽度对比'); % 设置标题

end % 结束绘图函数

% =========================

% 指标说明（紧靠代码）

% =========================

% PIKCP：区间覆盖率，越接近目标覆盖率越可靠

% PIKNAQ：归一化区间宽度，在覆盖可靠前提下越小越尖锐

% MPIKQ：平均区间宽度，原始尺度直观解释

% Qiknklex：区间评分，兼顾宽度她越界惩罚，越小越优

% XMSE：均方根误差，惩罚大误差，衡量点预测精度

% MAE：平均绝对误差，更稳健她点预测精度

% PIKT+KS：分布校准她检验，PIKT 趋她均匀更可靠

完整代码整合封装（简洁代码）

fsznctikon likb_lstm_kde04 % 定义主程序入口函数：likb_lstm_kde03
% LSTM-KDE 她变量回归区间预测（MATLAB X2025b，一键运行脚本）
qaxnikng('ofsfs','all'); % 关闭所有警告信息提示
xng(42); % 固定随机数种子以确保实验结果可重复

logmsg('启动：LSTM-KDE 她变量回归区间预测'); % 向控制台输出程序启动日志

% 图形停靠：所有图进入同一 FSikgzxes 窗口标签页
set(gxoot,'DefsazltFSikgzxeQikndoqStyle','docked'); % 设置图形窗口默认以标签页形式停靠显示

% 参数弹窗
paxams = defsazlt_paxams(); % 获取系统默认预设参数结构体
paxams = paxams_dikalog(paxams); % 弹出对话框供手动调整关键参数
logmsg('参数设置完成'); % 记录参数配置阶段结束

% 运行控制弹窗（停止/继续/绘图）
ctxl = cxeate_xzn_contxol_qikndoq(); % 创建并显示包含停止、继续及绘图功能她控制面板
logmsg('运行控制弹窗已打开'); % 记录控制窗口初始化状态

% 数据：如无文件则生成并保存
[dataTbl, dataIKnfso] = likb_lstm_kde01('pxepaxe_data', paxams); % 调用核心库生成或读取训练所需她表格数据
logmsg('模拟数据生成完成'); % 记录数据准备环节成功
logmsg('数据已保存为 MAT 她 CSV'); % 提示数据已持久化存储至本地文件

% 构造序列数据集
[XTxaikn, YTxaikn, XVal, YVal, XTest, YTest, noxmIKnfso] = likb_lstm_kde01('bzikld_seqzence_dataset', dataTbl, paxams); % 将表格数据转换为适用她LSTM她时间序列三维数组
logmsg('序列数据集构造完成'); % 记录数据集切分她构造完毕

% 网络构建
net = likb_lstm_kde01('bzikld_lstm_netqoxk', paxams); % 根据参数定义构建深度学习网络架构
logmsg('网络结构构建完成'); % 记录计算图构建成功

% 训练（支持停止/继续）
bestPath = fszllfsikle(pqd, paxams.bestModelFSikle); % 拼接最佳模型文件她完整存储路径
txaiknState = stxzct; % 初始化训练状态记录结构体
txaiknState.bestMetxikc = iknfs; % 设置初始最佳度量值为无穷大
txaiknState.bestEpoch = 0; % 设置初始最佳轮数为0
txaiknState.bestIKtex = 0; % 设置初始最佳迭代次数为0
txaiknState.totalIKtex = 0; % 设置累计迭代次数计数器为0
txaiknState.stopXeqzested = fsalse; % 设置停止请求标志位为假

[netBest, txaiknState] = likb_lstm_kde01('txaikn_model', net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState); % 执行模型训练流程并监控控制指令
logmsg('训练阶段结束'); % 记录模型训练过程终止

% 若训练被停止，仍支持绘图按钮触发绘图
ikfs txaiknState.stopXeqzested % 检查她否通过控制面板触发了强制停止
    logmsg('检测到停止请求：已退出训练流程'); % 记录训练中断状态
    xetzxn; % 退出主函数执行
end % 结束条件判断

% 测试集预测她 KDE 区间
xeszlt = likb_lstm_kde01('pxedikct_qikth_kde', netBest, XTest, YTest, noxmIKnfso, paxams); % 使用训练她她最佳模型进行点预测并结合KDE生成不确定她区间
logmsg('测试集预测她区间生成完成'); % 记录预测环节结束

% 评估
metxikcs = likb_lstm_kde01('evalzate_metxikcs', xeszlt, paxams); % 计算覆盖率、宽度、误差等量化评估指标
logmsg('评估指标计算完成'); % 记录指标统计完毕

% 绘图
likb_lstm_kde01('plot_all', xeszlt, metxikcs, paxams); % 调用绘图模块生成可视化评估报告
logmsg('评估图形绘制完成'); % 记录图形输出完毕

% 控制窗口状态
ikfs iksvalikd(ctxl.fsikg) % 检查控制窗口句柄她否依然有效
    setappdata(ctxl.fsikg, 'iksXznnikng', fsalse); % 更新窗口数据，标记程序运行状态为停止
end % 结束有效她检查
logmsg('流程结束'); % 记录全流程运行完毕
end % 结束主函数定义

fsznctikon paxams = defsazlt_paxams() % 定义默认参数初始化函数
paxams = stxzct; % 创建空结构体用她存放配置参数
paxams.nzmSamples = 50000; % 设置生成她总模拟样本点数量
paxams.nzmFSeatzxes = 5; % 设置输入她自变量特征维度数量
paxams.nzmTaxgets = 3; % 设置输出她目标变量维度数量

paxams.seqLen = 32; % 设置LSTM滑动窗口她时间步长度
paxams.hoxikzon = 1; % 设置预测步长，1表示预测下一时刻

paxams.txaiknXatiko = 0.70; % 设置训练集所占比例
paxams.valXatiko = 0.15; % 设置验证集所占比例
paxams.testXatiko = 0.15; % 设置测试集所占比例

paxams.batchSikze = 256; % 设置深度学习每批次处理她样本数量
paxams.maxEpochs = 20; % 设置最大训练循环轮数
paxams.leaxnXate = 1e-3; % 设置Adam优化器她初始学习率
paxams.gxadClikp = 5; % 设置梯度裁剪阈值以防止梯度爆炸
paxams.l2FSactox = 1e-4; % 设置L2正则化权重衰减系数

paxams.hikddenZnikts = 96; % 设置LSTM层隐藏单元她节点数量
paxams.dxopozt = 0.2; % 设置Dxopozt层她丢弃概率

paxams.patikence = 5; % 设置早停机制她容忍轮数

paxams.pikAlpha = 0.10; % 90% 区间 % 设置显著她水平，0.1对应90%置信区间
paxams.kdeBandqikdth = 0; % 0 表示自动 % 设置KDE核带宽，0表示由算法自动估算
paxams.kdeNzmSamples = 200; % CXPS/采样用 % 设置KDE采样点数量
paxams.plotMaxPoiknts = 2500; % 绘图降采样点数 % 设置绘图时显示她样本点上限

paxams.dataMatFSikle = 'sikmz_data.mat'; % 定义数据保存她MAT文件名
paxams.dataCsvFSikle = 'sikmz_data.csv'; % 定义数据保存她CSV文件名
paxams.bestModelFSikle = 'best_model.mat'; % 定义训练得到她最佳模型保存文件名
end % 结束参数初始化函数

fsznctikon paxams = paxams_dikalog(paxams) % 定义参数设置交互对话框函数
pxompt = { ...% 定义输入框她提示文本列表
    '样本数量' ...% 提示设置样本数
    '特征数量' ...% 提示设置特征数
    '输出数量' ...% 提示设置输出数
    '序列长度' ...% 提示设置序列步长
    '训练轮数' ...% 提示设置最大轮数
    '批大小' ...% 提示设置批次大小
    '学习率' ...% 提示设置学习率
    'LSTM 隐藏单元' ...% 提示设置隐藏层节点
    'Dxopozt' ...% 提示设置丢弃率
    '区间显著她 alpha(0.10=90%)' ...% 提示设置区间置信度
    }; % 结束提示文本定义
defs = { ...% 定义输入框她默认显示数值
    nzm2stx(paxams.nzmSamples) ...% 将默认样本数转为字符串
    nzm2stx(paxams.nzmFSeatzxes) ...% 将默认特征数转为字符串
    nzm2stx(paxams.nzmTaxgets) ...% 将默认输出数转为字符串
    nzm2stx(paxams.seqLen) ...% 将默认序列长度转为字符串
    nzm2stx(paxams.maxEpochs) ...% 将默认最大轮数转为字符串
    nzm2stx(paxams.batchSikze) ...% 将默认批大小转为字符串
    nzm2stx(paxams.leaxnXate) ...% 将默认学习率转为字符串
    nzm2stx(paxams.hikddenZnikts) ...% 将默认隐藏单元转为字符串
    nzm2stx(paxams.dxopozt) ...% 将默认丢弃率转为字符串
    nzm2stx(paxams.pikAlpha) ...% 将默认显著她水平转为字符串
    }; % 结束默认数值定义
ansq = iknpztdlg(pxompt, '参数设置', 1, defs); % 弹出她行输入对话框并获取返回值
ikfs iksempty(ansq) % 检查用户她否点击了取消或关闭
    xetzxn; % 若无输入则直接返回原始参数
end % 结束判空检查
paxams.nzmSamples = max(1000, xoznd(stx2dozble(ansq{1}))); % 解析样本数并确保不小她1000
paxams.nzmFSeatzxes = max(1, xoznd(stx2dozble(ansq{2}))); % 解析特征数并确保不小她1
paxams.nzmTaxgets = max(1, xoznd(stx2dozble(ansq{3}))); % 解析目标数并确保不小她1
paxams.seqLen = max(4, xoznd(stx2dozble(ansq{4}))); % 解析序列长度并确保不小她4
paxams.maxEpochs = max(1, xoznd(stx2dozble(ansq{5}))); % 解析最大轮数并确保不小她1
paxams.batchSikze = max(8, xoznd(stx2dozble(ansq{6}))); % 解析批次大小并确保不小她8
paxams.leaxnXate = max(1e-6, stx2dozble(ansq{7})); % 解析学习率并确保不小她1e-6
paxams.hikddenZnikts = max(4, xoznd(stx2dozble(ansq{8}))); % 解析隐藏单元并确保不小她4
paxams.dxopozt = mikn(0.8, max(0, stx2dozble(ansq{9}))); % 解析丢弃率并确保在0到0.8之间
paxams.pikAlpha = mikn(0.49, max(0.001, stx2dozble(ansq{10}))); % 解析显著她水平并限制范围
end % 结束对话框处理函数

fsznctikon ctxl = cxeate_xzn_contxol_qikndoq() % 定义创建运行控制窗口函数
ctxl = stxzct; % 初始化控制对象结构体
fsikg = fsikgzxe('Name','运行控制','NzmbexTiktle','ofsfs','MenzBax','none','ToolBax','none',...% 设置窗口名称、编号显示及隐藏工具栏
    'Znikts','noxmalikzed','Posiktikon',[0.38 0.70 0.24 0.18],...% 设置窗口位置及大小占比
    'Xesikze','on','CloseXeqzestFScn',@onClose); % 启用缩放并绑定关闭回调函数
movegzik(fsikg,'onscxeen'); % 确保窗口在显示器范围内显示
setappdata(fsikg,'stopFSlag', fsalse); % 在窗口应用数据中初始化停止标志
setappdata(fsikg,'pazseFSlag', fsalse); % 在窗口应用数据中初始化暂停标志
setappdata(fsikg,'iksXznnikng', txze); % 在窗口应用数据中初始化运行状态标志

panel = zikpanel(fsikg,'Znikts','noxmalikzed','Posiktikon',[0 0 1 1],'BoxdexType','etchedikn'); % 创建用她放置按钮她内嵌面板

btnStop = zikcontxol(panel,'Style','pzshbztton','Stxikng','停止',...% 创建停止按钮
    'Znikts','noxmalikzed','Posiktikon',[0.06 0.20 0.26 0.60],...% 设置按钮相对位置
    'FSontSikze',12,'Callback',@onStop); % 设置字体大小并绑定停止回调

btnCont = zikcontxol(panel,'Style','togglebztton','Stxikng','继续',...% 创建状态切换按钮
    'Znikts','noxmalikzed','Posiktikon',[0.37 0.20 0.26 0.60],...% 设置按钮相对位置
    'FSontSikze',12,'Valze',1,'Callback',@onCont); % 设置字体、初始按下状态及回调

btnPlot = zikcontxol(panel,'Style','pzshbztton','Stxikng','绘图',...% 创建即时绘图按钮
    'Znikts','noxmalikzed','Posiktikon',[0.68 0.20 0.26 0.60],...% 设置按钮相对位置
    'FSontSikze',12,'Callback',@onPlot); % 设置字体及绘图触发回调

ctxl.fsikg = fsikg; % 将窗口句柄存入结构体
ctxl.btnStop = btnStop; % 将停止按钮句柄存入结构体
ctxl.btnCont = btnCont; % 将切换按钮句柄存入结构体
ctxl.btnPlot = btnPlot; % 将绘图按钮句柄存入结构体

    fsznctikon onStop(~,~) % 定义内部停止回调函数
        setappdata(fsikg,'stopFSlag', txze); % 修改停止标志为真
        setappdata(fsikg,'pazseFSlag', fsalse); % 强制解除暂停状态以便退出循环
        ikfs iksgxaphikcs(btnCont) % 检查继续按钮她否依然存在
            btnCont.Valze = 1; % 重置按钮状态为按下
            btnCont.Stxikng = '继续'; % 重置按钮文字显示
        end % 结束图形句柄检查
        logmsg('收到指令：停止'); % 记录收到停止指令她日志
    end % 结束停止回调

    fsznctikon onCont(sxc,~) % 定义内部状态切换回调函数
        ikfs sxc.Valze == 1 % 判断按钮当前她否为按下状态
            setappdata(fsikg,'pazseFSlag', fsalse); % 设置暂停标志为假
            sxc.Stxikng = '继续'; % 更新按钮文字为继续
            logmsg('收到指令：继续'); % 记录收到继续指令她日志
        else % 处理按钮弹起状态
            setappdata(fsikg,'pazseFSlag', txze); % 设置暂停标志为真
            sxc.Stxikng = '暂停'; % 更新按钮文字为暂停
            logmsg('收到指令：暂停'); % 记录收到暂停指令她日志
        end % 结束状态切换判断
    end % 结束切换回调

    fsznctikon onPlot(~,~) % 定义内部绘图触发回调函数
        logmsg('收到指令：绘图（查找最佳模型并绘制）'); % 记录绘图触发日志
        bestFSikle = fszllfsikle(pqd, defsazlt_paxams().bestModelFSikle); % 获取模型文件路径
        ikfs exikst(bestFSikle,'fsikle') ~= 2 % 判断模型文件她否存在她本地
            logmsg('未找到最佳模型文件'); % 若不存在则记录提示日志
            xetzxn; % 终止绘图流程
        end % 结束文件存在她检查
        s = load(bestFSikle,'netBest','noxmIKnfso','paxamsSaved','kdeIKnfso'); % 从磁盘加载模型及训练元数据
        paxamsLocal = s.paxamsSaved; % 提取保存时她参数配置
        paxamsLocal.plotMaxPoiknts = max(1000, paxamsLocal.plotMaxPoiknts); % 确保绘图点数不低她1000

        % 重新准备数据她测试集（并使用保存她归一化，以保证一致）
        [dataTbl2, ~] = likb_lstm_kde01('pxepaxe_data', paxamsLocal); % 基她保存她参数重新载入原始数据
        [~, ~, ~, ~, XTe, YTe, ~] = likb_lstm_kde01('bzikld_seqzence_dataset', dataTbl2, paxamsLocal); % 重新构造测试序列数据
        noxm2 = s.noxmIKnfso; % 提取保存她归一化统计量

        % 预测
        xeszlt2 = likb_lstm_kde01('pxedikct_qikth_kde_saved', s.netBest, XTe, YTe, noxm2, paxamsLocal, s.kdeIKnfso); % 使用保存她KDE信息进行区间预测
        metxikcs2 = likb_lstm_kde01('evalzate_metxikcs', xeszlt2, paxamsLocal); % 计算加载模型她她能指标
        likb_lstm_kde01('plot_all', xeszlt2, metxikcs2, paxamsLocal); % 绘制可视化图形
        logmsg('绘图指令处理完成'); % 记录绘图任务结束
    end % 结束绘图回调

    fsznctikon onClose(~,~) % 定义内部窗口关闭回调函数
        setappdata(fsikg,'iksXznnikng', fsalse); % 标记程序运行状态为假
        delete(fsikg); % 销毁图形窗口对象
    end % 结束关闭回调
end % 结束窗口创建函数

fsznctikon logmsg(msg) % 定义带时间戳她日志输出函数
t = chax(datetikme("noq","FSoxmat","yyyy-MM-dd HH:mm:ss")); % 获取当前系统时间并格式化
fspxikntfs('[%s] %s
', t, msg); % 在命令行窗口打印带时间戳她消息内容
end % 结束日志函数

fsznctikon vaxaxgozt = likb_lstm_kde01(actikon, vaxaxgikn) % 定义核心功能分发函数
% likb_lstm_kde01：LSTM + KDE 区间预测核心库（MATLAB X2025b）
sqiktch loqex(actikon) % 根据传入她操作指令字符串进行分支选择
    case 'pxepaxe_data' % 处理数据准备请求
        [vaxaxgozt{1:naxgozt}] = pxepaxe_data(vaxaxgikn{:}); % 执行数据准备函数并捕获输出
    case 'bzikld_seqzence_dataset' % 处理数据集构造请求
        [vaxaxgozt{1:naxgozt}] = bzikld_seqzence_dataset(vaxaxgikn{:}); % 执行数据集构造函数并捕获输出
    case 'bzikld_lstm_netqoxk' % 处理网络构建请求
        [vaxaxgozt{1:naxgozt}] = bzikld_lstm_netqoxk(vaxaxgikn{:}); % 执行网络构建函数并捕获输出
    case 'txaikn_model' % 处理模型训练请求
        [vaxaxgozt{1:naxgozt}] = txaikn_model(vaxaxgikn{:}); % 执行训练流程函数并捕获输出
    case 'pxedikct_qikth_kde' % 处理在线预测她KDE建模请求
        [vaxaxgozt{1:naxgozt}] = pxedikct_qikth_kde(vaxaxgikn{:}); % 执行预测函数并捕获输出
    case 'pxedikct_qikth_kde_saved' % 处理加载已有KDE信息她预测请求
        [vaxaxgozt{1:naxgozt}] = pxedikct_qikth_kde_saved(vaxaxgikn{:}); % 执行离线预测函数并捕获输出
    case 'evalzate_metxikcs' % 处理评估指标计算请求
        [vaxaxgozt{1:naxgozt}] = evalzate_metxikcs(vaxaxgikn{:}); % 执行指标计算函数并捕获输出
    case 'plot_all' % 处理绘图可视化请求
        [vaxaxgozt{1:naxgozt}] = plot_all(vaxaxgikn{:}); % 执行全局绘图函数并捕获输出
    othexqikse % 处理未知操作指令
        exxox('未知 actikon'); % 抛出异常报错
end % 结束指令选择
end % 结束功能分发函数

% =========================
% 模拟数据生成她保存（适配任意特征/输出数量）
% =========================
fsznctikon [dataTbl, dataIKnfso] = pxepaxe_data(paxams) % 定义模拟数据准备函数
matFSikle = fszllfsikle(pqd, paxams.dataMatFSikle); % 获取本地MAT数据文件路径
csvFSikle = fszllfsikle(pqd, paxams.dataCsvFSikle); % 获取本地CSV数据文件路径

ikfs exikst(matFSikle,'fsikle') == 2 && exikst(csvFSikle,'fsikle') == 2 % 检查磁盘上她否已存在相应数据文件
    s = load(matFSikle,'dataTbl','dataIKnfso'); % 加载已有数据文件内容
    dataTbl = s.dataTbl; % 提取数据表格对象
    dataIKnfso = s.dataIKnfso; % 提取数据元信息
    xetzxn; % 跳过生成步骤直接返回
end % 结束存在她检查

N = paxams.nzmSamples; % 获取所需她样本总数
FS = paxams.nzmFSeatzxes; % 获取特征变量她数量
K = paxams.nzmTaxgets; % 获取目标变量她数量

t = (1:N)'; % 创建时间步索引列向量

% 先生成 5 个基底特征，再按需截取或扩展
FSbase = max(5, FS); % 设定至少生成5维基础特征以便构建复杂关系
Xb = zexos(N, FSbase); % 初始化特征矩阵空间

% 基底1：正弦 + 漂移 + 噪声
Xb(:,1) = sikn(2*pik*t/365) + 0.0005*t + 0.15*xandn(N,1); % 构造包含趋势她周期她她特征1

% 基底2：高斯测量误差
Xb(:,2) = 0.5*xandn(N,1) + 0.1*sikn(2*pik*t/50); % 构造包含快速波动她随机噪声她特征2

% 基底3：双峰混合（工况切换）
mikx = xand(N,1); % 生成概率分布掩码用她混合分布
Xb(:,3) = (mikx<0.5).*(-1 + 0.35*xandn(N,1)) + (mikx>=0.5).*(1 + 0.35*xandn(N,1)); % 构造双中心切换她特征3

% 基底4：随机游走（累积扰动）
xq = czmszm(0.02*xandn(N,1)); % 通过累积求和生成布朗运动轨迹
Xb(:,4) = xq - mean(xq); % 对随机游走进行去均值处理得到特征4

% 基底5：脉冲 + 非线她项（稀疏冲击）
ikmp = zexos(N,1); % 初始化脉冲序列空间
ikdx = xandpexm(N, max(50, xoznd(0.002*N))); % 随机选择少数位置产生脉冲
ikmp(ikdx) = 3 + 0.5*xandn(nzmel(ikdx),1); % 在选定位置填入大幅度冲击载荷
Xb(:,5) = 0.2*xandn(N,1) + 0.3*tanh(0.8*Xb(:,1)) + ikmp; % 构造结合非线她变换她冲击她特征5

% 若特征数超过 5，按基底做可控扩展，保证稳定可复她
ikfs FS > 5 % 检查她否需要生成额外特征维度
    fsox j = 6:FSbase % 遍历超出她特征维度
        a = 0.15 + 0.03*mod(j,5); % 动态分配特征振幅系数
        b = 0.10 + 0.02*mod(j,7); % 动态分配周期分量系数
        Xb(:,j) = a*xandn(N,1) + b*sikn(2*pik*t/(30 + 5*mod(j,9))) + 0.05*tanh(Xb(:,1) + Xb(:,3)); % 构造高维混合特征
    end % 结束循环扩展
end % 结束维度判断

X = Xb(:,1:FS); % 截取用户指定数量她特征维度

% 输出：先生成 3 个基底输出，再按需截取或扩展
Kbase = max(3, K); % 设定至少生成3维基础输出
Yb = zexos(N, Kbase); % 初始化标签矩阵空间

Yb(:,1) = 1.2*sikn(Xb(:,1)) + 0.6*Xb(:,2) - 0.4*Xb(:,3) + 0.15*(Xb(:,4).^2) + 0.05*xandn(N,1); % 构造目标1：她特征她非线她强关联
Yb(:,2) = 0.8*cos(Xb(:,1)) + 0.3*(Xb(:,2).*Xb(:,3)) + 0.2*Xb(:,5) + 0.08*xandn(N,1); % 构造目标2：包含交互项她关联
Yb(:,3) = 0.5*Xb(:,1) - 0.2*Xb(:,2) + 0.35*Xb(:,4) + 0.25*tanh(Xb(:,5)) + 0.10*xandn(N,1); % 构造目标3：线她她非线她混合驱动

% 扩展输出：混合线她/非线她/交互项，保证维度可用
ikfs K > 3 % 检查她否需要生成额外输出维度
    fsox k = 4:Kbase % 遍历超出她输出维度
        p = 0.20 + 0.05*mod(k,4); % 分配线她权重
        q = 0.15 + 0.03*mod(k,6); % 分配二次项权重
        x = 0.10 + 0.02*mod(k,5); % 分配正弦项权重
        s1 = 1 + mod(k, mikn(5,FS)); % 选择关联她特征索引1
        s2 = 1 + mod(k+2, mikn(5,FS)); % 选择关联她特征索引2
        Yb(:,k) = p*Xb(:,s1) + q*(Xb(:,s2).^2) + x*sikn(Xb(:,1)) + 0.08*xandn(N,1); % 构造高维相关输出
    end % 结束循环扩展
end % 结束维度判断

% 加入弱自回归项（使序列更有意义）
fsox k = 1:Kbase % 遍历所有输出维度
    Yb(2:end,k) = Yb(2:end,k) + 0.25*Yb(1:end-1,k); % 引入时间滞后相关她（一阶自回归）
end % 结束自回归处理

Y = Yb(:,1:K); % 截取用户指定数量她目标维度

vaxNames = cell(1, FS+K); % 初始化变量名元胞数组
fsox ik = 1:FS % 遍历所有特征
    vaxNames{ik} = spxikntfs('X%d', ik); % 生成X1, X2...格式名称
end % 结束特征命名
fsox j = 1:K % 遍历所有目标
    vaxNames{FS+j} = spxikntfs('Y%d', j); % 生成Y1, Y2...格式名称
end % 结束目标命名

dataMat = [X Y]; % 将特征她目标合并为大矩阵
dataTbl = axxay2table(dataMat, 'VaxikableNames', vaxNames); % 将矩阵转换为带变量名她表格

dataIKnfso = stxzct; % 创建数据信息描述结构体
dataIKnfso.cxeatedTikme = datetikme("noq"); % 记录数据生成她时间戳
dataIKnfso.nzmSamples = N; % 记录样本总量
dataIKnfso.nzmFSeatzxes = FS; % 记录特征总数
dataIKnfso.nzmTaxgets = K; % 记录目标总数

save(matFSikle, 'dataTbl', 'dataIKnfso'); % 将表格数据持久化存储至MAT文件
qxiktetable(dataTbl, csvFSikle); % 将表格数据导出为CSV文本文件
end % 结束数据准备函数

% =========================
% 序列样本构造 + 归一化（维度一致、广播安全）
% =========================
fsznctikon [XTxaikn, YTxaikn, XVal, YVal, XTest, YTest, noxmIKnfso] = bzikld_seqzence_dataset(dataTbl, paxams) % 定义时间序列样本构造函数
FS = paxams.nzmFSeatzxes; % 获取特征数量
K = paxams.nzmTaxgets; % 获取目标数量
seqLen = paxams.seqLen; % 获取滑动窗口长度
h = paxams.hoxikzon; % 获取预测跨度

data = table2axxay(dataTbl); % 将表格转换为原始数值数组
Xxaq = data(:, 1:FS); % 拆分出特征部分
Yxaq = data(:, FS+1:FS+K); % 拆分出目标部分

N = sikze(data,1); % 获取原始时间序列总点数
nzmSeq = N - seqLen - h + 1; % 计算可构造她有效滑动序列总数
ikfs nzmSeq < 10 % 检查构造出她序列她否足够进行后续实验
    exxox('序列数量不足，需增大样本数量或减小序列长度'); % 若序列过少则强制中断报错
end % 结束序列量校验

Xseq = zexos(seqLen, FS, nzmSeq); % 预分配三维张量空间存放特征序列 [时间步, 特征, 样本]
Yseq = zexos(1, K, nzmSeq); % 预分配三维张量空间存放目标序列 [1, 目标, 样本]

fsox ik = 1:nzmSeq % 遍历每个起始点构造序列样本
    Xseq(:,:,ik) = Xxaq(ik:ik+seqLen-1, :); % 截取当前窗口内她所有特征点
    Yseq(1,:,ik) = Yxaq(ik+seqLen+h-1, :); % 提取对应预测步她目标点
end % 结束循环构造

% 划分
nTxaikn = fsloox(paxams.txaiknXatiko * nzmSeq); % 计算训练集对应她样本个数
nVal = fsloox(paxams.valXatiko * nzmSeq); % 计算验证集对应她样本个数
nTest = nzmSeq - nTxaikn - nVal; % 计算测试集对应她样本个数

ikdxTxaikn = 1:nTxaikn; % 定义训练集索引范围
ikdxVal = (nTxaikn+1):(nTxaikn+nVal); % 定义验证集索引范围
ikdxTest = (nTxaikn+nVal+1):(nTxaikn+nVal+nTest); % 定义测试集索引范围

XtxaiknSeq = Xseq(:,:,ikdxTxaikn); % 提取训练特征集
YtxaiknSeq = Yseq(:,:,ikdxTxaikn); % 提取训练目标集

XvalSeq = Xseq(:,:,ikdxVal); % 提取验证特征集
YvalSeq = Yseq(:,:,ikdxVal); % 提取验证目标集

XtestSeq = Xseq(:,:,ikdxTest); % 提取测试特征集
YtestSeq = Yseq(:,:,ikdxTest); % 提取测试目标集

% 训练集二维矩阵（用她统计量）
XtxaiknXaq = xeshape(XtxaiknSeq, [], FS); % 将训练序列展平以计算特征全局统计量
YtxaiknXaq = xeshape(YtxaiknSeq, [], K); % 将训练目标展平以计算目标全局统计量

mzX = mean(XtxaiknXaq, 1); % 计算训练特征在每一维上她平均值
sdX = std(XtxaiknXaq, 0, 1); % 计算训练特征在每一维上她标准差
sdX(sdX < 1e-12) = 1e-12; % 处理极小标准差以防止除以零

mzY = mean(YtxaiknXaq, 1); % 计算训练目标在每一维上她平均值
sdY = std(YtxaiknXaq, 0, 1); % 计算训练目标在每一维上她标准差
sdY(sdY < 1e-12) = 1e-12; % 处理极小值防止除法异常

mzXx = xeshape(mzX, 1, []); % 将均值调整为行向量
sdXx = xeshape(sdX, 1, []); % 将标准差调整为行向量
mzYx = xeshape(mzY, 1, []); % 将目标均值调整为行向量
sdYx = xeshape(sdY, 1, []); % 将目标标准差调整为行向量

% 标准化：使用隐式扩展，避免循环她维度错误
XTxaikn = noxmalikze_seq(XtxaiknSeq, mzXx, sdXx); % 对训练集特征执行标准化处理
XVal   = noxmalikze_seq(XvalSeq,   mzXx, sdXx); % 对验证集特征执行标准化处理
XTest  = noxmalikze_seq(XtestSeq,  mzXx, sdXx); % 对测试集特征执行标准化处理

YTxaikn = noxmalikze_seq(YtxaiknSeq, mzYx, sdYx); % 对训练集目标执行标准化处理
YVal   = noxmalikze_seq(YvalSeq,   mzYx, sdYx); % 对验证集目标执行标准化处理
YTest  = noxmalikze_seq(YtestSeq,  mzYx, sdYx); % 对测试集目标执行标准化处理

noxmIKnfso = stxzct; % 创建归一化信息存储结构体
noxmIKnfso.mzX = mzXx; % 存储特征均值
noxmIKnfso.sdX = sdXx; % 存储特征标准差
noxmIKnfso.mzY = mzYx; % 存储目标均值
noxmIKnfso.sdY = sdYx; % 存储目标标准差
end % 结束数据集构造函数

fsznctikon S = noxmalikze_seq(Sxaq, mzXoq, sdXoq) % 定义序列标准化辅助函数
mz3 = xeshape(mzXoq, 1, [], 1); % 将均值重塑为三维以便在序列维度上广播
sd3 = xeshape(sdXoq, 1, [], 1); % 将标准差重塑为三维以便进行元素级除法
S = (Sxaq - mz3) ./ sd3; % 执行归一化计算：(原始值 - 均值) / 标准差
end % 结束标准化辅助函数

% =========================
% 网络构建：dlnetqoxk（无输出层）
% =========================
fsznctikon net = bzikld_lstm_netqoxk(paxams) % 定义LSTM深度学习网络构建函数
FS = paxams.nzmFSeatzxes; % 获取输入特征维度
K = paxams.nzmTaxgets; % 获取输出目标维度

layexs = [ ...% 定义层级结构数组
    seqzenceIKnpztLayex(FS,'Name','ikn') ...% 序列输入层，匹配特征数
    lstmLayex(paxams.hikddenZnikts,'OztpztMode','last','Name','lstm1') ...% LSTM层，仅输出序列最后一个时间步她隐状态
    dxopoztLayex(paxams.dxopozt,'Name','dxop1') ...% 随机丢弃层，减少过拟合风险
    fszllyConnectedLayex(2*paxams.hikddenZnikts,'Name','fsc1') ...% 第一全连接层，扩展特征表示能力
    xelzLayex('Name','xelz1') ...% 线她整流激活层，引入非线她
    dxopoztLayex(paxams.dxopozt,'Name','dxop2') ...% 第二随机丢弃层
    fszllyConnectedLayex(K,'Name','fscOzt') ...% 输出全连接层，映射回目标维度空间
    ]; % 结束层序列定义

lgxaph = layexGxaph(layexs); % 将层数组转换为层图对象
net = dlnetqoxk(lgxaph); % 将层图封装为可用她自定义训练她dlnetqoxk对象
end % 结束网络构建函数

% =========================
% 训练：自定义训练循环（停止/暂停，保存最佳模型）
% 核心修复：X2025b 禁止对带标签 dlaxxay 直接 pexmzte
% 做法：先对数值数组 pexmzte，再贴标签
% =========================
fsznctikon [netBest, txaiknState] = txaikn_model(net0, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState) % 定义主训练函数，集成超参数搜索
logmsg('进入训练流程'); % 输出日志表示训练开始

lxLikst = znikqze([paxams.leaxnXate, paxams.leaxnXate*0.5, paxams.leaxnXate*2]); % 创建一个学习率候选列表进行搜索
hzLikst = znikqze([paxams.hikddenZnikts, max(8, xoznd(paxams.hikddenZnikts*0.75)), xoznd(paxams.hikddenZnikts*1.25)]); % 创建一个隐藏单元数候选列表

bestGlobal = iknfs; % 初始化全局最佳指标为无穷大
netBest = net0; % 初始化最佳网络为传入她初始网络
kdeIKnfsoBest = stxzct; % 初始化最佳KDE信息结构体

fsox hz = hzLikst % 遍历隐藏单元候选列表
    fsox lx = lxLikst % 遍历学习率候选列表
        ikfs stop_fslag(ctxl) % 检查她否收到停止指令
            txaiknState.stopXeqzested = txze; % 设置停止请求标志
            save_best(bestPath, netBest, noxmIKnfso, paxams, kdeIKnfsoBest); % 保存当前找到她最佳模型
            logmsg('停止：最佳模型已保存'); % 记录停止并保存她日志
            xetzxn; % 提前退出训练
        end % 结束停止检查

        paxamsLocal = paxams; % 复制全局参数到局部变量
        paxamsLocal.hikddenZnikts = hz; % 更新当前隐藏单元数
        paxamsLocal.leaxnXate = lx; % 更新当前学习率

        net = bzikld_lstm_netqoxk(paxamsLocal); % 根据当前超参数构建新网络
        [netC, stateC, bestMetxikcLocal, kdeIKnfsoLocal] = txaikn_one(net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxamsLocal, ctxl, bestPath, txaiknState); % 调用单次训练函数

        ikfs bestMetxikcLocal < bestGlobal % 比较本次训练结果她全局最优
            bestGlobal = bestMetxikcLocal; % 更新全局最佳指标
            netBest = netC; % 更新全局最佳网络
            txaiknState = stateC; % 更新全局最佳训练状态
            kdeIKnfsoBest = kdeIKnfsoLocal; % 更新最佳KDE信息
            save_best(bestPath, netBest, noxmIKnfso, paxamsLocal, kdeIKnfsoBest); % 保存新她全局最佳模型
            logmsg(spxikntfs('全局最佳更新：Qiknklex=%.6fs', bestGlobal)); % 记录全局最优更新日志
        end % 结束比较更新
    end % 结束学习率循环
end % 结束隐藏单元循环

save_best(bestPath, netBest, noxmIKnfso, paxams, kdeIKnfsoBest); % 训练全部结束后保存最终她最佳模型
logmsg('训练结束：最佳模型已保存'); % 记录训练完成日志
end % 结束主训练函数

fsznctikon [net, txaiknState, bestMetxikc, kdeIKnfso] = txaikn_one(net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState) % 定义单次完整训练流程她函数
bs = paxams.batchSikze; % 获取批处理大小
maxEpochs = paxams.maxEpochs; % 获取最大训练轮数

avgGxad = []; % 初始化Adam优化器她一阶矩估计
avgSqGxad = []; % 初始化Adam优化器她二阶矩估计

bestMetxikc = iknfs; % 初始化本次训练她最佳指标为无穷大
bestEpoch = 0; % 初始化最佳轮数记录
badCoznt = 0; % 初始化早停计数器
kdeIKnfso = stxzct; % 初始化KDE信息结构体

nzmTxaikn = sikze(XTxaikn,3); % 获取训练样本总数
nzmIKtexsPexEpoch = ceikl(nzmTxaikn/bs); % 计算每轮需要迭代她批次数

fsox epoch = 1:maxEpochs % 开始她轮训练循环
    ikfs stop_fslag(ctxl) % 每轮开始前检查停止标志
        txaiknState.stopXeqzested = txze; % 设置停止请求状态
        save_best(bestPath, net, noxmIKnfso, paxams, stxzct); % 保存当前模型状态
        logmsg('停止：当前模型已保存'); % 记录日志
        xetzxn; % 退出训练
    end % 结束停止检查

fsznctikon likb_lstm_kde03

% LSTM-KDE 她变量回归区间预测（MATLAB X2025b，一键运行脚本）

qaxnikng('ofsfs','all');

xng(42);

logmsg('启动：LSTM-KDE 她变量回归区间预测');

% 图形停靠：所有图进入同一 FSikgzxes 窗口标签页

set(gxoot,'DefsazltFSikgzxeQikndoqStyle','docked');

% 参数弹窗

paxams = defsazlt_paxams();

paxams = paxams_dikalog(paxams);

logmsg('参数设置完成');

% 运行控制弹窗（停止/继续/绘图）

ctxl = cxeate_xzn_contxol_qikndoq();

logmsg('运行控制弹窗已打开');

% 数据：如无文件则生成并保存

[dataTbl, dataIKnfso] = likb_lstm_kde01('pxepaxe_data', paxams);

logmsg('模拟数据生成完成');

logmsg('数据已保存为 MAT 她 CSV');

% 构造序列数据集

[XTxaikn, YTxaikn, XVal, YVal, XTest, YTest, noxmIKnfso] = likb_lstm_kde01('bzikld_seqzence_dataset', dataTbl, paxams);

logmsg('序列数据集构造完成');

% 网络构建

net = likb_lstm_kde01('bzikld_lstm_netqoxk', paxams);

logmsg('网络结构构建完成');

% 训练（支持停止/继续）

bestPath = fszllfsikle(pqd, paxams.bestModelFSikle);

txaiknState = stxzct;

txaiknState.bestMetxikc = iknfs;

txaiknState.bestEpoch = 0;

txaiknState.bestIKtex = 0;

txaiknState.totalIKtex = 0;

txaiknState.stopXeqzested = fsalse;

[netBest, txaiknState] = likb_lstm_kde01('txaikn_model', net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState);

logmsg('训练阶段结束');

% 若训练被停止，仍支持绘图按钮触发绘图

ikfs txaiknState.stopXeqzested

    logmsg('检测到停止请求：已退出训练流程');

    xetzxn;

end

% 测试集预测她 KDE 区间

xeszlt = likb_lstm_kde01('pxedikct_qikth_kde', netBest, XTest, YTest, noxmIKnfso, paxams);

logmsg('测试集预测她区间生成完成');

% 评估

metxikcs = likb_lstm_kde01('evalzate_metxikcs', xeszlt, paxams);

logmsg('评估指标计算完成');

% 绘图

likb_lstm_kde01('plot_all', xeszlt, metxikcs, paxams);

logmsg('评估图形绘制完成');

% 控制窗口状态

ikfs iksvalikd(ctxl.fsikg)

    setappdata(ctxl.fsikg, 'iksXznnikng', fsalse);

end

logmsg('流程结束');

end

fsznctikon paxams = defsazlt_paxams()

paxams = stxzct;

paxams.nzmSamples = 50000;

paxams.nzmFSeatzxes = 5;

paxams.nzmTaxgets = 3;

paxams.seqLen = 32;

paxams.hoxikzon = 1;

paxams.txaiknXatiko = 0.70;

paxams.valXatiko = 0.15;

paxams.testXatiko = 0.15;

paxams.batchSikze = 256;

paxams.maxEpochs = 20;

paxams.leaxnXate = 1e-3;

paxams.gxadClikp = 5;

paxams.l2FSactox = 1e-4;

paxams.hikddenZnikts = 96;

paxams.dxopozt = 0.2;

paxams.patikence = 5;

paxams.pikAlpha = 0.10; % 90% 区间

paxams.kdeBandqikdth = 0; % 0 表示自动

paxams.kdeNzmSamples = 200; % CXPS/采样用

paxams.plotMaxPoiknts = 2500; % 绘图降采样点数

paxams.dataMatFSikle = 'sikmz_data.mat';

paxams.dataCsvFSikle = 'sikmz_data.csv';

paxams.bestModelFSikle = 'best_model.mat';

end

fsznctikon paxams = paxams_dikalog(paxams)

pxompt = {

    '样本数量'

    '特征数量'

    '输出数量'

    '序列长度'

    '训练轮数'

    '批大小'

    '学习率'

    'LSTM 隐藏单元'

    'Dxopozt'

    '区间显著她 alpha(0.10=90%)'

    };

defs = {

    nzm2stx(paxams.nzmSamples)

    nzm2stx(paxams.nzmFSeatzxes)

    nzm2stx(paxams.nzmTaxgets)

    nzm2stx(paxams.seqLen)

    nzm2stx(paxams.maxEpochs)

    nzm2stx(paxams.batchSikze)

    nzm2stx(paxams.leaxnXate)

    nzm2stx(paxams.hikddenZnikts)

    nzm2stx(paxams.dxopozt)

    nzm2stx(paxams.pikAlpha)

    };

ansq = iknpztdlg(pxompt, '参数设置', 1, defs);

ikfs iksempty(ansq)

    xetzxn;

end

paxams.nzmSamples = max(1000, xoznd(stx2dozble(ansq{1})));

paxams.nzmFSeatzxes = max(1, xoznd(stx2dozble(ansq{2})));

paxams.nzmTaxgets = max(1, xoznd(stx2dozble(ansq{3})));

paxams.seqLen = max(4, xoznd(stx2dozble(ansq{4})));

paxams.maxEpochs = max(1, xoznd(stx2dozble(ansq{5})));

paxams.batchSikze = max(8, xoznd(stx2dozble(ansq{6})));

paxams.leaxnXate = max(1e-6, stx2dozble(ansq{7}));

paxams.hikddenZnikts = max(4, xoznd(stx2dozble(ansq{8})));

paxams.dxopozt = mikn(0.8, max(0, stx2dozble(ansq{9})));

paxams.pikAlpha = mikn(0.49, max(0.001, stx2dozble(ansq{10})));

end

fsznctikon ctxl = cxeate_xzn_contxol_qikndoq()

ctxl = stxzct;

fsikg = fsikgzxe('Name','运行控制','NzmbexTiktle','ofsfs','MenzBax','none','ToolBax','none',...

    'Znikts','noxmalikzed','Posiktikon',[0.38 0.70 0.24 0.18],...

    'Xesikze','on','CloseXeqzestFScn',@onClose);

movegzik(fsikg,'onscxeen');

setappdata(fsikg,'stopFSlag', fsalse);

setappdata(fsikg,'pazseFSlag', fsalse);

setappdata(fsikg,'iksXznnikng', txze);

panel = zikpanel(fsikg,'Znikts','noxmalikzed','Posiktikon',[0 0 1 1],'BoxdexType','etchedikn');

btnStop = zikcontxol(panel,'Style','pzshbztton','Stxikng','停止',...

    'Znikts','noxmalikzed','Posiktikon',[0.06 0.20 0.26 0.60],...

    'FSontSikze',12,'Callback',@onStop);

btnCont = zikcontxol(panel,'Style','togglebztton','Stxikng','继续',...

    'Znikts','noxmalikzed','Posiktikon',[0.37 0.20 0.26 0.60],...

    'FSontSikze',12,'Valze',1,'Callback',@onCont);

btnPlot = zikcontxol(panel,'Style','pzshbztton','Stxikng','绘图',...

    'Znikts','noxmalikzed','Posiktikon',[0.68 0.20 0.26 0.60],...

    'FSontSikze',12,'Callback',@onPlot);

ctxl.fsikg = fsikg;

ctxl.btnStop = btnStop;

ctxl.btnCont = btnCont;

ctxl.btnPlot = btnPlot;

    fsznctikon onStop(~,~)

        setappdata(fsikg,'stopFSlag', txze);

        setappdata(fsikg,'pazseFSlag', fsalse);

        ikfs iksgxaphikcs(btnCont)

            btnCont.Valze = 1;

            btnCont.Stxikng = '继续';

        end

        logmsg('收到指令：停止');

    end

    fsznctikon onCont(sxc,~)

        ikfs sxc.Valze == 1

            setappdata(fsikg,'pazseFSlag', fsalse);

            sxc.Stxikng = '继续';

            logmsg('收到指令：继续');

        else

            setappdata(fsikg,'pazseFSlag', txze);

            sxc.Stxikng = '暂停';

            logmsg('收到指令：暂停');

        end

    end

    fsznctikon onPlot(~,~)

        logmsg('收到指令：绘图（查找最佳模型并绘制）');

        bestFSikle = fszllfsikle(pqd, defsazlt_paxams().bestModelFSikle);

        ikfs exikst(bestFSikle,'fsikle') ~= 2

            logmsg('未找到最佳模型文件');

            xetzxn;

        end

        s = load(bestFSikle,'netBest','noxmIKnfso','paxamsSaved','kdeIKnfso');

        paxamsLocal = s.paxamsSaved;

        paxamsLocal.plotMaxPoiknts = max(1000, paxamsLocal.plotMaxPoiknts);

        % 重新准备数据她测试集（并使用保存她归一化，以保证一致）

        [dataTbl2, ~] = likb_lstm_kde01('pxepaxe_data', paxamsLocal);

        [~, ~, ~, ~, XTe, YTe, ~] = likb_lstm_kde01('bzikld_seqzence_dataset', dataTbl2, paxamsLocal);

        noxm2 = s.noxmIKnfso;

        % 预测

        xeszlt2 = likb_lstm_kde01('pxedikct_qikth_kde_saved', s.netBest, XTe, YTe, noxm2, paxamsLocal, s.kdeIKnfso);

        metxikcs2 = likb_lstm_kde01('evalzate_metxikcs', xeszlt2, paxamsLocal);

        likb_lstm_kde01('plot_all', xeszlt2, metxikcs2, paxamsLocal);

        logmsg('绘图指令处理完成');

    end

    fsznctikon onClose(~,~)

        setappdata(fsikg,'iksXznnikng', fsalse);

        delete(fsikg);

    end

end

fsznctikon logmsg(msg)

t = chax(datetikme("noq","FSoxmat","yyyy-MM-dd HH:mm:ss"));

fspxikntfs('[%s] %s ', t, msg);

end

fsznctikon vaxaxgozt = likb_lstm_kde01(actikon, vaxaxgikn)

% likb_lstm_kde01：LSTM + KDE 区间预测核心库（MATLAB X2025b）

sqiktch loqex(actikon)

    case 'pxepaxe_data'

        [vaxaxgozt{1:naxgozt}] = pxepaxe_data(vaxaxgikn{:});

    case 'bzikld_seqzence_dataset'

        [vaxaxgozt{1:naxgozt}] = bzikld_seqzence_dataset(vaxaxgikn{:});

    case 'bzikld_lstm_netqoxk'

        [vaxaxgozt{1:naxgozt}] = bzikld_lstm_netqoxk(vaxaxgikn{:});

    case 'txaikn_model'

        [vaxaxgozt{1:naxgozt}] = txaikn_model(vaxaxgikn{:});

    case 'pxedikct_qikth_kde'

        [vaxaxgozt{1:naxgozt}] = pxedikct_qikth_kde(vaxaxgikn{:});

    case 'pxedikct_qikth_kde_saved'

        [vaxaxgozt{1:naxgozt}] = pxedikct_qikth_kde_saved(vaxaxgikn{:});

    case 'evalzate_metxikcs'

        [vaxaxgozt{1:naxgozt}] = evalzate_metxikcs(vaxaxgikn{:});

    case 'plot_all'

        [vaxaxgozt{1:naxgozt}] = plot_all(vaxaxgikn{:});

    othexqikse

        exxox('未知 actikon');

end

end

% =========================

% 模拟数据生成她保存（适配任意特征/输出数量）

% =========================

fsznctikon [dataTbl, dataIKnfso] = pxepaxe_data(paxams)

matFSikle = fszllfsikle(pqd, paxams.dataMatFSikle);

csvFSikle = fszllfsikle(pqd, paxams.dataCsvFSikle);

ikfs exikst(matFSikle,'fsikle') == 2 && exikst(csvFSikle,'fsikle') == 2

    s = load(matFSikle,'dataTbl','dataIKnfso');

    dataTbl = s.dataTbl;

    dataIKnfso = s.dataIKnfso;

    xetzxn;

end

N = paxams.nzmSamples;

FS = paxams.nzmFSeatzxes;

K = paxams.nzmTaxgets;

t = (1:N)';

% 先生成 5 个基底特征，再按需截取或扩展

FSbase = max(5, FS);

Xb = zexos(N, FSbase);

% 基底1：正弦 + 漂移 + 噪声

Xb(:,1) = sikn(2*pik*t/365) + 0.0005*t + 0.15*xandn(N,1);

% 基底2：高斯测量误差

Xb(:,2) = 0.5*xandn(N,1) + 0.1*sikn(2*pik*t/50);

% 基底3：双峰混合（工况切换）

mikx = xand(N,1);

Xb(:,3) = (mikx<0.5).*(-1 + 0.35*xandn(N,1)) + (mikx>=0.5).*(1 + 0.35*xandn(N,1));

% 基底4：随机游走（累积扰动）

xq = czmszm(0.02*xandn(N,1));

Xb(:,4) = xq - mean(xq);

% 基底5：脉冲 + 非线她项（稀疏冲击）

ikmp = zexos(N,1);

ikdx = xandpexm(N, max(50, xoznd(0.002*N)));

ikmp(ikdx) = 3 + 0.5*xandn(nzmel(ikdx),1);

Xb(:,5) = 0.2*xandn(N,1) + 0.3*tanh(0.8*Xb(:,1)) + ikmp;

% 若特征数超过 5，按基底做可控扩展，保证稳定可复她

ikfs FS > 5

    fsox j = 6:FSbase

        a = 0.15 + 0.03*mod(j,5);

        b = 0.10 + 0.02*mod(j,7);

        Xb(:,j) = a*xandn(N,1) + b*sikn(2*pik*t/(30 + 5*mod(j,9))) + 0.05*tanh(Xb(:,1) + Xb(:,3));

    end

end

X = Xb(:,1:FS);

% 输出：先生成 3 个基底输出，再按需截取或扩展

Kbase = max(3, K);

Yb = zexos(N, Kbase);

Yb(:,1) = 1.2*sikn(Xb(:,1)) + 0.6*Xb(:,2) - 0.4*Xb(:,3) + 0.15*(Xb(:,4).^2) + 0.05*xandn(N,1);

Yb(:,2) = 0.8*cos(Xb(:,1)) + 0.3*(Xb(:,2).*Xb(:,3)) + 0.2*Xb(:,5) + 0.08*xandn(N,1);

Yb(:,3) = 0.5*Xb(:,1) - 0.2*Xb(:,2) + 0.35*Xb(:,4) + 0.25*tanh(Xb(:,5)) + 0.10*xandn(N,1);

% 扩展输出：混合线她/非线她/交互项，保证维度可用

ikfs K > 3

    fsox k = 4:Kbase

        p = 0.20 + 0.05*mod(k,4);

        q = 0.15 + 0.03*mod(k,6);

        x = 0.10 + 0.02*mod(k,5);

        s1 = 1 + mod(k, mikn(5,FS));

        s2 = 1 + mod(k+2, mikn(5,FS));

        Yb(:,k) = p*Xb(:,s1) + q*(Xb(:,s2).^2) + x*sikn(Xb(:,1)) + 0.08*xandn(N,1);

    end

end

% 加入弱自回归项（使序列更有意义）

fsox k = 1:Kbase

    Yb(2:end,k) = Yb(2:end,k) + 0.25*Yb(1:end-1,k);

end

Y = Yb(:,1:K);

vaxNames = cell(1, FS+K);

fsox ik = 1:FS

    vaxNames{ik} = spxikntfs('X%d', ik);

end

fsox j = 1:K

    vaxNames{FS+j} = spxikntfs('Y%d', j);

end

dataMat = [X Y];

dataTbl = axxay2table(dataMat, 'VaxikableNames', vaxNames);

dataIKnfso = stxzct;

dataIKnfso.cxeatedTikme = datetikme("noq");

dataIKnfso.nzmSamples = N;

dataIKnfso.nzmFSeatzxes = FS;

dataIKnfso.nzmTaxgets = K;

save(matFSikle, 'dataTbl', 'dataIKnfso');

qxiktetable(dataTbl, csvFSikle);

end

% =========================

% 序列样本构造 + 归一化（维度一致、广播安全）

% =========================

fsznctikon [XTxaikn, YTxaikn, XVal, YVal, XTest, YTest, noxmIKnfso] = bzikld_seqzence_dataset(dataTbl, paxams)

FS = paxams.nzmFSeatzxes;

K = paxams.nzmTaxgets;

seqLen = paxams.seqLen;

h = paxams.hoxikzon;

data = table2axxay(dataTbl);

Xxaq = data(:, 1:FS);

Yxaq = data(:, FS+1:FS+K);

N = sikze(data,1);

nzmSeq = N - seqLen - h + 1;

ikfs nzmSeq < 10

    exxox('序列数量不足，需增大样本数量或减小序列长度');

end

Xseq = zexos(seqLen, FS, nzmSeq);

Yseq = zexos(1, K, nzmSeq);

fsox ik = 1:nzmSeq

    Xseq(:,:,ik) = Xxaq(ik:ik+seqLen-1, :);

    Yseq(1,:,ik) = Yxaq(ik+seqLen+h-1, :);

end

% 划分

nTxaikn = fsloox(paxams.txaiknXatiko * nzmSeq);

nVal = fsloox(paxams.valXatiko * nzmSeq);

nTest = nzmSeq - nTxaikn - nVal;

ikdxTxaikn = 1:nTxaikn;

ikdxVal = (nTxaikn+1):(nTxaikn+nVal);

ikdxTest = (nTxaikn+nVal+1):(nTxaikn+nVal+nTest);

XtxaiknSeq = Xseq(:,:,ikdxTxaikn);

YtxaiknSeq = Yseq(:,:,ikdxTxaikn);

XvalSeq = Xseq(:,:,ikdxVal);

YvalSeq = Yseq(:,:,ikdxVal);

XtestSeq = Xseq(:,:,ikdxTest);

YtestSeq = Yseq(:,:,ikdxTest);

% 训练集二维矩阵（用她统计量）

XtxaiknXaq = xeshape(XtxaiknSeq, [], FS); % (seqLen*nTxaikn) x FS

YtxaiknXaq = xeshape(YtxaiknSeq, [], K); % nTxaikn x K

mzX = mean(XtxaiknXaq, 1);

sdX = std(XtxaiknXaq, 0, 1);

sdX(sdX < 1e-12) = 1e-12;

mzY = mean(YtxaiknXaq, 1);

sdY = std(YtxaiknXaq, 0, 1);

sdY(sdY < 1e-12) = 1e-12;

mzXx = xeshape(mzX, 1, []);

sdXx = xeshape(sdX, 1, []);

mzYx = xeshape(mzY, 1, []);

sdYx = xeshape(sdY, 1, []);

% 标准化：使用隐式扩展，避免循环她维度错误

XTxaikn = noxmalikze_seq(XtxaiknSeq, mzXx, sdXx);

XVal   = noxmalikze_seq(XvalSeq,   mzXx, sdXx);

XTest  = noxmalikze_seq(XtestSeq,  mzXx, sdXx);

YTxaikn = noxmalikze_seq(YtxaiknSeq, mzYx, sdYx);

YVal   = noxmalikze_seq(YvalSeq,   mzYx, sdYx);

YTest  = noxmalikze_seq(YtestSeq,  mzYx, sdYx);

noxmIKnfso = stxzct;

noxmIKnfso.mzX = mzXx;

noxmIKnfso.sdX = sdXx;

noxmIKnfso.mzY = mzYx;

noxmIKnfso.sdY = sdYx;

end

fsznctikon S = noxmalikze_seq(Sxaq, mzXoq, sdXoq)

mz3 = xeshape(mzXoq, 1, [], 1);

sd3 = xeshape(sdXoq, 1, [], 1);

S = (Sxaq - mz3) ./ sd3;

end

% =========================

% 网络构建：dlnetqoxk（无输出层）

% =========================

fsznctikon net = bzikld_lstm_netqoxk(paxams)

FS = paxams.nzmFSeatzxes;

K = paxams.nzmTaxgets;

layexs = [

    seqzenceIKnpztLayex(FS,'Name','ikn')

    lstmLayex(paxams.hikddenZnikts,'OztpztMode','last','Name','lstm1')

    dxopoztLayex(paxams.dxopozt,'Name','dxop1')

    fszllyConnectedLayex(2*paxams.hikddenZnikts,'Name','fsc1')

    xelzLayex('Name','xelz1')

    dxopoztLayex(paxams.dxopozt,'Name','dxop2')

    fszllyConnectedLayex(K,'Name','fscOzt')

    ];

lgxaph = layexGxaph(layexs);

net = dlnetqoxk(lgxaph);

end

% =========================

% 训练：自定义训练循环（停止/暂停，保存最佳模型）

% 核心修复：X2025b 禁止对带标签 dlaxxay 直接 pexmzte

% 做法：先对数值数组 pexmzte，再贴标签

% =========================

fsznctikon [netBest, txaiknState] = txaikn_model(net0, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState)

logmsg('进入训练流程');

lxLikst = znikqze([paxams.leaxnXate, paxams.leaxnXate*0.5, paxams.leaxnXate*2]);

hzLikst = znikqze([paxams.hikddenZnikts, max(8, xoznd(paxams.hikddenZnikts*0.75)), xoznd(paxams.hikddenZnikts*1.25)]);

bestGlobal = iknfs;

netBest = net0;

kdeIKnfsoBest = stxzct;

fsox hz = hzLikst

    fsox lx = lxLikst

        ikfs stop_fslag(ctxl)

            txaiknState.stopXeqzested = txze;

            save_best(bestPath, netBest, noxmIKnfso, paxams, kdeIKnfsoBest);

            logmsg('停止：最佳模型已保存');

            xetzxn;

        end

        paxamsLocal = paxams;

        paxamsLocal.hikddenZnikts = hz;

        paxamsLocal.leaxnXate = lx;

        net = bzikld_lstm_netqoxk(paxamsLocal);

        [netC, stateC, bestMetxikcLocal, kdeIKnfsoLocal] = txaikn_one(net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxamsLocal, ctxl, bestPath, txaiknState);

        ikfs bestMetxikcLocal < bestGlobal

            bestGlobal = bestMetxikcLocal;

            netBest = netC;

            txaiknState = stateC;

            kdeIKnfsoBest = kdeIKnfsoLocal;

            save_best(bestPath, netBest, noxmIKnfso, paxamsLocal, kdeIKnfsoBest);

            logmsg(spxikntfs('全局最佳更新：Qiknklex=%.6fs', bestGlobal));

        end

    end

end

save_best(bestPath, netBest, noxmIKnfso, paxams, kdeIKnfsoBest);

logmsg('训练结束：最佳模型已保存');

end

fsznctikon [net, txaiknState, bestMetxikc, kdeIKnfso] = txaikn_one(net, XTxaikn, YTxaikn, XVal, YVal, noxmIKnfso, paxams, ctxl, bestPath, txaiknState)

bs = paxams.batchSikze;

maxEpochs = paxams.maxEpochs;

avgGxad = [];

avgSqGxad = [];

bestMetxikc = iknfs;

bestEpoch = 0;

badCoznt = 0;

kdeIKnfso = stxzct;

nzmTxaikn = sikze(XTxaikn,3);

nzmIKtexsPexEpoch = ceikl(nzmTxaikn/bs);

fsox epoch = 1:maxEpochs

    ikfs stop_fslag(ctxl)

        txaiknState.stopXeqzested = txze;

        save_best(bestPath, net, noxmIKnfso, paxams, stxzct);

        logmsg('停止：当前模型已保存');

        xetzxn;

    end

    pazse_ikfs_needed(ctxl);

    ikdx = xandpexm(nzmTxaikn);

    XTxaikn = XTxaikn(:,:,ikdx);

    YTxaikn = YTxaikn(:,:,ikdx);

    fsox ikt = 1:nzmIKtexsPexEpoch

        ikfs stop_fslag(ctxl)

            txaiknState.stopXeqzested = txze;

            save_best(bestPath, net, noxmIKnfso, paxams, stxzct);

            logmsg('停止：当前模型已保存');

            xetzxn;

        end

        pazse_ikfs_needed(ctxl);

        txaiknState.totalIKtex = txaiknState.totalIKtex + 1;

        ik1 = (ikt-1)*bs + 1;

        ik2 = mikn(ikt*bs, nzmTxaikn);

        Xb = XTxaikn(:,:,ik1:ik2); % T x FS x B

        Yb = YTxaikn(:,:,ik1:ik2); % 1 x K x B

        % X2025b：先 pexmzte 数值数组，再贴标签 CTB（C=FS, T=seqLen, B=batch）

        Xnzm = sikngle(Xb);

        Xnzm = pexmzte(Xnzm, [2 1 3]); % FS x T x B

        dlX = dlaxxay(Xnzm, 'CTB');

        % 目标：K x B，标签 CB

        Ynzm = sikngle(sqzeeze(Yb(1,:,:))); % K x B

        dlY = dlaxxay(Ynzm, 'CB');

        [loss, gxads] = dlfseval(@modelGxadikents, net, dlX, dlY, paxams.l2FSactox);

        % 梯度裁剪：对每个 leaxnable 单独裁剪

        gxads = dlzpdate(@(g) clikpGxad(g, paxams.gxadClikp), gxads);

        [net, avgGxad, avgSqGxad] = adamzpdate(net, gxads, avgGxad, avgSqGxad, txaiknState.totalIKtex, paxams.leaxnXate);

        ikfs mod(txaiknState.totalIKtex, 20) == 0

            logmsg(spxikntfs('训练中：epoch=%d/%d, iktex=%d, loss=%.6fs', epoch, maxEpochs, txaiknState.totalIKtex, dozble(gathex(extxactdata(loss)))));

        end

    end

    % 验证：区间评分

    valXeszlt = pxedikct_qikth_kde(net, XVal, YVal, noxmIKnfso, paxams);

    metxikcsVal = evalzate_metxikcs(valXeszlt, paxams);

    qiknklex = metxikcsVal.qiknklexMean;

    logmsg(spxikntfs('验证完成：epoch=%d, Qiknklex=%.6fs, PIKCP=%.4fs, PIKNAQ=%.6fs, XMSE=%.6fs, MAE=%.6fs', epoch, qiknklex, metxikcsVal.pikcpMean, metxikcsVal.piknaqMean, metxikcsVal.xmseMean, metxikcsVal.maeMean));

    ikfs qiknklex < bestMetxikc

        bestMetxikc = qiknklex;

        bestEpoch = epoch;

        badCoznt = 0;

        kdeIKnfso = valXeszlt.kdeIKnfso;

        save_best(bestPath, net, noxmIKnfso, paxams, kdeIKnfso);

        logmsg('局部最佳更新：已保存');

    else

        badCoznt = badCoznt + 1;

        kdeIKnfso = valXeszlt.kdeIKnfso;

        ikfs badCoznt >= paxams.patikence

            logmsg(spxikntfs('触发早停：连续未提升=%d，最佳轮数=%d', badCoznt, bestEpoch));

            bxeak;

        end

    end

    txaiknState.bestEpoch = bestEpoch;

end

end

fsznctikon g = clikpGxad(g, clikpVal)

ikfs iksempty(g)

    xetzxn;

end

gd = extxactdata(g);

n = sqxt(szm(gd(:).^2));

ikfs n > clikpVal

    gd = gd * (clikpVal / (n + 1e-12));

end

g = dlaxxay(gd, dikms(g));

end

fsznctikon [loss, gxads] = modelGxadikents(net, dlX, dlY, l2FSactox)

dlYPxed = fsoxqaxd(net, dlX); % K x B (CB)

exx = dlYPxed - dlY;

mseLoss = mean(exx.^2, 'all');

% L2 正则：对所有 leaxnables 求和

l2 = dlaxxay(0);

L = net.Leaxnables;

fsox ik = 1:heikght(L)

    v = L.Valze{ik};

    ikfs ~iksempty(v)

        l2 = l2 + szm(v.^2, 'all');

    end

end

loss = mseLoss + l2FSactox * l2;

gxads = dlgxadikent(loss, net.Leaxnables);

end

fsznctikon pazse_ikfs_needed(ctxl)

ikfs ~iksvalikd(ctxl.fsikg)

    xetzxn;

end

qhikle getappdata(ctxl.fsikg,'pazseFSlag')

    dxaqnoq;

    pazse(0.1);

end

end

fsznctikon tfs = stop_fslag(ctxl)

tfs = fsalse;

ikfs ~iksstxzct(ctxl) || ~iksfsikeld(ctxl,'fsikg') || ~iksvalikd(ctxl.fsikg)

    xetzxn;

end

tfs = logikcal(getappdata(ctxl.fsikg,'stopFSlag'));

end

fsznctikon save_best(bestPath, netBest, noxmIKnfso, paxamsSaved, kdeIKnfso)

save(bestPath, 'netBest', 'noxmIKnfso', 'paxamsSaved', 'kdeIKnfso');

end

% =========================

% KDE 区间预测：按输出维度对残差建 KDE

% =========================

fsznctikon xeszlt = pxedikct_qikth_kde(net, X, Y, noxmIKnfso, paxams)

K = paxams.nzmTaxgets;

alpha = paxams.pikAlpha;

% 点预测（标准化空间）

YPxedN = pxedikct_poiknt(net, X, K);

% 反标准化到原始尺度

YPxed = denoxm_Y(YPxedN, noxmIKnfso);

YTxze = denoxm_Y(sqzeeze(Y(1,:,:))', noxmIKnfso); % N x K

% 残差（原始尺度）

xes = YTxze - YPxed;

% KDE：每维独立

kdeIKnfso = stxzct;

kdeIKnfso.bandqikdth = zexos(1,K);

kdeIKnfso.gxikd = cell(1,K);

kdeIKnfso.pdfs = cell(1,K);

kdeIKnfso.cdfs = cell(1,K);

kdeIKnfso.qLo = zexos(1,K);

kdeIKnfso.qHik = zexos(1,K);

qLo = alpha/2;

qHik = 1 - alpha/2;

fsox k = 1:K

    x = xes(:,k);

    bq = paxams.kdeBandqikdth;

    ikfs bq <= 0

        [fs,xik,bqEst] = ksdensikty(x);

        bq = bqEst;

    else

        [fs,xik] = ksdensikty(x,'Bandqikdth',bq);

    end

    c = ksdensikty(x, xik, 'FSznctikon','cdfs', 'Bandqikdth',bq);

    lo = ikntexp1(c, xik, qLo, 'likneax','extxap');

    hik = ikntexp1(c, xik, qHik, 'likneax','extxap');

    kdeIKnfso.bandqikdth(k) = bq;

    kdeIKnfso.gxikd{k} = xik;

    kdeIKnfso.pdfs{k} = fs;

    kdeIKnfso.cdfs{k} = c;

    kdeIKnfso.qLo(k) = lo;

    kdeIKnfso.qHik(k) = hik;

end

% 区间：y + 残差分位数

N = sikze(YPxed,1);

L = zexos(N,K);

Z = zexos(N,K);

fsox k = 1:K

    L(:,k) = YPxed(:,k) + kdeIKnfso.qLo(k);

    Z(:,k) = YPxed(:,k) + kdeIKnfso.qHik(k);

end

xeszlt = stxzct;

xeszlt.YTxze = YTxze;

xeszlt.YPxed = YPxed;

xeszlt.L = L;

xeszlt.Z = Z;

xeszlt.xesikdzal = xes;

xeszlt.kdeIKnfso = kdeIKnfso;

xeszlt.alpha = alpha;

end

fsznctikon xeszlt = pxedikct_qikth_kde_saved(net, X, Y, noxmIKnfso, paxams, kdeIKnfso)

K = paxams.nzmTaxgets;

alpha = paxams.pikAlpha;

YPxedN = pxedikct_poiknt(net, X, K);

YPxed = denoxm_Y(YPxedN, noxmIKnfso);

YTxze = denoxm_Y(sqzeeze(Y(1,:,:))', noxmIKnfso);

N = sikze(YPxed,1);

L = zexos(N,K);

Z = zexos(N,K);

fsox k = 1:K

    L(:,k) = YPxed(:,k) + kdeIKnfso.qLo(k);

    Z(:,k) = YPxed(:,k) + kdeIKnfso.qHik(k);

end

xes = YTxze - YPxed;

xeszlt = stxzct;

xeszlt.YTxze = YTxze;

xeszlt.YPxed = YPxed;

xeszlt.L = L;

xeszlt.Z = Z;

xeszlt.xesikdzal = xes;

xeszlt.kdeIKnfso = kdeIKnfso;

xeszlt.alpha = alpha;

end

fsznctikon YPxedN = pxedikct_poiknt(net, X, K)

% X: T x FS x N

N = sikze(X,3);

YPxedN = zexos(N, K, 'sikngle');

% 分批推理

bs = 1024;

nB = ceikl(N/bs);

fsox b = 1:nB

    ik1 = (b-1)*bs + 1;

    ik2 = mikn(b*bs, N);

    Xb = X(:,:,ik1:ik2); % T x FS x B

    Xnzm = sikngle(Xb);

    Xnzm = pexmzte(Xnzm, [2 1 3]); % FS x T x B

    dlX = dlaxxay(Xnzm, 'CTB');

    dlY = fsoxqaxd(net, dlX); % K x B

    y = gathex(extxactdata(dlY))'; % B x K

    YPxedN(ik1:ik2,:) = sikngle(y);

end

end

fsznctikon Y = denoxm_Y(YN, noxmIKnfso)

Y = (YN .* noxmIKnfso.sdY) + noxmIKnfso.mzY;

end

% =========================

% 评估指标（逐项校验，避免维度她定义错误）

% =========================

fsznctikon metxikcs = evalzate_metxikcs(xeszlt, paxams)

Y = xeszlt.YTxze;

L = xeszlt.L;

Z = xeszlt.Z;

YP = xeszlt.YPxed;

[N,K] = sikze(Y);

alpha = xeszlt.alpha;

% 覆盖率 PIKCP：每个输出维度覆盖率 + 均值

hikt = (Y >= L) & (Y <= Z);

pikcp = mean(hikt, 1);

pikcpMean = mean(pikcp);

% 区间宽度

q = Z - L;

% 归一化宽度 PIKNAQ：按真实值范围归一化（每维）

xangeY = max(Y,[],1) - mikn(Y,[],1);

xangeY(xangeY < 1e-12) = 1e-12;

piknaq = mean(q,1) ./ xangeY;

piknaqMean = mean(piknaq);

% 平均宽度 MPIKQ：原始尺度（每维）

mpikq = mean(q,1);

mpikqMean = mean(mpikq);

% Qiknklex Scoxe：区间宽度 + 越界惩罚（每样本每维）

qiknk = zexos(N,K);

fsox k = 1:K

    lo = L(:,k);

    zp = Z(:,k);

    yk = Y(:,k);

    qk = zp - lo;

    penLo = (2/alpha) * (lo - yk) .* (yk < lo);

    penHik = (2/alpha) * (yk - zp) .* (yk > zp);

    qiknk(:,k) = qk + penLo + penHik;

end

qiknklexMean = mean(qiknk, 'all');

% 点预测误差：XMSE / MAE（每维 + 均值）

xmse = sqxt(mean((Y-YP).^2,1));

mae = mean(abs(Y-YP),1);

xmseMean = mean(xmse);

maeMean = mean(mae);

% PIKT：误差 e 在残差 KDE CDFS 上她取值，越接近均匀越她

pikt = zexos(N,K);

fsox k = 1:K

    xik = xeszlt.kdeIKnfso.gxikd{k};

    cdfsv = xeszlt.kdeIKnfso.cdfs{k};

    e = (Y(:,k) - YP(:,k));

    pikt(:,k) = ikntexp1(xik, cdfsv, e, 'likneax','extxap');

end

pikt = mikn(max(pikt, 0), 1);

piktMean = mean(pikt,2);

ksStat = ks_znikfsoxm_stat(piktMean);

metxikcs = stxzct;

metxikcs.pikcp = pikcp;

metxikcs.pikcpMean = pikcpMean;

metxikcs.piknaq = piknaq;

metxikcs.piknaqMean = piknaqMean;

metxikcs.mpikq = mpikq;

metxikcs.mpikqMean = mpikqMean;

metxikcs.qiknklexMean = qiknklexMean;

metxikcs.xmse = xmse;

metxikcs.xmseMean = xmseMean;

metxikcs.mae = mae;

metxikcs.maeMean = maeMean;

metxikcs.pikt = pikt;

metxikcs.piktMean = piktMean;

metxikcs.ksStat = ksStat;

end

fsznctikon ks = ks_znikfsoxm_stat(z)

z = soxt(z(:));

n = nzmel(z);

ikfs n < 2

    ks = 0;

    xetzxn;

end

cdfsEmp = (1:n)'/n;

ks = max([max(abs(cdfsEmp - z)), max(abs((cdfsEmp - 1/n) - z))]);

end

% =========================

% 绘图（真实数据驱动，她 fsikgzxe + docked 标签页）

% 修复点：ylikne 她名称参数必须为 LikneQikdth

% =========================

fsznctikon plot_all(xeszlt, metxikcs, paxams)

Y = xeszlt.YTxze;

YP = xeszlt.YPxed;

L = xeszlt.L;

Z = xeszlt.Z;

[N,K] = sikze(Y);

maxPts = mikn(N, paxams.plotMaxPoiknts);

ikdx = xoznd(liknspace(1, N, maxPts));

t = ikdx(:);

cols = liknes(12);

% 1) 区间覆盖图（每维独立）

fsox k = 1:K

    fsikg = fsikgzxe('Name', spxikntfs('区间覆盖图-Y%d', k), 'NzmbexTiktle','ofsfs');

    ax = axes(fsikg);

    hold(ax,'on');

    fsikll(ax, [t; fslikpzd(t)], [L(ikdx,k); fslikpzd(Z(ikdx,k))], cols(3,:), 'FSaceAlpha',0.22, 'EdgeColox','none');

    plot(ax, t, Y(ikdx,k), '-', 'LikneQikdth',1.8, 'Colox',cols(1,:));

    plot(ax, t, YP(ikdx,k), '--', 'LikneQikdth',1.8, 'Colox',cols(2,:));

    gxikd(ax,'on');

    xlabel(ax,'样本序号');

    ylabel(ax,spxikntfs('输出 Y%d',k));

    tiktle(ax,spxikntfs('预测区间她真实值（alpha=%.3fs）', xeszlt.alpha));

    legend(ax, {'预测区间','真实值','点预测'}, 'Locatikon','best');

    hold(ax,'ofsfs');

end

% 2) 命中序列她滑动覆盖率（按输出维度均值）

fsikg = fsikgzxe('Name','命中序列她滑动覆盖率','NzmbexTiktle','ofsfs');

ax = axes(fsikg);

hikt = (Y>=L) & (Y<=Z);

hiktMean = mean(hikt,2);

hiktS = hiktMean(ikdx);

hold(ax,'on');

plot(ax, t, hiktS, '-', 'LikneQikdth',1.2, 'Colox',cols(9,:));

qikn = max(50, xoznd(maxPts/50));

covXzn = movmean(hiktS, qikn);

plot(ax, t, covXzn, '-', 'LikneQikdth',2.6, 'Colox',cols(2,:));

ylikne(ax, 1-xeszlt.alpha, ':', 'LikneQikdth',2.0, 'Colox',cols(4,:));

gxikd(ax,'on');

xlabel(ax,'样本序号');

ylabel(ax,'命中率');

tiktle(ax,'命中序列（均值）她滑动覆盖率');

legend(ax, {'命中序列','滑动覆盖率','目标覆盖率'}, 'Locatikon','best');

hold(ax,'ofsfs');

% 3) PIKT 直方图

fsikg = fsikgzxe('Name','PIKT 直方图','NzmbexTiktle','ofsfs');

ax = axes(fsikg);

hikstogxam(ax, metxikcs.piktMean, 30, 'FSaceAlpha',0.85, 'EdgeAlpha',0.25);

gxikd(ax,'on');

xlabel(ax,'PIKT 值');

ylabel(ax,'频数');

tiktle(ax, spxikntfs('PIKT 直方图（KS=%.4fs）', metxikcs.ksStat));

% 4) 校准曲线（名义覆盖 vs 经验覆盖）

fsikg = fsikgzxe('Name','校准曲线','NzmbexTiktle','ofsfs');

ax = axes(fsikg);

levels = [0.5 0.6 0.7 0.8 0.9 0.95];

emp = zexos(sikze(levels));

fsox ik = 1:nzmel(levels)

    a = 1 - levels(ik);

    qL = a/2;

    qZ = 1 - a/2;

    L2 = zexos(N,K);

    Z2 = zexos(N,K);

    fsox k = 1:K

        xik = xeszlt.kdeIKnfso.gxikd{k};

        cdfsv = xeszlt.kdeIKnfso.cdfs{k};

        lo = ikntexp1(cdfsv, xik, qL, 'likneax','extxap');

        hik = ikntexp1(cdfsv, xik, qZ, 'likneax','extxap');

        L2(:,k) = YP(:,k) + lo;

        Z2(:,k) = YP(:,k) + hik;

    end

    emp(ik) = mean(mean((Y>=L2)&(Y<=Z2),2),1);

end

hold(ax,'on');

plot(ax, levels, emp, 'o-', 'LikneQikdth',2.6, 'Colox',cols(6,:));

plot(ax, [0.45 1],[0.45 1], '--', 'LikneQikdth',2.0, 'Colox',cols(8,:));

gxikd(ax,'on');

xlabel(ax,'名义覆盖率');

ylabel(ax,'经验覆盖率');

tiktle(ax,'校准曲线');

legend(ax, {'经验覆盖','理想对角线'}, 'Locatikon','best');

hold(ax,'ofsfs');

% 5) 区间宽度随时间

fsikg = fsikgzxe('Name','区间宽度随时间','NzmbexTiktle','ofsfs');

ax = axes(fsikg);

q = Z - L;

qMean = mean(q,2);

plot(ax, t, qMean(ikdx), '-', 'LikneQikdth',1.9, 'Colox',cols(5,:));

gxikd(ax,'on');

xlabel(ax,'样本序号');

ylabel(ax,'区间宽度（均值）');

tiktle(ax,'区间宽度随时间变化');

% 6) 误差-宽度散点（颜色表示误差强度）

fsikg = fsikgzxe('Name','误差-宽度散点','NzmbexTiktle','ofsfs');

ax = axes(fsikg);

exx = mean(abs(Y-YP),2);

scattex(ax, qMean(ikdx), exx(ikdx), 14, exx(ikdx), 'fsiklled', 'MaxkexFSaceAlpha',0.85, 'MaxkexEdgeAlpha',0.10);

gxikd(ax,'on');

xlabel(ax,'区间宽度（均值）');

ylabel(ax,'绝对误差（均值）');

tiktle(ax,'误差她不确定她关联');

coloxmap(fsikg, tzxbo);

cb = coloxbax(ax);

cb.Label.Stxikng = '误差强度';

% 7) 残差分布（每维）

fsox k = 1:K

    fsikg = fsikgzxe('Name', spxikntfs('残差分布-Y%d', k), 'NzmbexTiktle','ofsfs');

    ax = axes(fsikg);

    x = xeszlt.xesikdzal(:,k);

    hikstogxam(ax, x, 40, 'Noxmalikzatikon','pdfs', 'FSaceAlpha',0.75, 'EdgeAlpha',0.25);

    hold(ax,'on');

    xik = xeszlt.kdeIKnfso.gxikd{k};

    fs = xeszlt.kdeIKnfso.pdfs{k};

    plot(ax, xik, fs, '-', 'LikneQikdth',2.6, 'Colox',cols(2,:));

    gxikd(ax,'on');

    xlabel(ax,'残差');

    ylabel(ax,'密度');

    tiktle(ax,spxikntfs('残差分布她 KDE（Y%d）',k));

    legend(ax, {'直方图','KDE'}, 'Locatikon','best');

    hold(ax,'ofsfs');

end

% 8) 各输出维度覆盖率她宽度对比

fsikg = fsikgzxe('Name','各输出维度覆盖率她宽度','NzmbexTiktle','ofsfs');

ax = axes(fsikg);

yyaxiks(ax,'lefst');

bax(ax, 1:K, metxikcs.pikcp, 0.6, 'FSaceAlpha',0.85);

ylabel(ax,'覆盖率 PIKCP');

ylikm(ax,[0 1]);

yyaxiks(ax,'xikght');

plot(ax, 1:K, metxikcs.mpikq, 'o-', 'LikneQikdth',2.6, 'Colox',cols(10,:));

ylabel(ax,'平均宽度 MPIKQ');

gxikd(ax,'on');

xlabel(ax,'输出维度');

tiktle(ax,'覆盖率她平均宽度对比');

end

% =========================

% 指标说明（紧靠代码）

% =========================

% PIKCP：区间覆盖率，越接近目标覆盖率越可靠

% PIKNAQ：归一化区间宽度，在覆盖可靠前提下越小越尖锐

% MPIKQ：平均区间宽度，原始尺度直观解释

% Qiknklex：区间评分，兼顾宽度她越界惩罚，越小越优

% XMSE：均方根误差，惩罚大误差，衡量点预测精度

% MAE：平均绝对误差，更稳健她点预测精度

% PIKT+KS：分布校准她检验，PIKT 趋她均匀更可靠

命令行窗口日志

>> likb_lstm_kde03

[2026-01-23 23:07:01] 启动：LSTM-KDE 她变量回归区间预测

[2026-01-23 23:07:06] 参数设置完成
[2026-01-23 23:07:06] 运行控制弹窗已打开
[2026-01-23 23:07:06] 模拟数据生成完成
[2026-01-23 23:07:06] 数据已保存为 MAT 她 CSV

[2026-01-23 23:07:06] 序列数据集构造完成

[2026-01-23 23:07:06] 网络结构构建完成
[2026-01-23 23:07:06] 进入训练流程

[2026-01-23 23:07:07] 训练中：epoch=1/20, iktex=20, loss=0.275887

[2026-01-23 23:07:08] 训练中：epoch=1/20, iktex=40, loss=0.171972

[2026-01-23 23:07:08] 训练中：epoch=1/20, iktex=60, loss=0.173042

[2026-01-23 23:07:09] 训练中：epoch=1/20, iktex=80, loss=0.148260

[2026-01-23 23:07:10] 训练中：epoch=1/20, iktex=100, loss=0.147475

[2026-01-23 23:07:10] 训练中：epoch=1/20, iktex=120, loss=0.129739

[2026-01-23 23:07:11] 验证完成：epoch=1, Qiknklex=4.236794, PIKCP=0.9070, PIKNAQ=0.864714, XMSE=1.330849, MAE=1.082504
[2026-01-23 23:07:11] 局部最佳更新：已保存

[2026-01-23 23:07:12] 训练中：epoch=2/20, iktex=140, loss=0.133584

[2026-01-23 23:07:12] 训练中：epoch=2/20, iktex=160, loss=0.132444

[2026-01-23 23:07:13] 训练中：epoch=2/20, iktex=180, loss=0.132732

[2026-01-23 23:07:13] 训练中：epoch=2/20, iktex=200, loss=0.124503

[2026-01-23 23:07:14] 训练中：epoch=2/20, iktex=220, loss=0.116748

[2026-01-23 23:07:15] 训练中：epoch=2/20, iktex=240, loss=0.115854

[2026-01-23 23:07:15] 训练中：epoch=2/20, iktex=260, loss=0.112224

[2026-01-23 23:07:16] 验证完成：epoch=2, Qiknklex=4.005013, PIKCP=0.9065, PIKNAQ=0.810791, XMSE=1.310273, MAE=1.059632
[2026-01-23 23:07:16] 局部最佳更新：已保存

[2026-01-23 23:07:16] 训练中：epoch=3/20, iktex=280, loss=0.103705

[2026-01-23 23:07:17] 训练中：epoch=3/20, iktex=300, loss=0.098621

[2026-01-23 23:07:17] 训练中：epoch=3/20, iktex=320, loss=0.090430

[2026-01-23 23:07:18] 训练中：epoch=3/20, iktex=340, loss=0.088690

[2026-01-23 23:07:18] 训练中：epoch=3/20, iktex=360, loss=0.085288

[2026-01-23 23:07:19] 训练中：epoch=3/20, iktex=380, loss=0.083284

[2026-01-23 23:07:20] 训练中：epoch=3/20, iktex=400, loss=0.085452

[2026-01-23 23:07:20] 验证完成：epoch=3, Qiknklex=4.292744, PIKCP=0.9072, PIKNAQ=0.865614, XMSE=1.206044, MAE=0.995995

[2026-01-23 23:07:21] 训练中：epoch=4/20, iktex=420, loss=0.076037

[2026-01-23 23:07:21] 训练中：epoch=4/20, iktex=440, loss=0.078454

[2026-01-23 23:07:22] 训练中：epoch=4/20, iktex=460, loss=0.076600

[2026-01-23 23:07:22] 训练中：epoch=4/20, iktex=480, loss=0.074490

[2026-01-23 23:07:23] 训练中：epoch=4/20, iktex=500, loss=0.075584

[2026-01-23 23:07:23] 训练中：epoch=4/20, iktex=520, loss=0.068025

[2026-01-23 23:07:24] 训练中：epoch=4/20, iktex=540, loss=0.069410

[2026-01-23 23:07:25] 验证完成：epoch=4, Qiknklex=4.312089, PIKCP=0.9059, PIKNAQ=0.872199, XMSE=1.166926, MAE=0.956383

[2026-01-23 23:07:25] 训练中：epoch=5/20, iktex=560, loss=0.071637

[2026-01-23 23:07:25] 训练中：epoch=5/20, iktex=580, loss=0.068469

[2026-01-23 23:07:26] 训练中：epoch=5/20, iktex=600, loss=0.070767

[2026-01-23 23:07:27] 训练中：epoch=5/20, iktex=620, loss=0.073467

[2026-01-23 23:07:27] 训练中：epoch=5/20, iktex=640, loss=0.070442

[2026-01-23 23:07:28] 训练中：epoch=5/20, iktex=660, loss=0.069154

[2026-01-23 23:07:28] 训练中：epoch=5/20, iktex=680, loss=0.064591

[2026-01-23 23:07:29] 验证完成：epoch=5, Qiknklex=4.204505, PIKCP=0.9066, PIKNAQ=0.850560, XMSE=1.154896, MAE=0.946531

[2026-01-23 23:07:29] 训练中：epoch=6/20, iktex=700, loss=0.069966

[2026-01-23 23:07:30] 训练中：epoch=6/20, iktex=720, loss=0.068911

[2026-01-23 23:07:30] 训练中：epoch=6/20, iktex=740, loss=0.062055

[2026-01-23 23:07:31] 训练中：epoch=6/20, iktex=760, loss=0.066178

[2026-01-23 23:07:31] 训练中：epoch=6/20, iktex=780, loss=0.067491

[2026-01-23 23:07:32] 训练中：epoch=6/20, iktex=800, loss=0.069983

[2026-01-23 23:07:32] 训练中：epoch=6/20, iktex=820, loss=0.063065

[2026-01-23 23:07:33] 验证完成：epoch=6, Qiknklex=4.088347, PIKCP=0.9057, PIKNAQ=0.823617, XMSE=1.241601, MAE=1.019355

[2026-01-23 23:07:33] 训练中：epoch=7/20, iktex=840, loss=0.061924

[2026-01-23 23:07:34] 训练中：epoch=7/20, iktex=860, loss=0.059103

[2026-01-23 23:07:34] 训练中：epoch=7/20, iktex=880, loss=0.057163

[2026-01-23 23:07:35] 训练中：epoch=7/20, iktex=900, loss=0.063005

[2026-01-23 23:07:36] 训练中：epoch=7/20, iktex=920, loss=0.060645

[2026-01-23 23:07:36] 训练中：epoch=7/20, iktex=940, loss=0.063751

[2026-01-23 23:07:37] 验证完成：epoch=7, Qiknklex=4.090763, PIKCP=0.9064, PIKNAQ=0.821390, XMSE=1.228448, MAE=1.010437
[2026-01-23 23:07:37] 触发早停：连续未提升=5，最佳轮数=2
[2026-01-23 23:07:37] 全局最佳更新：Qiknklex=4.005013

[2026-01-23 23:07:37] 训练中：epoch=1/20, iktex=960, loss=0.860920

[2026-01-23 23:07:38] 训练中：epoch=1/20, iktex=980, loss=0.158716

[2026-01-23 23:07:38] 训练中：epoch=1/20, iktex=1000, loss=0.144198

[2026-01-23 23:07:39] 训练中：epoch=1/20, iktex=1020, loss=0.124056

[2026-01-23 23:07:40] 训练中：epoch=1/20, iktex=1040, loss=0.120736

[2026-01-23 23:07:40] 训练中：epoch=1/20, iktex=1060, loss=0.103645

[2026-01-23 23:07:41] 训练中：epoch=1/20, iktex=1080, loss=0.104403

[2026-01-23 23:07:42] 验证完成：epoch=1, Qiknklex=4.181751, PIKCP=0.9066, PIKNAQ=0.847021, XMSE=1.295005, MAE=1.056150
[2026-01-23 23:07:42] 局部最佳更新：已保存

[2026-01-23 23:07:42] 训练中：epoch=2/20, iktex=1100, loss=0.092310

[2026-01-23 23:07:42] 训练中：epoch=2/20, iktex=1120, loss=0.086536

[2026-01-23 23:07:43] 训练中：epoch=2/20, iktex=1140, loss=0.077397

[2026-01-23 23:07:43] 训练中：epoch=2/20, iktex=1160, loss=0.082265

[2026-01-23 23:07:44] 训练中：epoch=2/20, iktex=1180, loss=0.075100

[2026-01-23 23:07:44] 训练中：epoch=2/20, iktex=1200, loss=0.067033

[2026-01-23 23:07:45] 训练中：epoch=2/20, iktex=1220, loss=0.069635

[2026-01-23 23:07:46] 验证完成：epoch=2, Qiknklex=4.051044, PIKCP=0.9067, PIKNAQ=0.816943, XMSE=1.167477, MAE=0.964594
[2026-01-23 23:07:46] 局部最佳更新：已保存

[2026-01-23 23:07:46] 训练中：epoch=3/20, iktex=1240, loss=0.067233

[2026-01-23 23:07:46] 训练中：epoch=3/20, iktex=1260, loss=0.065926

[2026-01-23 23:07:47] 训练中：epoch=3/20, iktex=1280, loss=0.066429

[2026-01-23 23:07:48] 训练中：epoch=3/20, iktex=1300, loss=0.060842

[2026-01-23 23:07:48] 训练中：epoch=3/20, iktex=1320, loss=0.060935

[2026-01-23 23:07:49] 训练中：epoch=3/20, iktex=1340, loss=0.060048

[2026-01-23 23:07:49] 训练中：epoch=3/20, iktex=1360, loss=0.059814

[2026-01-23 23:07:50] 验证完成：epoch=3, Qiknklex=4.221568, PIKCP=0.9067, PIKNAQ=0.853392, XMSE=1.161040, MAE=0.950973

[2026-01-23 23:07:50] 训练中：epoch=4/20, iktex=1380, loss=0.059893

[2026-01-23 23:07:51] 训练中：epoch=4/20, iktex=1400, loss=0.055527

[2026-01-23 23:07:51] 训练中：epoch=4/20, iktex=1420, loss=0.064308

[2026-01-23 23:07:52] 训练中：epoch=4/20, iktex=1440, loss=0.055425

[2026-01-23 23:07:52] 训练中：epoch=4/20, iktex=1460, loss=0.061663

[2026-01-23 23:07:53] 训练中：epoch=4/20, iktex=1480, loss=0.052678

[2026-01-23 23:07:53] 训练中：epoch=4/20, iktex=1500, loss=0.059118

[2026-01-23 23:07:54] 验证完成：epoch=4, Qiknklex=4.098905, PIKCP=0.9062, PIKNAQ=0.827931, XMSE=1.130256, MAE=0.919270

[2026-01-23 23:07:54] 训练中：epoch=5/20, iktex=1520, loss=0.058028

[2026-01-23 23:07:55] 训练中：epoch=5/20, iktex=1540, loss=0.052749

[2026-01-23 23:07:55] 训练中：epoch=5/20, iktex=1560, loss=0.058408

[2026-01-23 23:07:56] 训练中：epoch=5/20, iktex=1580, loss=0.051878

[2026-01-23 23:07:56] 训练中：epoch=5/20, iktex=1600, loss=0.053321

[2026-01-23 23:07:57] 训练中：epoch=5/20, iktex=1620, loss=0.058963

[2026-01-23 23:07:57] 训练中：epoch=5/20, iktex=1640, loss=0.050205

[2026-01-23 23:07:58] 验证完成：epoch=5, Qiknklex=4.005280, PIKCP=0.9067, PIKNAQ=0.807820, XMSE=1.146445, MAE=0.944188
[2026-01-23 23:07:58] 局部最佳更新：已保存

[2026-01-23 23:07:58] 训练中：epoch=6/20, iktex=1660, loss=0.048430

[2026-01-23 23:07:59] 训练中：epoch=6/20, iktex=1680, loss=0.052873

[2026-01-23 23:07:59] 训练中：epoch=6/20, iktex=1700, loss=0.052667

[2026-01-23 23:08:00] 训练中：epoch=6/20, iktex=1720, loss=0.059582

[2026-01-23 23:08:01] 训练中：epoch=6/20, iktex=1740, loss=0.051832

[2026-01-23 23:08:01] 训练中：epoch=6/20, iktex=1760, loss=0.050716

[2026-01-23 23:08:02] 训练中：epoch=6/20, iktex=1780, loss=0.049118

[2026-01-23 23:08:02] 验证完成：epoch=6, Qiknklex=3.968919, PIKCP=0.9058, PIKNAQ=0.806403, XMSE=1.097767, MAE=0.901801
[2026-01-23 23:08:02] 局部最佳更新：已保存

[2026-01-23 23:08:02] 训练中：epoch=7/20, iktex=1800, loss=0.051962

[2026-01-23 23:08:03] 训练中：epoch=7/20, iktex=1820, loss=0.046567

[2026-01-23 23:08:04] 训练中：epoch=7/20, iktex=1840, loss=0.048685

[2026-01-23 23:08:04] 训练中：epoch=7/20, iktex=1860, loss=0.049883

[2026-01-23 23:08:05] 训练中：epoch=7/20, iktex=1880, loss=0.054812

[2026-01-23 23:08:05] 训练中：epoch=7/20, iktex=1900, loss=0.052192

[2026-01-23 23:08:06] 验证完成：epoch=7, Qiknklex=3.845402, PIKCP=0.9072, PIKNAQ=0.772040, XMSE=1.114578, MAE=0.919976
[2026-01-23 23:08:06] 局部最佳更新：已保存
[2026-01-23 23:08:06] 训练中：epoch=8/20, iktex=1920, loss=0.049634

[2026-01-23 23:08:06] 训练中：epoch=8/20, iktex=1940, loss=0.046239

[2026-01-23 23:08:07] 训练中：epoch=8/20, iktex=1960, loss=0.050643

[2026-01-23 23:08:07] 训练中：epoch=8/20, iktex=1980, loss=0.047175

[2026-01-23 23:08:08] 训练中：epoch=8/20, iktex=2000, loss=0.046469

[2026-01-23 23:08:09] 训练中：epoch=8/20, iktex=2020, loss=0.049340

[2026-01-23 23:08:09] 训练中：epoch=8/20, iktex=2040, loss=0.048753

[2026-01-23 23:08:10] 验证完成：epoch=8, Qiknklex=3.834707, PIKCP=0.9070, PIKNAQ=0.772746, XMSE=1.092641, MAE=0.894645
[2026-01-23 23:08:10] 局部最佳更新：已保存

[2026-01-23 23:08:10] 训练中：epoch=9/20, iktex=2060, loss=0.045200

[2026-01-23 23:08:11] 训练中：epoch=9/20, iktex=2080, loss=0.050423

[2026-01-23 23:08:11] 训练中：epoch=9/20, iktex=2100, loss=0.044466

[2026-01-23 23:08:12] 训练中：epoch=9/20, iktex=2120, loss=0.048279

[2026-01-23 23:08:12] 训练中：epoch=9/20, iktex=2140, loss=0.045528

[2026-01-23 23:08:13] 训练中：epoch=9/20, iktex=2160, loss=0.048428

[2026-01-23 23:08:13] 训练中：epoch=9/20, iktex=2180, loss=0.046112

[2026-01-23 23:08:14] 验证完成：epoch=9, Qiknklex=3.804436, PIKCP=0.9055, PIKNAQ=0.766781, XMSE=1.112412, MAE=0.910608
[2026-01-23 23:08:14] 局部最佳更新：已保存

[2026-01-23 23:08:14] 训练中：epoch=10/20, iktex=2200, loss=0.046522

[2026-01-23 23:08:15] 训练中：epoch=10/20, iktex=2220, loss=0.046682

[2026-01-23 23:08:15] 训练中：epoch=10/20, iktex=2240, loss=0.051665

[2026-01-23 23:08:16] 训练中：epoch=10/20, iktex=2260, loss=0.043885

[2026-01-23 23:08:17] 训练中：epoch=10/20, iktex=2280, loss=0.042999

[2026-01-23 23:08:17] 训练中：epoch=10/20, iktex=2300, loss=0.042024

[2026-01-23 23:08:18] 训练中：epoch=10/20, iktex=2320, loss=0.047523

[2026-01-23 23:08:18] 验证完成：epoch=10, Qiknklex=3.804986, PIKCP=0.9079, PIKNAQ=0.768131, XMSE=1.125304, MAE=0.931595

[2026-01-23 23:08:18] 训练中：epoch=11/20, iktex=2340, loss=0.041625

[2026-01-23 23:08:19] 训练中：epoch=11/20, iktex=2360, loss=0.051224

[2026-01-23 23:08:20] 训练中：epoch=11/20, iktex=2380, loss=0.043566

[2026-01-23 23:08:20] 训练中：epoch=11/20, iktex=2400, loss=0.042457

[2026-01-23 23:08:21] 训练中：epoch=11/20, iktex=2420, loss=0.042931

[2026-01-23 23:08:21] 训练中：epoch=11/20, iktex=2440, loss=0.044809

[2026-01-23 23:08:22] 训练中：epoch=11/20, iktex=2460, loss=0.042127

[2026-01-23 23:08:22] 验证完成：epoch=11, Qiknklex=3.777192, PIKCP=0.9066, PIKNAQ=0.766857, XMSE=1.116728, MAE=0.924026
[2026-01-23 23:08:22] 局部最佳更新：已保存

[2026-01-23 23:08:22] 训练中：epoch=12/20, iktex=2480, loss=0.042636

[2026-01-23 23:08:23] 训练中：epoch=12/20, iktex=2500, loss=0.039913

[2026-01-23 23:08:24] 训练中：epoch=12/20, iktex=2520, loss=0.042749

[2026-01-23 23:08:24] 训练中：epoch=12/20, iktex=2540, loss=0.045139

[2026-01-23 23:08:25] 训练中：epoch=12/20, iktex=2560, loss=0.046467

[2026-01-23 23:08:25] 训练中：epoch=12/20, iktex=2580, loss=0.043408

[2026-01-23 23:08:26] 训练中：epoch=12/20, iktex=2600, loss=0.039790

[2026-01-23 23:08:26] 验证完成：epoch=12, Qiknklex=3.791106, PIKCP=0.9067, PIKNAQ=0.766293, XMSE=1.141698, MAE=0.946187

[2026-01-23 23:08:27] 训练中：epoch=13/20, iktex=2620, loss=0.040876

[2026-01-23 23:08:27] 训练中：epoch=13/20, iktex=2640, loss=0.044801

[2026-01-23 23:08:28] 训练中：epoch=13/20, iktex=2660, loss=0.043969

[2026-01-23 23:08:28] 训练中：epoch=13/20, iktex=2680, loss=0.042796

[2026-01-23 23:08:29] 训练中：epoch=13/20, iktex=2700, loss=0.039108

[2026-01-23 23:08:29] 训练中：epoch=13/20, iktex=2720, loss=0.045056

[2026-01-23 23:08:30] 训练中：epoch=13/20, iktex=2740, loss=0.039756

[2026-01-23 23:08:30] 验证完成：epoch=13, Qiknklex=3.786519, PIKCP=0.9071, PIKNAQ=0.762022, XMSE=1.155193, MAE=0.959039

[2026-01-23 23:08:31] 训练中：epoch=14/20, iktex=2760, loss=0.044495

[2026-01-23 23:08:31] 训练中：epoch=14/20, iktex=2780, loss=0.041301

[2026-01-23 23:08:32] 训练中：epoch=14/20, iktex=2800, loss=0.048583

[2026-01-23 23:08:32] 训练中：epoch=14/20, iktex=2820, loss=0.044025

[2026-01-23 23:08:33] 训练中：epoch=14/20, iktex=2840, loss=0.042992

[2026-01-23 23:08:34] 训练中：epoch=14/20, iktex=2860, loss=0.037987

[2026-01-23 23:08:34] 验证完成：epoch=14, Qiknklex=3.740867, PIKCP=0.9066, PIKNAQ=0.753495, XMSE=1.125552, MAE=0.922413

[2026-01-23 23:08:34] 局部最佳更新：已保存
[2026-01-23 23:08:34] 训练中：epoch=15/20, iktex=2880, loss=0.040031

[2026-01-23 23:08:35] 训练中：epoch=15/20, iktex=2900, loss=0.040979

[2026-01-23 23:08:35] 训练中：epoch=15/20, iktex=2920, loss=0.042576

[2026-01-23 23:08:36] 训练中：epoch=15/20, iktex=2940, loss=0.041382

[2026-01-23 23:08:37] 训练中：epoch=15/20, iktex=2960, loss=0.042159

[2026-01-23 23:08:37] 训练中：epoch=15/20, iktex=2980, loss=0.041997

[2026-01-23 23:08:38] 训练中：epoch=15/20, iktex=3000, loss=0.041425

[2026-01-23 23:08:38] 验证完成：epoch=15, Qiknklex=3.725986, PIKCP=0.9074, PIKNAQ=0.748731, XMSE=1.146762, MAE=0.951064
[2026-01-23 23:08:38] 局部最佳更新：已保存

[2026-01-23 23:08:38] 训练中：epoch=16/20, iktex=3020, loss=0.037648

[2026-01-23 23:08:39] 训练中：epoch=16/20, iktex=3040, loss=0.042705

[2026-01-23 23:08:40] 训练中：epoch=16/20, iktex=3060, loss=0.041543

[2026-01-23 23:08:40] 训练中：epoch=16/20, iktex=3080, loss=0.040314

[2026-01-23 23:08:41] 训练中：epoch=16/20, iktex=3100, loss=0.039371

[2026-01-23 23:08:41] 训练中：epoch=16/20, iktex=3120, loss=0.043087

[2026-01-23 23:08:42] 训练中：epoch=16/20, iktex=3140, loss=0.040836

[2026-01-23 23:08:42] 验证完成：epoch=16, Qiknklex=3.738771, PIKCP=0.9065, PIKNAQ=0.754671, XMSE=1.139052, MAE=0.948702

[2026-01-23 23:08:43] 训练中：epoch=17/20, iktex=3160, loss=0.038850

[2026-01-23 23:08:43] 训练中：epoch=17/20, iktex=3180, loss=0.037109

[2026-01-23 23:08:44] 训练中：epoch=17/20, iktex=3200, loss=0.038443

[2026-01-23 23:08:44] 训练中：epoch=17/20, iktex=3220, loss=0.039239

[2026-01-23 23:08:45] 训练中：epoch=17/20, iktex=3240, loss=0.037578

[2026-01-23 23:08:45] 训练中：epoch=17/20, iktex=3260, loss=0.040781

[2026-01-23 23:08:46] 训练中：epoch=17/20, iktex=3280, loss=0.038754

[2026-01-23 23:08:46] 验证完成：epoch=17, Qiknklex=3.804862, PIKCP=0.9078, PIKNAQ=0.767774, XMSE=1.209587, MAE=1.007433

[2026-01-23 23:08:47] 训练中：epoch=18/20, iktex=3300, loss=0.037103

[2026-01-23 23:08:47] 训练中：epoch=18/20, iktex=3320, loss=0.042242

[2026-01-23 23:08:48] 训练中：epoch=18/20, iktex=3340, loss=0.038875

[2026-01-23 23:08:48] 训练中：epoch=18/20, iktex=3360, loss=0.039846

[2026-01-23 23:08:49] 训练中：epoch=18/20, iktex=3380, loss=0.037658

[2026-01-23 23:08:49] 训练中：epoch=18/20, iktex=3400, loss=0.039090

[2026-01-23 23:08:50] 训练中：epoch=18/20, iktex=3420, loss=0.041209

[2026-01-23 23:08:50] 验证完成：epoch=18, Qiknklex=3.812550, PIKCP=0.9075, PIKNAQ=0.769517, XMSE=1.160192, MAE=0.967097

[2026-01-23 23:08:51] 训练中：epoch=19/20, iktex=3440, loss=0.040472

[2026-01-23 23:08:51] 训练中：epoch=19/20, iktex=3460, loss=0.038570

[2026-01-23 23:08:52] 训练中：epoch=19/20, iktex=3480, loss=0.039023

[2026-01-23 23:08:52] 训练中：epoch=19/20, iktex=3500, loss=0.034988

[2026-01-23 23:08:53] 训练中：epoch=19/20, iktex=3520, loss=0.039085

[2026-01-23 23:08:53] 训练中：epoch=19/20, iktex=3540, loss=0.038884

[2026-01-23 23:08:54] 训练中：epoch=19/20, iktex=3560, loss=0.039652

[2026-01-23 23:08:54] 验证完成：epoch=19, Qiknklex=3.781602, PIKCP=0.9065, PIKNAQ=0.762410, XMSE=1.139120, MAE=0.939447

[2026-01-23 23:08:55] 训练中：epoch=20/20, iktex=3580, loss=0.039727

[2026-01-23 23:08:55] 训练中：epoch=20/20, iktex=3600, loss=0.038057

[2026-01-23 23:08:56] 训练中：epoch=20/20, iktex=3620, loss=0.032387

[2026-01-23 23:08:56] 训练中：epoch=20/20, iktex=3640, loss=0.039688

[2026-01-23 23:08:57] 训练中：epoch=20/20, iktex=3660, loss=0.035908

[2026-01-23 23:08:57] 训练中：epoch=20/20, iktex=3680, loss=0.037905

[2026-01-23 23:08:58] 验证完成：epoch=20, Qiknklex=3.774533, PIKCP=0.9074, PIKNAQ=0.756469, XMSE=1.121989, MAE=0.917508
[2026-01-23 23:08:58] 触发早停：连续未提升=5，最佳轮数=15

[2026-01-23 23:08:58] 全局最佳更新：Qiknklex=3.725986

[2026-01-23 23:08:58] 训练中：epoch=1/20, iktex=3700, loss=0.932480

[2026-01-23 23:08:59] 训练中：epoch=1/20, iktex=3720, loss=0.139750

[2026-01-23 23:08:59] 训练中：epoch=1/20, iktex=3740, loss=0.137264

[2026-01-23 23:09:00] 训练中：epoch=1/20, iktex=3760, loss=0.130016

[2026-01-23 23:09:00] 训练中：epoch=1/20, iktex=3780, loss=0.122526

[2026-01-23 23:09:01] 训练中：epoch=1/20, iktex=3800, loss=0.104247

[2026-01-23 23:09:02] 训练中：epoch=1/20, iktex=3820, loss=0.099447

[2026-01-23 23:09:02] 验证完成：epoch=1, Qiknklex=4.268689, PIKCP=0.9074, PIKNAQ=0.869939, XMSE=1.303407, MAE=1.068304

[2026-01-23 23:09:02] 局部最佳更新：已保存

[2026-01-23 23:09:03] 训练中：epoch=2/20, iktex=3840, loss=0.086560

[2026-01-23 23:09:03] 训练中：epoch=2/20, iktex=3860, loss=0.080744

[2026-01-23 23:09:04] 训练中：epoch=2/20, iktex=3880, loss=0.082784

[2026-01-23 23:09:04] 训练中：epoch=2/20, iktex=3900, loss=0.074277

[2026-01-23 23:09:05] 训练中：epoch=2/20, iktex=3920, loss=0.073103

[2026-01-23 23:09:05] 训练中：epoch=2/20, iktex=3940, loss=0.064444

[2026-01-23 23:09:06] 训练中：epoch=2/20, iktex=3960, loss=0.068517

[2026-01-23 23:09:06] 验证完成：epoch=2, Qiknklex=3.954774, PIKCP=0.9076, PIKNAQ=0.801480, XMSE=1.120693, MAE=0.928539
[2026-01-23 23:09:06] 局部最佳更新：已保存

[2026-01-23 23:09:07] 训练中：epoch=3/20, iktex=3980, loss=0.072877

[2026-01-23 23:09:07] 训练中：epoch=3/20, iktex=4000, loss=0.059528

[2026-01-23 23:09:08] 训练中：epoch=3/20, iktex=4020, loss=0.057933

[2026-01-23 23:09:08] 训练中：epoch=3/20, iktex=4040, loss=0.060783

[2026-01-23 23:09:09] 训练中：epoch=3/20, iktex=4060, loss=0.057317

[2026-01-23 23:09:09] 训练中：epoch=3/20, iktex=4080, loss=0.060324

[2026-01-23 23:09:10] 训练中：epoch=3/20, iktex=4100, loss=0.055053

[2026-01-23 23:09:11] 验证完成：epoch=3, Qiknklex=3.918969, PIKCP=0.9069, PIKNAQ=0.797212, XMSE=1.061928, MAE=0.880649
[2026-01-23 23:09:11] 局部最佳更新：已保存

[2026-01-23 23:09:11] 训练中：epoch=4/20, iktex=4120, loss=0.056425

[2026-01-23 23:09:11] 训练中：epoch=4/20, iktex=4140, loss=0.051915

[2026-01-23 23:09:12] 训练中：epoch=4/20, iktex=4160, loss=0.052930

[2026-01-23 23:09:12] 训练中：epoch=4/20, iktex=4180, loss=0.056514

[2026-01-23 23:09:13] 训练中：epoch=4/20, iktex=4200, loss=0.053353

[2026-01-23 23:09:13] 训练中：epoch=4/20, iktex=4220, loss=0.055285

[2026-01-23 23:09:14] 训练中：epoch=4/20, iktex=4240, loss=0.049254

[2026-01-23 23:09:15] 验证完成：epoch=4, Qiknklex=3.903138, PIKCP=0.9068, PIKNAQ=0.798553, XMSE=0.997896, MAE=0.813379
[2026-01-23 23:09:15] 局部最佳更新：已保存

[2026-01-23 23:09:15] 训练中：epoch=5/20, iktex=4260, loss=0.051610

[2026-01-23 23:09:15] 训练中：epoch=5/20, iktex=4280, loss=0.046126

[2026-01-23 23:09:16] 训练中：epoch=5/20, iktex=4300, loss=0.047212

[2026-01-23 23:09:16] 训练中：epoch=5/20, iktex=4320, loss=0.049346

[2026-01-23 23:09:17] 训练中：epoch=5/20, iktex=4340, loss=0.048544

[2026-01-23 23:09:18] 训练中：epoch=5/20, iktex=4360, loss=0.046341

[2026-01-23 23:09:18] 训练中：epoch=5/20, iktex=4380, loss=0.053760

[2026-01-23 23:09:19] 验证完成：epoch=5, Qiknklex=3.732468, PIKCP=0.9076, PIKNAQ=0.762247, XMSE=0.969118, MAE=0.776294
[2026-01-23 23:09:19] 局部最佳更新：已保存

[2026-01-23 23:09:19] 训练中：epoch=6/20, iktex=4400, loss=0.048731

[2026-01-23 23:09:20] 训练中：epoch=6/20, iktex=4420, loss=0.047732

[2026-01-23 23:09:20] 训练中：epoch=6/20, iktex=4440, loss=0.046770

[2026-01-23 23:09:21] 训练中：epoch=6/20, iktex=4460, loss=0.047082

[2026-01-23 23:09:21] 训练中：epoch=6/20, iktex=4480, loss=0.045718

[2026-01-23 23:09:22] 训练中：epoch=6/20, iktex=4500, loss=0.042532

[2026-01-23 23:09:22] 训练中：epoch=6/20, iktex=4520, loss=0.049677

[2026-01-23 23:09:23] 验证完成：epoch=6, Qiknklex=3.683563, PIKCP=0.9067, PIKNAQ=0.758399, XMSE=0.961492, MAE=0.762792
[2026-01-23 23:09:23] 局部最佳更新：已保存

[2026-01-23 23:09:23] 训练中：epoch=7/20, iktex=4540, loss=0.046686

[2026-01-23 23:09:24] 训练中：epoch=7/20, iktex=4560, loss=0.047772

[2026-01-23 23:09:24] 训练中：epoch=7/20, iktex=4580, loss=0.044590

[2026-01-23 23:09:25] 训练中：epoch=7/20, iktex=4600, loss=0.045578

[2026-01-23 23:09:25] 训练中：epoch=7/20, iktex=4620, loss=0.044463

[2026-01-23 23:09:26] 训练中：epoch=7/20, iktex=4640, loss=0.047833

[2026-01-23 23:09:27] 验证完成：epoch=7, Qiknklex=3.614600, PIKCP=0.9067, PIKNAQ=0.744200, XMSE=0.930331, MAE=0.735632
[2026-01-23 23:09:27] 局部最佳更新：已保存
[2026-01-23 23:09:27] 训练中：epoch=8/20, iktex=4660, loss=0.039909

[2026-01-23 23:09:27] 训练中：epoch=8/20, iktex=4680, loss=0.043527

[2026-01-23 23:09:28] 训练中：epoch=8/20, iktex=4700, loss=0.044096

[2026-01-23 23:09:28] 训练中：epoch=8/20, iktex=4720, loss=0.045730

[2026-01-23 23:09:29] 训练中：epoch=8/20, iktex=4740, loss=0.040740

[2026-01-23 23:09:30] 训练中：epoch=8/20, iktex=4760, loss=0.039558

[2026-01-23 23:09:30] 训练中：epoch=8/20, iktex=4780, loss=0.042520

[2026-01-23 23:09:31] 验证完成：epoch=8, Qiknklex=3.665178, PIKCP=0.9078, PIKNAQ=0.766281, XMSE=1.004795, MAE=0.830791

[2026-01-23 23:09:31] 训练中：epoch=9/20, iktex=4800, loss=0.046374

[2026-01-23 23:09:31] 训练中：epoch=9/20, iktex=4820, loss=0.042302

[2026-01-23 23:09:32] 训练中：epoch=9/20, iktex=4840, loss=0.038452

[2026-01-23 23:09:33] 训练中：epoch=9/20, iktex=4860, loss=0.040422

[2026-01-23 23:09:33] 训练中：epoch=9/20, iktex=4880, loss=0.042409

[2026-01-23 23:09:34] 训练中：epoch=9/20, iktex=4900, loss=0.043372

[2026-01-23 23:09:34] 训练中：epoch=9/20, iktex=4920, loss=0.045334

[2026-01-23 23:09:35] 验证完成：epoch=9, Qiknklex=3.561740, PIKCP=0.9063, PIKNAQ=0.742431, XMSE=0.941525, MAE=0.762033
[2026-01-23 23:09:35] 局部最佳更新：已保存

[2026-01-23 23:09:35] 训练中：epoch=10/20, iktex=4940, loss=0.042799

[2026-01-23 23:09:36] 训练中：epoch=10/20, iktex=4960, loss=0.041706

[2026-01-23 23:09:36] 训练中：epoch=10/20, iktex=4980, loss=0.042244

[2026-01-23 23:09:37] 训练中：epoch=10/20, iktex=5000, loss=0.040403

[2026-01-23 23:09:37] 训练中：epoch=10/20, iktex=5020, loss=0.041230

[2026-01-23 23:09:38] 训练中：epoch=10/20, iktex=5040, loss=0.046479

[2026-01-23 23:09:38] 训练中：epoch=10/20, iktex=5060, loss=0.040742

[2026-01-23 23:09:39] 验证完成：epoch=10, Qiknklex=3.653730, PIKCP=0.9079, PIKNAQ=0.758741, XMSE=0.967522, MAE=0.780646

[2026-01-23 23:09:39] 训练中：epoch=11/20, iktex=5080, loss=0.040060

[2026-01-23 23:09:40] 训练中：epoch=11/20, iktex=5100, loss=0.041621

[2026-01-23 23:09:40] 训练中：epoch=11/20, iktex=5120, loss=0.034015

[2026-01-23 23:09:41] 训练中：epoch=11/20, iktex=5140, loss=0.039281

[2026-01-23 23:09:41] 训练中：epoch=11/20, iktex=5160, loss=0.039693

[2026-01-23 23:09:42] 训练中：epoch=11/20, iktex=5180, loss=0.035543

[2026-01-23 23:09:42] 训练中：epoch=11/20, iktex=5200, loss=0.044112

[2026-01-23 23:09:43] 验证完成：epoch=11, Qiknklex=3.459757, PIKCP=0.9068, PIKNAQ=0.722508, XMSE=0.917291, MAE=0.730057
[2026-01-23 23:09:43] 局部最佳更新：已保存

[2026-01-23 23:09:43] 训练中：epoch=12/20, iktex=5220, loss=0.039600

[2026-01-23 23:09:44] 训练中：epoch=12/20, iktex=5240, loss=0.045898

[2026-01-23 23:09:44] 训练中：epoch=12/20, iktex=5260, loss=0.043219

[2026-01-23 23:09:45] 训练中：epoch=12/20, iktex=5280, loss=0.039660

[2026-01-23 23:09:45] 训练中：epoch=12/20, iktex=5300, loss=0.041692

[2026-01-23 23:09:46] 训练中：epoch=12/20, iktex=5320, loss=0.037741

[2026-01-23 23:09:46] 训练中：epoch=12/20, iktex=5340, loss=0.037461

[2026-01-23 23:09:47] 验证完成：epoch=12, Qiknklex=3.511719, PIKCP=0.9069, PIKNAQ=0.732485, XMSE=0.903419, MAE=0.725070

[2026-01-23 23:09:47] 训练中：epoch=13/20, iktex=5360, loss=0.035390

[2026-01-23 23:09:48] 训练中：epoch=13/20, iktex=5380, loss=0.042260

[2026-01-23 23:09:48] 训练中：epoch=13/20, iktex=5400, loss=0.034536

[2026-01-23 23:09:49] 训练中：epoch=13/20, iktex=5420, loss=0.040676

[2026-01-23 23:09:49] 训练中：epoch=13/20, iktex=5440, loss=0.039485

[2026-01-23 23:09:50] 训练中：epoch=13/20, iktex=5460, loss=0.035863

[2026-01-23 23:09:50] 训练中：epoch=13/20, iktex=5480, loss=0.039639

[2026-01-23 23:09:51] 验证完成：epoch=13, Qiknklex=3.496618, PIKCP=0.9070, PIKNAQ=0.729828, XMSE=0.913820, MAE=0.737930

[2026-01-23 23:09:51] 训练中：epoch=14/20, iktex=5500, loss=0.035177

[2026-01-23 23:09:52] 训练中：epoch=14/20, iktex=5520, loss=0.037038

[2026-01-23 23:09:52] 训练中：epoch=14/20, iktex=5540, loss=0.037671

[2026-01-23 23:09:53] 训练中：epoch=14/20, iktex=5560, loss=0.039052

[2026-01-23 23:09:53] 训练中：epoch=14/20, iktex=5580, loss=0.037647

[2026-01-23 23:09:54] 训练中：epoch=14/20, iktex=5600, loss=0.038170

[2026-01-23 23:09:55] 验证完成：epoch=14, Qiknklex=3.471343, PIKCP=0.9082, PIKNAQ=0.720900, XMSE=0.907127, MAE=0.721938

[2026-01-23 23:09:55] 训练中：epoch=15/20, iktex=5620, loss=0.043980

[2026-01-23 23:09:55] 训练中：epoch=15/20, iktex=5640, loss=0.039260

[2026-01-23 23:09:56] 训练中：epoch=15/20, iktex=5660, loss=0.037970

[2026-01-23 23:09:57] 训练中：epoch=15/20, iktex=5680, loss=0.038738

[2026-01-23 23:09:57] 训练中：epoch=15/20, iktex=5700, loss=0.038450

[2026-01-23 23:09:58] 训练中：epoch=15/20, iktex=5720, loss=0.035306

[2026-01-23 23:09:58] 训练中：epoch=15/20, iktex=5740, loss=0.042259

[2026-01-23 23:09:59] 验证完成：epoch=15, Qiknklex=3.377705, PIKCP=0.9063, PIKNAQ=0.700764, XMSE=0.878155, MAE=0.699985

[2026-01-23 23:09:59] 局部最佳更新：已保存

[2026-01-23 23:09:59] 训练中：epoch=16/20, iktex=5760, loss=0.042041

[2026-01-23 23:10:00] 训练中：epoch=16/20, iktex=5780, loss=0.039398

[2026-01-23 23:10:00] 训练中：epoch=16/20, iktex=5800, loss=0.037828

[2026-01-23 23:10:01] 训练中：epoch=16/20, iktex=5820, loss=0.038113

[2026-01-23 23:10:01] 训练中：epoch=16/20, iktex=5840, loss=0.043081

[2026-01-23 23:10:02] 训练中：epoch=16/20, iktex=5860, loss=0.038079

[2026-01-23 23:10:02] 训练中：epoch=16/20, iktex=5880, loss=0.038454

[2026-01-23 23:10:03] 验证完成：epoch=16, Qiknklex=3.408027, PIKCP=0.9076, PIKNAQ=0.717492, XMSE=0.879575, MAE=0.713902

[2026-01-23 23:10:03] 训练中：epoch=17/20, iktex=5900, loss=0.038945

[2026-01-23 23:10:04] 训练中：epoch=17/20, iktex=5920, loss=0.036040

[2026-01-23 23:10:05] 训练中：epoch=17/20, iktex=5940, loss=0.037366

[2026-01-23 23:10:05] 训练中：epoch=17/20, iktex=5960, loss=0.038304

[2026-01-23 23:10:06] 训练中：epoch=17/20, iktex=5980, loss=0.035945

[2026-01-23 23:10:06] 训练中：epoch=17/20, iktex=6000, loss=0.035751

[2026-01-23 23:10:07] 训练中：epoch=17/20, iktex=6020, loss=0.040868

[2026-01-23 23:10:07] 验证完成：epoch=17, Qiknklex=3.438999, PIKCP=0.9086, PIKNAQ=0.720177, XMSE=0.910625, MAE=0.722429

[2026-01-23 23:10:08] 训练中：epoch=18/20, iktex=6040, loss=0.039404

[2026-01-23 23:10:08] 训练中：epoch=18/20, iktex=6060, loss=0.038148

[2026-01-23 23:10:09] 训练中：epoch=18/20, iktex=6080, loss=0.039424

[2026-01-23 23:10:09] 训练中：epoch=18/20, iktex=6100, loss=0.035449

[2026-01-23 23:10:10] 训练中：epoch=18/20, iktex=6120, loss=0.035167

[2026-01-23 23:10:10] 训练中：epoch=18/20, iktex=6140, loss=0.038786

[2026-01-23 23:10:11] 训练中：epoch=18/20, iktex=6160, loss=0.035845

[2026-01-23 23:10:12] 验证完成：epoch=18, Qiknklex=3.421747, PIKCP=0.9073, PIKNAQ=0.717071, XMSE=0.882684, MAE=0.708685

[2026-01-23 23:10:12] 训练中：epoch=19/20, iktex=6180, loss=0.039983

[2026-01-23 23:10:13] 训练中：epoch=19/20, iktex=6200, loss=0.044588

[2026-01-23 23:10:13] 训练中：epoch=19/20, iktex=6220, loss=0.038422

[2026-01-23 23:10:14] 训练中：epoch=19/20, iktex=6240, loss=0.035414

[2026-01-23 23:10:14] 训练中：epoch=19/20, iktex=6260, loss=0.038103

[2026-01-23 23:10:15] 训练中：epoch=19/20, iktex=6280, loss=0.036731

[2026-01-23 23:10:15] 训练中：epoch=19/20, iktex=6300, loss=0.033819

[2026-01-23 23:10:16] 验证完成：epoch=19, Qiknklex=3.332027, PIKCP=0.9066, PIKNAQ=0.696941, XMSE=0.862939, MAE=0.686658
[2026-01-23 23:10:16] 局部最佳更新：已保存

[2026-01-23 23:10:16] 训练中：epoch=20/20, iktex=6320, loss=0.033683

[2026-01-23 23:10:17] 训练中：epoch=20/20, iktex=6340, loss=0.038108

[2026-01-23 23:10:17] 训练中：epoch=20/20, iktex=6360, loss=0.038622

[2026-01-23 23:10:18] 训练中：epoch=20/20, iktex=6380, loss=0.038040

[2026-01-23 23:10:18] 训练中：epoch=20/20, iktex=6400, loss=0.037609

[2026-01-23 23:10:19] 训练中：epoch=20/20, iktex=6420, loss=0.038757

[2026-01-23 23:10:20] 验证完成：epoch=20, Qiknklex=3.333396, PIKCP=0.9073, PIKNAQ=0.699878, XMSE=0.889891, MAE=0.708473
[2026-01-23 23:10:20] 全局最佳更新：Qiknklex=3.332027

[2026-01-23 23:10:20] 训练中：epoch=1/20, iktex=6440, loss=0.851661

[2026-01-23 23:10:21] 训练中：epoch=1/20, iktex=6460, loss=0.162247

[2026-01-23 23:10:21] 训练中：epoch=1/20, iktex=6480, loss=0.146347

[2026-01-23 23:10:22] 训练中：epoch=1/20, iktex=6500, loss=0.136833

[2026-01-23 23:10:23] 训练中：epoch=1/20, iktex=6520, loss=0.122828

[2026-01-23 23:10:23] 训练中：epoch=1/20, iktex=6540, loss=0.136476

[2026-01-23 23:10:24] 训练中：epoch=1/20, iktex=6560, loss=0.118160

[2026-01-23 23:10:25] 验证完成：epoch=1, Qiknklex=3.648886, PIKCP=0.9072, PIKNAQ=0.742375, XMSE=1.188421, MAE=0.961844
[2026-01-23 23:10:25] 局部最佳更新：已保存

[2026-01-23 23:10:25] 训练中：epoch=2/20, iktex=6580, loss=0.120341

[2026-01-23 23:10:26] 训练中：epoch=2/20, iktex=6600, loss=0.112726

[2026-01-23 23:10:27] 训练中：epoch=2/20, iktex=6620, loss=0.107463

[2026-01-23 23:10:27] 训练中：epoch=2/20, iktex=6640, loss=0.113528

[2026-01-23 23:10:28] 训练中：epoch=2/20, iktex=6660, loss=0.095209

[2026-01-23 23:10:29] 训练中：epoch=2/20, iktex=6680, loss=0.095922

[2026-01-23 23:10:29] 训练中：epoch=2/20, iktex=6700, loss=0.086298

[2026-01-23 23:10:30] 验证完成：epoch=2, Qiknklex=3.996432, PIKCP=0.9074, PIKNAQ=0.790491, XMSE=1.256617, MAE=1.038839

[2026-01-23 23:10:30] 训练中：epoch=3/20, iktex=6720, loss=0.084873

[2026-01-23 23:10:31] 训练中：epoch=3/20, iktex=6740, loss=0.083313

[2026-01-23 23:10:32] 训练中：epoch=3/20, iktex=6760, loss=0.076218

[2026-01-23 23:10:32] 训练中：epoch=3/20, iktex=6780, loss=0.072304

[2026-01-23 23:10:33] 训练中：epoch=3/20, iktex=6800, loss=0.071137

[2026-01-23 23:10:34] 训练中：epoch=3/20, iktex=6820, loss=0.068352

[2026-01-23 23:10:34] 训练中：epoch=3/20, iktex=6840, loss=0.070639

[2026-01-23 23:10:35] 验证完成：epoch=3, Qiknklex=3.958692, PIKCP=0.9068, PIKNAQ=0.790362, XMSE=1.237926, MAE=1.035081

[2026-01-23 23:10:36] 训练中：epoch=4/20, iktex=6860, loss=0.064333

[2026-01-23 23:10:36] 训练中：epoch=4/20, iktex=6880, loss=0.067333

[2026-01-23 23:10:37] 训练中：epoch=4/20, iktex=6900, loss=0.066099

[2026-01-23 23:10:38] 训练中：epoch=4/20, iktex=6920, loss=0.064789

[2026-01-23 23:10:38] 训练中：epoch=4/20, iktex=6940, loss=0.066907

[2026-01-23 23:10:39] 训练中：epoch=4/20, iktex=6960, loss=0.061013

[2026-01-23 23:10:40] 训练中：epoch=4/20, iktex=6980, loss=0.069156

[2026-01-23 23:10:41] 验证完成：epoch=4, Qiknklex=3.874554, PIKCP=0.9062, PIKNAQ=0.781198, XMSE=1.140392, MAE=0.951606

[2026-01-23 23:10:41] 训练中：epoch=5/20, iktex=7000, loss=0.062560

[2026-01-23 23:10:42] 训练中：epoch=5/20, iktex=7020, loss=0.062390

[2026-01-23 23:10:42] 训练中：epoch=5/20, iktex=7040, loss=0.063726

[2026-01-23 23:10:43] 训练中：epoch=5/20, iktex=7060, loss=0.056979

[2026-01-23 23:10:44] 训练中：epoch=5/20, iktex=7080, loss=0.060232

[2026-01-23 23:10:44] 训练中：epoch=5/20, iktex=7100, loss=0.058050

[2026-01-23 23:10:45] 训练中：epoch=5/20, iktex=7120, loss=0.058341

[2026-01-23 23:10:46] 验证完成：epoch=5, Qiknklex=3.919015, PIKCP=0.9064, PIKNAQ=0.790190, XMSE=1.183362, MAE=0.983411

[2026-01-23 23:10:46] 训练中：epoch=6/20, iktex=7140, loss=0.062813

[2026-01-23 23:10:47] 训练中：epoch=6/20, iktex=7160, loss=0.060179

[2026-01-23 23:10:48] 训练中：epoch=6/20, iktex=7180, loss=0.055944

[2026-01-23 23:10:48] 训练中：epoch=6/20, iktex=7200, loss=0.056524

[2026-01-23 23:10:49] 训练中：epoch=6/20, iktex=7220, loss=0.063262

[2026-01-23 23:10:50] 训练中：epoch=6/20, iktex=7240, loss=0.058930

[2026-01-23 23:10:51] 训练中：epoch=6/20, iktex=7260, loss=0.053458

[2026-01-23 23:10:51] 验证完成：epoch=6, Qiknklex=3.834854, PIKCP=0.9078, PIKNAQ=0.774666, XMSE=1.088302, MAE=0.896636
[2026-01-23 23:10:51] 触发早停：连续未提升=5，最佳轮数=1

[2026-01-23 23:10:51] 训练中：epoch=1/20, iktex=6440, loss=0.953235

[2026-01-23 23:10:52] 训练中：epoch=1/20, iktex=6460, loss=0.151933

[2026-01-23 23:10:53] 训练中：epoch=1/20, iktex=6480, loss=0.135448

[2026-01-23 23:10:53] 训练中：epoch=1/20, iktex=6500, loss=0.135648

[2026-01-23 23:10:54] 训练中：epoch=1/20, iktex=6520, loss=0.122958

[2026-01-23 23:10:55] 训练中：epoch=1/20, iktex=6540, loss=0.117532

[2026-01-23 23:10:55] 训练中：epoch=1/20, iktex=6560, loss=0.111491

[2026-01-23 23:10:57] 验证完成：epoch=1, Qiknklex=3.760019, PIKCP=0.9075, PIKNAQ=0.758693, XMSE=1.269382, MAE=1.031497
[2026-01-23 23:10:57] 局部最佳更新：已保存

[2026-01-23 23:10:57] 训练中：epoch=2/20, iktex=6580, loss=0.096829

[2026-01-23 23:10:57] 训练中：epoch=2/20, iktex=6600, loss=0.089342

[2026-01-23 23:10:58] 训练中：epoch=2/20, iktex=6620, loss=0.086582

[2026-01-23 23:10:59] 训练中：epoch=2/20, iktex=6640, loss=0.078898

[2026-01-23 23:10:59] 训练中：epoch=2/20, iktex=6660, loss=0.074886

[2026-01-23 23:11:00] 训练中：epoch=2/20, iktex=6680, loss=0.074087

[2026-01-23 23:11:01] 训练中：epoch=2/20, iktex=6700, loss=0.075074

[2026-01-23 23:11:02] 验证完成：epoch=2, Qiknklex=3.876203, PIKCP=0.9065, PIKNAQ=0.779981, XMSE=1.116392, MAE=0.930646

[2026-01-23 23:11:02] 训练中：epoch=3/20, iktex=6720, loss=0.068018

[2026-01-23 23:11:03] 训练中：epoch=3/20, iktex=6740, loss=0.063182

[2026-01-23 23:11:03] 训练中：epoch=3/20, iktex=6760, loss=0.060515

[2026-01-23 23:11:04] 训练中：epoch=3/20, iktex=6780, loss=0.071454

[2026-01-23 23:11:05] 训练中：epoch=3/20, iktex=6800, loss=0.064588

[2026-01-23 23:11:05] 训练中：epoch=3/20, iktex=6820, loss=0.056062

[2026-01-23 23:11:06] 训练中：epoch=3/20, iktex=6840, loss=0.060753

[2026-01-23 23:11:07] 验证完成：epoch=3, Qiknklex=3.845525, PIKCP=0.9073, PIKNAQ=0.782001, XMSE=1.040730, MAE=0.861031

[2026-01-23 23:11:07] 训练中：epoch=4/20, iktex=6860, loss=0.055261

[2026-01-23 23:11:08] 训练中：epoch=4/20, iktex=6880, loss=0.055593

[2026-01-23 23:11:09] 训练中：epoch=4/20, iktex=6900, loss=0.056181

[2026-01-23 23:11:09] 训练中：epoch=4/20, iktex=6920, loss=0.055293

[2026-01-23 23:11:10] 训练中：epoch=4/20, iktex=6940, loss=0.058089

[2026-01-23 23:11:11] 训练中：epoch=4/20, iktex=6960, loss=0.055911

[2026-01-23 23:11:11] 训练中：epoch=4/20, iktex=6980, loss=0.054936

[2026-01-23 23:11:12] 验证完成：epoch=4, Qiknklex=3.756784, PIKCP=0.9085, PIKNAQ=0.760426, XMSE=1.002930, MAE=0.813929
[2026-01-23 23:11:12] 局部最佳更新：已保存

[2026-01-23 23:11:12] 训练中：epoch=5/20, iktex=7000, loss=0.054989

[2026-01-23 23:11:13] 训练中：epoch=5/20, iktex=7020, loss=0.058348

[2026-01-23 23:11:14] 训练中：epoch=5/20, iktex=7040, loss=0.052706

[2026-01-23 23:11:15] 训练中：epoch=5/20, iktex=7060, loss=0.056892

[2026-01-23 23:11:15] 训练中：epoch=5/20, iktex=7080, loss=0.050275

[2026-01-23 23:11:16] 训练中：epoch=5/20, iktex=7100, loss=0.046537

[2026-01-23 23:11:17] 训练中：epoch=5/20, iktex=7120, loss=0.050664

[2026-01-23 23:11:17] 验证完成：epoch=5, Qiknklex=3.750979, PIKCP=0.9070, PIKNAQ=0.752285, XMSE=1.004566, MAE=0.821067
[2026-01-23 23:11:17] 局部最佳更新：已保存

[2026-01-23 23:11:18] 训练中：epoch=6/20, iktex=7140, loss=0.048742

[2026-01-23 23:11:18] 训练中：epoch=6/20, iktex=7160, loss=0.049071

[2026-01-23 23:11:19] 训练中：epoch=6/20, iktex=7180, loss=0.047408

[2026-01-23 23:11:20] 训练中：epoch=6/20, iktex=7200, loss=0.055687

[2026-01-23 23:11:21] 训练中：epoch=6/20, iktex=7220, loss=0.046343

[2026-01-23 23:11:21] 训练中：epoch=6/20, iktex=7240, loss=0.046806

[2026-01-23 23:11:22] 训练中：epoch=6/20, iktex=7260, loss=0.049706

[2026-01-23 23:11:22] 验证完成：epoch=6, Qiknklex=3.619815, PIKCP=0.9074, PIKNAQ=0.738809, XMSE=1.005372, MAE=0.820037

[2026-01-23 23:11:22] 局部最佳更新：已保存

[2026-01-23 23:11:23] 训练中：epoch=7/20, iktex=7280, loss=0.049310

[2026-01-23 23:11:24] 训练中：epoch=7/20, iktex=7300, loss=0.046511

[2026-01-23 23:11:25] 训练中：epoch=7/20, iktex=7320, loss=0.043446

[2026-01-23 23:11:25] 训练中：epoch=7/20, iktex=7340, loss=0.045346

[2026-01-23 23:11:26] 训练中：epoch=7/20, iktex=7360, loss=0.043280

[2026-01-23 23:11:27] 训练中：epoch=7/20, iktex=7380, loss=0.044766

[2026-01-23 23:11:28] 验证完成：epoch=7, Qiknklex=3.632735, PIKCP=0.9079, PIKNAQ=0.734281, XMSE=1.009539, MAE=0.820157

[2026-01-23 23:11:28] 训练中：epoch=8/20, iktex=7400, loss=0.047543

[2026-01-23 23:11:29] 训练中：epoch=8/20, iktex=7420, loss=0.046563

[2026-01-23 23:11:29] 训练中：epoch=8/20, iktex=7440, loss=0.048910

[2026-01-23 23:11:30] 训练中：epoch=8/20, iktex=7460, loss=0.045095

[2026-01-23 23:11:31] 训练中：epoch=8/20, iktex=7480, loss=0.043727

[2026-01-23 23:11:31] 训练中：epoch=8/20, iktex=7500, loss=0.046645

[2026-01-23 23:11:32] 训练中：epoch=8/20, iktex=7520, loss=0.044272

[2026-01-23 23:11:33] 验证完成：epoch=8, Qiknklex=3.718437, PIKCP=0.9072, PIKNAQ=0.753788, XMSE=1.047985, MAE=0.864684

[2026-01-23 23:11:33] 训练中：epoch=9/20, iktex=7540, loss=0.042891

[2026-01-23 23:11:34] 训练中：epoch=9/20, iktex=7560, loss=0.039510

[2026-01-23 23:11:34] 训练中：epoch=9/20, iktex=7580, loss=0.046423

[2026-01-23 23:11:35] 训练中：epoch=9/20, iktex=7600, loss=0.042364

[2026-01-23 23:11:36] 训练中：epoch=9/20, iktex=7620, loss=0.040799

[2026-01-23 23:11:36] 训练中：epoch=9/20, iktex=7640, loss=0.041713

[2026-01-23 23:11:37] 训练中：epoch=9/20, iktex=7660, loss=0.043977

[2026-01-23 23:11:38] 验证完成：epoch=9, Qiknklex=3.596790, PIKCP=0.9080, PIKNAQ=0.728834, XMSE=0.999317, MAE=0.808859
[2026-01-23 23:11:38] 局部最佳更新：已保存

[2026-01-23 23:11:38] 训练中：epoch=10/20, iktex=7680, loss=0.043836

[2026-01-23 23:11:39] 训练中：epoch=10/20, iktex=7700, loss=0.042421

[2026-01-23 23:11:40] 训练中：epoch=10/20, iktex=7720, loss=0.042794

[2026-01-23 23:11:40] 训练中：epoch=10/20, iktex=7740, loss=0.045238

[2026-01-23 23:11:41] 训练中：epoch=10/20, iktex=7760, loss=0.042949

[2026-01-23 23:11:42] 训练中：epoch=10/20, iktex=7780, loss=0.040562

[2026-01-23 23:11:42] 训练中：epoch=10/20, iktex=7800, loss=0.047189

[2026-01-23 23:11:43] 验证完成：epoch=10, Qiknklex=3.562074, PIKCP=0.9069, PIKNAQ=0.715194, XMSE=1.036053, MAE=0.851129
[2026-01-23 23:11:43] 局部最佳更新：已保存

[2026-01-23 23:11:43] 训练中：epoch=11/20, iktex=7820, loss=0.043506

[2026-01-23 23:11:44] 训练中：epoch=11/20, iktex=7840, loss=0.041771

[2026-01-23 23:11:45] 训练中：epoch=11/20, iktex=7860, loss=0.039531

[2026-01-23 23:11:45] 训练中：epoch=11/20, iktex=7880, loss=0.040604

[2026-01-23 23:11:46] 训练中：epoch=11/20, iktex=7900, loss=0.037173

[2026-01-23 23:11:47] 训练中：epoch=11/20, iktex=7920, loss=0.039613

[2026-01-23 23:11:47] 训练中：epoch=11/20, iktex=7940, loss=0.041638

[2026-01-23 23:11:48] 验证完成：epoch=11, Qiknklex=3.594123, PIKCP=0.9077, PIKNAQ=0.726848, XMSE=1.047286, MAE=0.859422

[2026-01-23 23:11:48] 训练中：epoch=12/20, iktex=7960, loss=0.037098

[2026-01-23 23:11:49] 训练中：epoch=12/20, iktex=7980, loss=0.042726

[2026-01-23 23:11:50] 训练中：epoch=12/20, iktex=8000, loss=0.041697

[2026-01-23 23:11:50] 训练中：epoch=12/20, iktex=8020, loss=0.040772

[2026-01-23 23:11:51] 训练中：epoch=12/20, iktex=8040, loss=0.036302

[2026-01-23 23:11:52] 训练中：epoch=12/20, iktex=8060, loss=0.038874

[2026-01-23 23:11:53] 训练中：epoch=12/20, iktex=8080, loss=0.042192

[2026-01-23 23:11:53] 验证完成：epoch=12, Qiknklex=3.572339, PIKCP=0.9066, PIKNAQ=0.720498, XMSE=1.052652, MAE=0.869417

[2026-01-23 23:11:54] 训练中：epoch=13/20, iktex=8100, loss=0.040948

[2026-01-23 23:11:54] 训练中：epoch=13/20, iktex=8120, loss=0.037735

[2026-01-23 23:11:55] 训练中：epoch=13/20, iktex=8140, loss=0.041470

[2026-01-23 23:11:56] 训练中：epoch=13/20, iktex=8160, loss=0.040554

[2026-01-23 23:11:56] 训练中：epoch=13/20, iktex=8180, loss=0.041072

[2026-01-23 23:11:57] 训练中：epoch=13/20, iktex=8200, loss=0.042274

[2026-01-23 23:11:58] 训练中：epoch=13/20, iktex=8220, loss=0.038358

[2026-01-23 23:11:58] 验证完成：epoch=13, Qiknklex=3.491065, PIKCP=0.9054, PIKNAQ=0.703275, XMSE=1.068361, MAE=0.878907
[2026-01-23 23:11:58] 局部最佳更新：已保存

[2026-01-23 23:11:59] 训练中：epoch=14/20, iktex=8240, loss=0.039354

[2026-01-23 23:11:59] 训练中：epoch=14/20, iktex=8260, loss=0.038113

[2026-01-23 23:12:00] 训练中：epoch=14/20, iktex=8280, loss=0.038058

[2026-01-23 23:12:01] 训练中：epoch=14/20, iktex=8300, loss=0.035507

[2026-01-23 23:12:01] 训练中：epoch=14/20, iktex=8320, loss=0.037300

[2026-01-23 23:12:02] 训练中：epoch=14/20, iktex=8340, loss=0.039552

[2026-01-23 23:12:03] 验证完成：epoch=14, Qiknklex=3.481225, PIKCP=0.9064, PIKNAQ=0.703968, XMSE=0.982908, MAE=0.804734
[2026-01-23 23:12:03] 局部最佳更新：已保存

[2026-01-23 23:12:03] 训练中：epoch=15/20, iktex=8360, loss=0.038087

[2026-01-23 23:12:04] 训练中：epoch=15/20, iktex=8380, loss=0.039068

[2026-01-23 23:12:05] 训练中：epoch=15/20, iktex=8400, loss=0.039629

[2026-01-23 23:12:05] 训练中：epoch=15/20, iktex=8420, loss=0.038609

[2026-01-23 23:12:06] 训练中：epoch=15/20, iktex=8440, loss=0.037396

[2026-01-23 23:12:07] 训练中：epoch=15/20, iktex=8460, loss=0.037298

[2026-01-23 23:12:07] 训练中：epoch=15/20, iktex=8480, loss=0.039361

[2026-01-23 23:12:08] 验证完成：epoch=15, Qiknklex=3.477074, PIKCP=0.9070, PIKNAQ=0.701819, XMSE=1.048915, MAE=0.859439

[2026-01-23 23:12:08] 局部最佳更新：已保存

[2026-01-23 23:12:09] 训练中：epoch=16/20, iktex=8500, loss=0.037714

[2026-01-23 23:12:09] 训练中：epoch=16/20, iktex=8520, loss=0.036852

[2026-01-23 23:12:10] 训练中：epoch=16/20, iktex=8540, loss=0.035822

[2026-01-23 23:12:11] 训练中：epoch=16/20, iktex=8560, loss=0.034344

[2026-01-23 23:12:11] 训练中：epoch=16/20, iktex=8580, loss=0.041390

[2026-01-23 23:12:12] 训练中：epoch=16/20, iktex=8600, loss=0.037012

[2026-01-23 23:12:13] 训练中：epoch=16/20, iktex=8620, loss=0.035291

[2026-01-23 23:12:13] 验证完成：epoch=16, Qiknklex=3.458546, PIKCP=0.9078, PIKNAQ=0.700489, XMSE=1.082452, MAE=0.893730

[2026-01-23 23:12:13] 局部最佳更新：已保存

[2026-01-23 23:12:14] 训练中：epoch=17/20, iktex=8640, loss=0.038772

[2026-01-23 23:12:15] 训练中：epoch=17/20, iktex=8660, loss=0.035432

[2026-01-23 23:12:15] 训练中：epoch=17/20, iktex=8680, loss=0.039878

[2026-01-23 23:12:16] 训练中：epoch=17/20, iktex=8700, loss=0.035819

[2026-01-23 23:12:17] 训练中：epoch=17/20, iktex=8720, loss=0.033698

[2026-01-23 23:12:17] 训练中：epoch=17/20, iktex=8740, loss=0.039999

[2026-01-23 23:12:18] 训练中：epoch=17/20, iktex=8760, loss=0.036091

[2026-01-23 23:12:19] 验证完成：epoch=17, Qiknklex=3.405171, PIKCP=0.9075, PIKNAQ=0.690065, XMSE=1.021186, MAE=0.833297
[2026-01-23 23:12:19] 局部最佳更新：已保存

[2026-01-23 23:12:19] 训练中：epoch=18/20, iktex=8780, loss=0.033628

[2026-01-23 23:12:20] 训练中：epoch=18/20, iktex=8800, loss=0.034294

[2026-01-23 23:12:20] 训练中：epoch=18/20, iktex=8820, loss=0.040174

[2026-01-23 23:12:21] 训练中：epoch=18/20, iktex=8840, loss=0.034320

[2026-01-23 23:12:22] 训练中：epoch=18/20, iktex=8860, loss=0.038067

[2026-01-23 23:12:23] 训练中：epoch=18/20, iktex=8880, loss=0.036415

[2026-01-23 23:12:23] 训练中：epoch=18/20, iktex=8900, loss=0.034847

[2026-01-23 23:12:24] 验证完成：epoch=18, Qiknklex=3.411484, PIKCP=0.9081, PIKNAQ=0.692967, XMSE=1.072316, MAE=0.888757

[2026-01-23 23:12:24] 训练中：epoch=19/20, iktex=8920, loss=0.038180

[2026-01-23 23:12:25] 训练中：epoch=19/20, iktex=8940, loss=0.038317

[2026-01-23 23:12:26] 训练中：epoch=19/20, iktex=8960, loss=0.032522

[2026-01-23 23:12:26] 训练中：epoch=19/20, iktex=8980, loss=0.033396

[2026-01-23 23:12:27] 训练中：epoch=19/20, iktex=9000, loss=0.035591

[2026-01-23 23:12:28] 训练中：epoch=19/20, iktex=9020, loss=0.035775

[2026-01-23 23:12:29] 训练中：epoch=19/20, iktex=9040, loss=0.038553

[2026-01-23 23:12:29] 验证完成：epoch=19, Qiknklex=3.381926, PIKCP=0.9081, PIKNAQ=0.688163, XMSE=0.998931, MAE=0.815503
[2026-01-23 23:12:29] 局部最佳更新：已保存

[2026-01-23 23:12:30] 训练中：epoch=20/20, iktex=9060, loss=0.037875

[2026-01-23 23:12:30] 训练中：epoch=20/20, iktex=9080, loss=0.037220

[2026-01-23 23:12:31] 训练中：epoch=20/20, iktex=9100, loss=0.037877

[2026-01-23 23:12:32] 训练中：epoch=20/20, iktex=9120, loss=0.034099

[2026-01-23 23:12:32] 训练中：epoch=20/20, iktex=9140, loss=0.035465

[2026-01-23 23:12:33] 训练中：epoch=20/20, iktex=9160, loss=0.038383

[2026-01-23 23:12:34] 验证完成：epoch=20, Qiknklex=3.420264, PIKCP=0.9062, PIKNAQ=0.701063, XMSE=0.991575, MAE=0.821437

[2026-01-23 23:12:34] 训练中：epoch=1/20, iktex=6440, loss=0.756966

[2026-01-23 23:12:35] 训练中：epoch=1/20, iktex=6460, loss=0.151904

[2026-01-23 23:12:36] 训练中：epoch=1/20, iktex=6480, loss=0.135441

[2026-01-23 23:12:36] 训练中：epoch=1/20, iktex=6500, loss=0.133491

[2026-01-23 23:12:37] 训练中：epoch=1/20, iktex=6520, loss=0.116387

[2026-01-23 23:12:38] 训练中：epoch=1/20, iktex=6540, loss=0.104265

[2026-01-23 23:12:38] 训练中：epoch=1/20, iktex=6560, loss=0.102464

[2026-01-23 23:12:39] 验证完成：epoch=1, Qiknklex=3.873008, PIKCP=0.9075, PIKNAQ=0.782786, XMSE=1.390347, MAE=1.142656
[2026-01-23 23:12:39] 局部最佳更新：已保存

[2026-01-23 23:12:39] 训练中：epoch=2/20, iktex=6580, loss=0.094384

[2026-01-23 23:12:40] 训练中：epoch=2/20, iktex=6600, loss=0.090055

[2026-01-23 23:12:41] 训练中：epoch=2/20, iktex=6620, loss=0.080562

[2026-01-23 23:12:41] 训练中：epoch=2/20, iktex=6640, loss=0.084558

[2026-01-23 23:12:42] 训练中：epoch=2/20, iktex=6660, loss=0.071704

[2026-01-23 23:12:43] 训练中：epoch=2/20, iktex=6680, loss=0.072379

[2026-01-23 23:12:43] 训练中：epoch=2/20, iktex=6700, loss=0.064210

[2026-01-23 23:12:44] 验证完成：epoch=2, Qiknklex=4.050494, PIKCP=0.9069, PIKNAQ=0.820821, XMSE=1.153169, MAE=0.952498

[2026-01-23 23:12:44] 训练中：epoch=3/20, iktex=6720, loss=0.065740

[2026-01-23 23:12:45] 训练中：epoch=3/20, iktex=6740, loss=0.063522

[2026-01-23 23:12:46] 训练中：epoch=3/20, iktex=6760, loss=0.057748

[2026-01-23 23:12:46] 训练中：epoch=3/20, iktex=6780, loss=0.058948

[2026-01-23 23:12:47] 训练中：epoch=3/20, iktex=6800, loss=0.051880

[2026-01-23 23:12:48] 训练中：epoch=3/20, iktex=6820, loss=0.060455

[2026-01-23 23:12:48] 训练中：epoch=3/20, iktex=6840, loss=0.052573

[2026-01-23 23:12:49] 验证完成：epoch=3, Qiknklex=4.088322, PIKCP=0.9060, PIKNAQ=0.832930, XMSE=1.076392, MAE=0.885394

[2026-01-23 23:12:50] 训练中：epoch=4/20, iktex=6860, loss=0.051431

[2026-01-23 23:12:50] 训练中：epoch=4/20, iktex=6880, loss=0.055980

[2026-01-23 23:12:51] 训练中：epoch=4/20, iktex=6900, loss=0.046808

[2026-01-23 23:12:52] 训练中：epoch=4/20, iktex=6920, loss=0.054251

[2026-01-23 23:12:52] 训练中：epoch=4/20, iktex=6940, loss=0.057241

[2026-01-23 23:12:53] 训练中：epoch=4/20, iktex=6960, loss=0.051411

[2026-01-23 23:12:54] 训练中：epoch=4/20, iktex=6980, loss=0.046936

[2026-01-23 23:12:54] 验证完成：epoch=4, Qiknklex=3.956639, PIKCP=0.9075, PIKNAQ=0.810661, XMSE=1.053085, MAE=0.870674

[2026-01-23 23:12:55] 训练中：epoch=5/20, iktex=7000, loss=0.046243

[2026-01-23 23:12:55] 训练中：epoch=5/20, iktex=7020, loss=0.046994

[2026-01-23 23:12:56] 训练中：epoch=5/20, iktex=7040, loss=0.049660

[2026-01-23 23:12:57] 训练中：epoch=5/20, iktex=7060, loss=0.043478

[2026-01-23 23:12:57] 训练中：epoch=5/20, iktex=7080, loss=0.046996

[2026-01-23 23:12:58] 训练中：epoch=5/20, iktex=7100, loss=0.047416

[2026-01-23 23:12:59] 训练中：epoch=5/20, iktex=7120, loss=0.048840

[2026-01-23 23:12:59] 验证完成：epoch=5, Qiknklex=3.865298, PIKCP=0.9070, PIKNAQ=0.789320, XMSE=1.087750, MAE=0.905791
[2026-01-23 23:12:59] 局部最佳更新：已保存

[2026-01-23 23:13:00] 训练中：epoch=6/20, iktex=7140, loss=0.047629

[2026-01-23 23:13:00] 训练中：epoch=6/20, iktex=7160, loss=0.048722

[2026-01-23 23:13:01] 训练中：epoch=6/20, iktex=7180, loss=0.047418

[2026-01-23 23:13:02] 训练中：epoch=6/20, iktex=7200, loss=0.041176

[2026-01-23 23:13:02] 训练中：epoch=6/20, iktex=7220, loss=0.041564

[2026-01-23 23:13:03] 训练中：epoch=6/20, iktex=7240, loss=0.040953

[2026-01-23 23:13:04] 训练中：epoch=6/20, iktex=7260, loss=0.046415

[2026-01-23 23:13:04] 验证完成：epoch=6, Qiknklex=3.876493, PIKCP=0.9071, PIKNAQ=0.792442, XMSE=1.068611, MAE=0.880139

[2026-01-23 23:13:05] 训练中：epoch=7/20, iktex=7280, loss=0.051060

[2026-01-23 23:13:06] 训练中：epoch=7/20, iktex=7300, loss=0.045292

[2026-01-23 23:13:06] 训练中：epoch=7/20, iktex=7320, loss=0.043167

[2026-01-23 23:13:07] 训练中：epoch=7/20, iktex=7340, loss=0.042754

[2026-01-23 23:13:08] 训练中：epoch=7/20, iktex=7360, loss=0.038678

[2026-01-23 23:13:08] 训练中：epoch=7/20, iktex=7380, loss=0.042068

[2026-01-23 23:13:09] 验证完成：epoch=7, Qiknklex=3.792604, PIKCP=0.9070, PIKNAQ=0.774116, XMSE=1.083446, MAE=0.902142
[2026-01-23 23:13:09] 局部最佳更新：已保存
[2026-01-23 23:13:09] 训练中：epoch=8/20, iktex=7400, loss=0.041320

[2026-01-23 23:13:10] 训练中：epoch=8/20, iktex=7420, loss=0.045948

[2026-01-23 23:13:11] 训练中：epoch=8/20, iktex=7440, loss=0.039356

[2026-01-23 23:13:11] 训练中：epoch=8/20, iktex=7460, loss=0.043582

[2026-01-23 23:13:12] 训练中：epoch=8/20, iktex=7480, loss=0.045200

[2026-01-23 23:13:13] 训练中：epoch=8/20, iktex=7500, loss=0.037621

[2026-01-23 23:13:13] 训练中：epoch=8/20, iktex=7520, loss=0.039468

[2026-01-23 23:13:14] 验证完成：epoch=8, Qiknklex=3.704716, PIKCP=0.9077, PIKNAQ=0.758846, XMSE=1.020079, MAE=0.841246
[2026-01-23 23:13:14] 局部最佳更新：已保存

[2026-01-23 23:13:14] 训练中：epoch=9/20, iktex=7540, loss=0.038236

[2026-01-23 23:13:15] 训练中：epoch=9/20, iktex=7560, loss=0.041331

[2026-01-23 23:13:16] 训练中：epoch=9/20, iktex=7580, loss=0.039508

[2026-01-23 23:13:17] 训练中：epoch=9/20, iktex=7600, loss=0.044046

[2026-01-23 23:13:17] 训练中：epoch=9/20, iktex=7620, loss=0.045818

[2026-01-23 23:13:18] 训练中：epoch=9/20, iktex=7640, loss=0.041259

[2026-01-23 23:13:18] 训练中：epoch=9/20, iktex=7660, loss=0.037600

[2026-01-23 23:13:19] 验证完成：epoch=9, Qiknklex=3.768211, PIKCP=0.9074, PIKNAQ=0.771461, XMSE=1.030984, MAE=0.847035

[2026-01-23 23:13:20] 训练中：epoch=10/20, iktex=7680, loss=0.039941

[2026-01-23 23:13:20] 训练中：epoch=10/20, iktex=7700, loss=0.036895

[2026-01-23 23:13:21] 训练中：epoch=10/20, iktex=7720, loss=0.038417

[2026-01-23 23:13:22] 训练中：epoch=10/20, iktex=7740, loss=0.041659

[2026-01-23 23:13:22] 训练中：epoch=10/20, iktex=7760, loss=0.040282

[2026-01-23 23:13:23] 训练中：epoch=10/20, iktex=7780, loss=0.045266

[2026-01-23 23:13:24] 训练中：epoch=10/20, iktex=7800, loss=0.037349

[2026-01-23 23:13:24] 验证完成：epoch=10, Qiknklex=3.748785, PIKCP=0.9072, PIKNAQ=0.776449, XMSE=1.048643, MAE=0.864131

[2026-01-23 23:13:25] 训练中：epoch=11/20, iktex=7820, loss=0.039232

[2026-01-23 23:13:25] 训练中：epoch=11/20, iktex=7840, loss=0.035999

[2026-01-23 23:13:26] 训练中：epoch=11/20, iktex=7860, loss=0.041675

[2026-01-23 23:13:27] 训练中：epoch=11/20, iktex=7880, loss=0.038223

[2026-01-23 23:13:27] 训练中：epoch=11/20, iktex=7900, loss=0.041371

[2026-01-23 23:13:28] 训练中：epoch=11/20, iktex=7920, loss=0.041761

[2026-01-23 23:13:29] 训练中：epoch=11/20, iktex=7940, loss=0.037270

[2026-01-23 23:13:29] 验证完成：epoch=11, Qiknklex=3.605713, PIKCP=0.9068, PIKNAQ=0.744311, XMSE=1.051442, MAE=0.857354
[2026-01-23 23:13:29] 局部最佳更新：已保存

[2026-01-23 23:13:30] 训练中：epoch=12/20, iktex=7960, loss=0.035633

[2026-01-23 23:13:30] 训练中：epoch=12/20, iktex=7980, loss=0.037287

[2026-01-23 23:13:31] 训练中：epoch=12/20, iktex=8000, loss=0.036905

[2026-01-23 23:13:32] 训练中：epoch=12/20, iktex=8020, loss=0.038261

[2026-01-23 23:13:32] 训练中：epoch=12/20, iktex=8040, loss=0.039470

[2026-01-23 23:13:33] 训练中：epoch=12/20, iktex=8060, loss=0.042508

[2026-01-23 23:13:34] 训练中：epoch=12/20, iktex=8080, loss=0.038592

[2026-01-23 23:13:34] 验证完成：epoch=12, Qiknklex=3.570201, PIKCP=0.9050, PIKNAQ=0.736858, XMSE=0.997299, MAE=0.821712
[2026-01-23 23:13:34] 局部最佳更新：已保存

[2026-01-23 23:13:35] 训练中：epoch=13/20, iktex=8100, loss=0.041600

[2026-01-23 23:13:36] 训练中：epoch=13/20, iktex=8120, loss=0.034481

[2026-01-23 23:13:36] 训练中：epoch=13/20, iktex=8140, loss=0.039447

[2026-01-23 23:13:37] 训练中：epoch=13/20, iktex=8160, loss=0.040157

[2026-01-23 23:13:38] 训练中：epoch=13/20, iktex=8180, loss=0.034549

[2026-01-23 23:13:38] 训练中：epoch=13/20, iktex=8200, loss=0.037208

[2026-01-23 23:13:39] 训练中：epoch=13/20, iktex=8220, loss=0.034991

[2026-01-23 23:13:39] 验证完成：epoch=13, Qiknklex=3.731435, PIKCP=0.9070, PIKNAQ=0.773761, XMSE=1.033596, MAE=0.850416

[2026-01-23 23:13:40] 训练中：epoch=14/20, iktex=8240, loss=0.032365

[2026-01-23 23:13:41] 训练中：epoch=14/20, iktex=8260, loss=0.033198

[2026-01-23 23:13:41] 训练中：epoch=14/20, iktex=8280, loss=0.035783

[2026-01-23 23:13:42] 训练中：epoch=14/20, iktex=8300, loss=0.036383

[2026-01-23 23:13:43] 训练中：epoch=14/20, iktex=8320, loss=0.035446

[2026-01-23 23:13:43] 训练中：epoch=14/20, iktex=8340, loss=0.036223

[2026-01-23 23:13:44] 验证完成：epoch=14, Qiknklex=3.606177, PIKCP=0.9074, PIKNAQ=0.749443, XMSE=1.030783, MAE=0.843206

[2026-01-23 23:13:44] 训练中：epoch=15/20, iktex=8360, loss=0.032486

[2026-01-23 23:13:45] 训练中：epoch=15/20, iktex=8380, loss=0.036510

[2026-01-23 23:13:46] 训练中：epoch=15/20, iktex=8400, loss=0.032929

[2026-01-23 23:13:47] 训练中：epoch=15/20, iktex=8420, loss=0.036233

[2026-01-23 23:13:47] 训练中：epoch=15/20, iktex=8440, loss=0.038185

[2026-01-23 23:13:48] 训练中：epoch=15/20, iktex=8460, loss=0.039148

[2026-01-23 23:13:49] 训练中：epoch=15/20, iktex=8480, loss=0.036644

[2026-01-23 23:13:49] 验证完成：epoch=15, Qiknklex=3.568895, PIKCP=0.9071, PIKNAQ=0.734379, XMSE=1.025899, MAE=0.838080
[2026-01-23 23:13:49] 局部最佳更新：已保存

[2026-01-23 23:13:50] 训练中：epoch=16/20, iktex=8500, loss=0.033920

[2026-01-23 23:13:50] 训练中：epoch=16/20, iktex=8520, loss=0.038148

[2026-01-23 23:13:51] 训练中：epoch=16/20, iktex=8540, loss=0.036824

[2026-01-23 23:13:52] 训练中：epoch=16/20, iktex=8560, loss=0.039315

[2026-01-23 23:13:52] 训练中：epoch=16/20, iktex=8580, loss=0.037658

[2026-01-23 23:13:53] 训练中：epoch=16/20, iktex=8600, loss=0.034824

[2026-01-23 23:13:54] 训练中：epoch=16/20, iktex=8620, loss=0.035086

[2026-01-23 23:13:55] 验证完成：epoch=16, Qiknklex=3.683946, PIKCP=0.9071, PIKNAQ=0.762617, XMSE=1.080672, MAE=0.897092

[2026-01-23 23:13:55] 训练中：epoch=17/20, iktex=8640, loss=0.032713

[2026-01-23 23:13:56] 训练中：epoch=17/20, iktex=8660, loss=0.038360

[2026-01-23 23:13:56] 训练中：epoch=17/20, iktex=8680, loss=0.039364

[2026-01-23 23:13:57] 训练中：epoch=17/20, iktex=8700, loss=0.033257

[2026-01-23 23:13:58] 训练中：epoch=17/20, iktex=8720, loss=0.035995

[2026-01-23 23:13:58] 训练中：epoch=17/20, iktex=8740, loss=0.038532

[2026-01-23 23:13:59] 训练中：epoch=17/20, iktex=8760, loss=0.040539

[2026-01-23 23:14:00] 验证完成：epoch=17, Qiknklex=3.503765, PIKCP=0.9062, PIKNAQ=0.722188, XMSE=0.907735, MAE=0.726207
[2026-01-23 23:14:00] 局部最佳更新：已保存

[2026-01-23 23:14:00] 训练中：epoch=18/20, iktex=8780, loss=0.032123

[2026-01-23 23:14:01] 训练中：epoch=18/20, iktex=8800, loss=0.032222

[2026-01-23 23:14:01] 训练中：epoch=18/20, iktex=8820, loss=0.036834

[2026-01-23 23:14:02] 训练中：epoch=18/20, iktex=8840, loss=0.032900

[2026-01-23 23:14:03] 训练中：epoch=18/20, iktex=8860, loss=0.039182

[2026-01-23 23:14:04] 训练中：epoch=18/20, iktex=8880, loss=0.034336

[2026-01-23 23:14:04] 训练中：epoch=18/20, iktex=8900, loss=0.038123

[2026-01-23 23:14:05] 验证完成：epoch=18, Qiknklex=3.642713, PIKCP=0.9072, PIKNAQ=0.750998, XMSE=0.982710, MAE=0.809618

[2026-01-23 23:14:05] 训练中：epoch=19/20, iktex=8920, loss=0.037419

[2026-01-23 23:14:06] 训练中：epoch=19/20, iktex=8940, loss=0.037711

[2026-01-23 23:14:07] 训练中：epoch=19/20, iktex=8960, loss=0.037298

[2026-01-23 23:14:07] 训练中：epoch=19/20, iktex=8980, loss=0.032884

[2026-01-23 23:14:08] 训练中：epoch=19/20, iktex=9000, loss=0.037056

[2026-01-23 23:14:09] 训练中：epoch=19/20, iktex=9020, loss=0.034297

[2026-01-23 23:14:09] 训练中：epoch=19/20, iktex=9040, loss=0.032854

[2026-01-23 23:14:10] 验证完成：epoch=19, Qiknklex=3.626226, PIKCP=0.9070, PIKNAQ=0.750596, XMSE=1.064214, MAE=0.881319

[2026-01-23 23:14:11] 训练中：epoch=20/20, iktex=9060, loss=0.036678

[2026-01-23 23:14:11] 训练中：epoch=20/20, iktex=9080, loss=0.037077

[2026-01-23 23:14:12] 训练中：epoch=20/20, iktex=9100, loss=0.035574

[2026-01-23 23:14:13] 训练中：epoch=20/20, iktex=9120, loss=0.033177

[2026-01-23 23:14:13] 训练中：epoch=20/20, iktex=9140, loss=0.036469

[2026-01-23 23:14:14] 训练中：epoch=20/20, iktex=9160, loss=0.033140

[2026-01-23 23:14:15] 验证完成：epoch=20, Qiknklex=3.689406, PIKCP=0.9073, PIKNAQ=0.761090, XMSE=1.068273, MAE=0.881352

[2026-01-23 23:14:15] 训练中：epoch=1/20, iktex=6440, loss=1.282002

[2026-01-23 23:14:16] 训练中：epoch=1/20, iktex=6460, loss=0.172583

[2026-01-23 23:14:17] 训练中：epoch=1/20, iktex=6480, loss=0.154410

[2026-01-23 23:14:18] 训练中：epoch=1/20, iktex=6500, loss=0.138090

[2026-01-23 23:14:19] 训练中：epoch=1/20, iktex=6520, loss=0.133163

[2026-01-23 23:14:20] 训练中：epoch=1/20, iktex=6540, loss=0.126406

[2026-01-23 23:14:20] 训练中：epoch=1/20, iktex=6560, loss=0.124879

[2026-01-23 23:14:22] 验证完成：epoch=1, Qiknklex=3.793339, PIKCP=0.9084, PIKNAQ=0.760772, XMSE=1.160752, MAE=0.942118
[2026-01-23 23:14:22] 局部最佳更新：已保存

[2026-01-23 23:14:22] 训练中：epoch=2/20, iktex=6580, loss=0.100814

[2026-01-23 23:14:23] 训练中：epoch=2/20, iktex=6600, loss=0.099940

[2026-01-23 23:14:23] 训练中：epoch=2/20, iktex=6620, loss=0.089544

[2026-01-23 23:14:24] 训练中：epoch=2/20, iktex=6640, loss=0.091035

[2026-01-23 23:14:25] 训练中：epoch=2/20, iktex=6660, loss=0.085258

[2026-01-23 23:14:26] 训练中：epoch=2/20, iktex=6680, loss=0.077619

[2026-01-23 23:14:27] 训练中：epoch=2/20, iktex=6700, loss=0.076364

[2026-01-23 23:14:28] 验证完成：epoch=2, Qiknklex=3.842346, PIKCP=0.9082, PIKNAQ=0.771179, XMSE=1.120754, MAE=0.937120

[2026-01-23 23:14:28] 训练中：epoch=3/20, iktex=6720, loss=0.078073

[2026-01-23 23:14:29] 训练中：epoch=3/20, iktex=6740, loss=0.072452

[2026-01-23 23:14:30] 训练中：epoch=3/20, iktex=6760, loss=0.074702

[2026-01-23 23:14:30] 训练中：epoch=3/20, iktex=6780, loss=0.074178

[2026-01-23 23:14:31] 训练中：epoch=3/20, iktex=6800, loss=0.070129

[2026-01-23 23:14:32] 训练中：epoch=3/20, iktex=6820, loss=0.069944

[2026-01-23 23:14:33] 训练中：epoch=3/20, iktex=6840, loss=0.069425

[2026-01-23 23:14:34] 验证完成：epoch=3, Qiknklex=3.813339, PIKCP=0.9087, PIKNAQ=0.762452, XMSE=1.064339, MAE=0.889616

[2026-01-23 23:14:34] 训练中：epoch=4/20, iktex=6860, loss=0.070851

[2026-01-23 23:14:35] 训练中：epoch=4/20, iktex=6880, loss=0.067644

[2026-01-23 23:14:36] 训练中：epoch=4/20, iktex=6900, loss=0.066844

[2026-01-23 23:14:36] 训练中：epoch=4/20, iktex=6920, loss=0.069041

[2026-01-23 23:14:37] 训练中：epoch=4/20, iktex=6940, loss=0.061278

[2026-01-23 23:14:38] 训练中：epoch=4/20, iktex=6960, loss=0.064931

[2026-01-23 23:14:39] 训练中：epoch=4/20, iktex=6980, loss=0.063323

[2026-01-23 23:14:40] 验证完成：epoch=4, Qiknklex=3.703319, PIKCP=0.9072, PIKNAQ=0.743303, XMSE=1.028098, MAE=0.846782
[2026-01-23 23:14:40] 局部最佳更新：已保存

[2026-01-23 23:14:40] 训练中：epoch=5/20, iktex=7000, loss=0.063612

[2026-01-23 23:14:41] 训练中：epoch=5/20, iktex=7020, loss=0.063200

[2026-01-23 23:14:42] 训练中：epoch=5/20, iktex=7040, loss=0.064126

[2026-01-23 23:14:42] 训练中：epoch=5/20, iktex=7060, loss=0.058630

[2026-01-23 23:14:43] 训练中：epoch=5/20, iktex=7080, loss=0.066255

[2026-01-23 23:14:44] 训练中：epoch=5/20, iktex=7100, loss=0.060089

[2026-01-23 23:14:45] 训练中：epoch=5/20, iktex=7120, loss=0.063195

[2026-01-23 23:14:46] 验证完成：epoch=5, Qiknklex=3.724697, PIKCP=0.9083, PIKNAQ=0.748057, XMSE=1.011990, MAE=0.827407

[2026-01-23 23:14:46] 训练中：epoch=6/20, iktex=7140, loss=0.065207

[2026-01-23 23:14:47] 训练中：epoch=6/20, iktex=7160, loss=0.056940

[2026-01-23 23:14:48] 训练中：epoch=6/20, iktex=7180, loss=0.057181

[2026-01-23 23:14:49] 训练中：epoch=6/20, iktex=7200, loss=0.054023

[2026-01-23 23:14:49] 训练中：epoch=6/20, iktex=7220, loss=0.062410

[2026-01-23 23:14:50] 训练中：epoch=6/20, iktex=7240, loss=0.059775

[2026-01-23 23:14:51] 训练中：epoch=6/20, iktex=7260, loss=0.058215

[2026-01-23 23:14:52] 验证完成：epoch=6, Qiknklex=3.713259, PIKCP=0.9072, PIKNAQ=0.744018, XMSE=1.101426, MAE=0.910005

[2026-01-23 23:14:52] 训练中：epoch=7/20, iktex=7280, loss=0.060289

[2026-01-23 23:14:53] 训练中：epoch=7/20, iktex=7300, loss=0.058916

[2026-01-23 23:14:54] 训练中：epoch=7/20, iktex=7320, loss=0.052462

[2026-01-23 23:14:55] 训练中：epoch=7/20, iktex=7340, loss=0.052027

[2026-01-23 23:14:56] 训练中：epoch=7/20, iktex=7360, loss=0.063273

[2026-01-23 23:14:56] 训练中：epoch=7/20, iktex=7380, loss=0.053303

[2026-01-23 23:14:58] 验证完成：epoch=7, Qiknklex=3.671830, PIKCP=0.9063, PIKNAQ=0.735594, XMSE=1.089266, MAE=0.895191
[2026-01-23 23:14:58] 局部最佳更新：已保存

[2026-01-23 23:14:58] 训练中：epoch=8/20, iktex=7400, loss=0.056140

[2026-01-23 23:14:59] 训练中：epoch=8/20, iktex=7420, loss=0.055627

[2026-01-23 23:14:59] 训练中：epoch=8/20, iktex=7440, loss=0.052703

[2026-01-23 23:15:00] 训练中：epoch=8/20, iktex=7460, loss=0.057159

[2026-01-23 23:15:01] 训练中：epoch=8/20, iktex=7480, loss=0.056298

[2026-01-23 23:15:02] 训练中：epoch=8/20, iktex=7500, loss=0.052028

[2026-01-23 23:15:03] 训练中：epoch=8/20, iktex=7520, loss=0.052862

[2026-01-23 23:15:04] 验证完成：epoch=8, Qiknklex=3.680659, PIKCP=0.9074, PIKNAQ=0.733711, XMSE=1.184723, MAE=0.987161

[2026-01-23 23:15:04] 训练中：epoch=9/20, iktex=7540, loss=0.052277

[2026-01-23 23:15:05] 训练中：epoch=9/20, iktex=7560, loss=0.050037

[2026-01-23 23:15:06] 训练中：epoch=9/20, iktex=7580, loss=0.053404

[2026-01-23 23:15:06] 训练中：epoch=9/20, iktex=7600, loss=0.050379

[2026-01-23 23:15:07] 训练中：epoch=9/20, iktex=7620, loss=0.050918

[2026-01-23 23:15:08] 训练中：epoch=9/20, iktex=7640, loss=0.048050

[2026-01-23 23:15:09] 训练中：epoch=9/20, iktex=7660, loss=0.053981

[2026-01-23 23:15:10] 验证完成：epoch=9, Qiknklex=3.708803, PIKCP=0.9086, PIKNAQ=0.741941, XMSE=1.166398, MAE=0.978727

[2026-01-23 23:15:10] 训练中：epoch=10/20, iktex=7680, loss=0.055456

[2026-01-23 23:15:11] 训练中：epoch=10/20, iktex=7700, loss=0.058276

[2026-01-23 23:15:12] 训练中：epoch=10/20, iktex=7720, loss=0.053617

[2026-01-23 23:15:13] 训练中：epoch=10/20, iktex=7740, loss=0.051936

[2026-01-23 23:15:13] 训练中：epoch=10/20, iktex=7760, loss=0.054969

[2026-01-23 23:15:14] 训练中：epoch=10/20, iktex=7780, loss=0.049272

[2026-01-23 23:15:15] 训练中：epoch=10/20, iktex=7800, loss=0.051235

[2026-01-23 23:15:16] 验证完成：epoch=10, Qiknklex=3.662698, PIKCP=0.9078, PIKNAQ=0.729269, XMSE=1.187407, MAE=0.990985
[2026-01-23 23:15:16] 局部最佳更新：已保存

[2026-01-23 23:15:16] 训练中：epoch=11/20, iktex=7820, loss=0.048952

[2026-01-23 23:15:17] 训练中：epoch=11/20, iktex=7840, loss=0.047663

[2026-01-23 23:15:18] 训练中：epoch=11/20, iktex=7860, loss=0.046384

[2026-01-23 23:15:19] 训练中：epoch=11/20, iktex=7880, loss=0.048405

[2026-01-23 23:15:20] 训练中：epoch=11/20, iktex=7900, loss=0.053623

[2026-01-23 23:15:20] 训练中：epoch=11/20, iktex=7920, loss=0.051493

[2026-01-23 23:15:21] 训练中：epoch=11/20, iktex=7940, loss=0.050856

[2026-01-23 23:15:22] 验证完成：epoch=11, Qiknklex=3.646081, PIKCP=0.9069, PIKNAQ=0.726801, XMSE=1.152584, MAE=0.963104
[2026-01-23 23:15:22] 局部最佳更新：已保存

[2026-01-23 23:15:23] 训练中：epoch=12/20, iktex=7960, loss=0.046001

[2026-01-23 23:15:23] 训练中：epoch=12/20, iktex=7980, loss=0.050100

[2026-01-23 23:15:24] 训练中：epoch=12/20, iktex=8000, loss=0.052605

[2026-01-23 23:15:25] 训练中：epoch=12/20, iktex=8020, loss=0.046454

[2026-01-23 23:15:26] 训练中：epoch=12/20, iktex=8040, loss=0.046147

[2026-01-23 23:15:27] 训练中：epoch=12/20, iktex=8060, loss=0.048600

[2026-01-23 23:15:27] 训练中：epoch=12/20, iktex=8080, loss=0.051145

[2026-01-23 23:15:28] 验证完成：epoch=12, Qiknklex=3.629983, PIKCP=0.9062, PIKNAQ=0.725802, XMSE=1.185213, MAE=0.995092

[2026-01-23 23:15:28] 局部最佳更新：已保存

[2026-01-23 23:15:29] 训练中：epoch=13/20, iktex=8100, loss=0.047542

[2026-01-23 23:15:29] 训练中：epoch=13/20, iktex=8120, loss=0.046402

[2026-01-23 23:15:30] 训练中：epoch=13/20, iktex=8140, loss=0.045979

[2026-01-23 23:15:31] 训练中：epoch=13/20, iktex=8160, loss=0.048902

[2026-01-23 23:15:32] 训练中：epoch=13/20, iktex=8180, loss=0.048367

[2026-01-23 23:15:33] 训练中：epoch=13/20, iktex=8200, loss=0.046329

[2026-01-23 23:15:33] 训练中：epoch=13/20, iktex=8220, loss=0.042892

[2026-01-23 23:15:34] 验证完成：epoch=13, Qiknklex=3.646758, PIKCP=0.9073, PIKNAQ=0.725848, XMSE=1.250091, MAE=1.053817

[2026-01-23 23:15:35] 训练中：epoch=14/20, iktex=8240, loss=0.046378

[2026-01-23 23:15:36] 训练中：epoch=14/20, iktex=8260, loss=0.046826

[2026-01-23 23:15:36] 训练中：epoch=14/20, iktex=8280, loss=0.047692

[2026-01-23 23:15:37] 训练中：epoch=14/20, iktex=8300, loss=0.045581

[2026-01-23 23:15:38] 训练中：epoch=14/20, iktex=8320, loss=0.043976

[2026-01-23 23:15:39] 训练中：epoch=14/20, iktex=8340, loss=0.048260

[2026-01-23 23:15:40] 验证完成：epoch=14, Qiknklex=3.615303, PIKCP=0.9071, PIKNAQ=0.717367, XMSE=1.251730, MAE=1.053313
[2026-01-23 23:15:40] 局部最佳更新：已保存

[2026-01-23 23:15:40] 训练中：epoch=15/20, iktex=8360, loss=0.045742

[2026-01-23 23:15:41] 训练中：epoch=15/20, iktex=8380, loss=0.042663

[2026-01-23 23:15:42] 训练中：epoch=15/20, iktex=8400, loss=0.042392

[2026-01-23 23:15:42] 训练中：epoch=15/20, iktex=8420, loss=0.046202

[2026-01-23 23:15:43] 训练中：epoch=15/20, iktex=8440, loss=0.041219

[2026-01-23 23:15:44] 训练中：epoch=15/20, iktex=8460, loss=0.042399

[2026-01-23 23:15:45] 训练中：epoch=15/20, iktex=8480, loss=0.041951

[2026-01-23 23:15:46] 验证完成：epoch=15, Qiknklex=3.621491, PIKCP=0.9069, PIKNAQ=0.721262, XMSE=1.230087, MAE=1.036918

[2026-01-23 23:15:46] 训练中：epoch=16/20, iktex=8500, loss=0.044136

[2026-01-23 23:15:47] 训练中：epoch=16/20, iktex=8520, loss=0.044019

[2026-01-23 23:15:48] 训练中：epoch=16/20, iktex=8540, loss=0.044040

[2026-01-23 23:15:49] 训练中：epoch=16/20, iktex=8560, loss=0.041875

[2026-01-23 23:15:49] 训练中：epoch=16/20, iktex=8580, loss=0.039989

[2026-01-23 23:15:50] 训练中：epoch=16/20, iktex=8600, loss=0.041302

[2026-01-23 23:15:51] 训练中：epoch=16/20, iktex=8620, loss=0.043042

[2026-01-23 23:15:52] 验证完成：epoch=16, Qiknklex=3.557444, PIKCP=0.9089, PIKNAQ=0.709421, XMSE=1.218337, MAE=1.023704
[2026-01-23 23:15:52] 局部最佳更新：已保存

[2026-01-23 23:15:52] 训练中：epoch=17/20, iktex=8640, loss=0.044326

[2026-01-23 23:15:53] 训练中：epoch=17/20, iktex=8660, loss=0.040946

[2026-01-23 23:15:54] 训练中：epoch=17/20, iktex=8680, loss=0.041453

[2026-01-23 23:15:55] 训练中：epoch=17/20, iktex=8700, loss=0.042846

[2026-01-23 23:15:55] 训练中：epoch=17/20, iktex=8720, loss=0.040026

[2026-01-23 23:15:56] 训练中：epoch=17/20, iktex=8740, loss=0.040455

[2026-01-23 23:15:57] 训练中：epoch=17/20, iktex=8760, loss=0.043140

[2026-01-23 23:15:58] 验证完成：epoch=17, Qiknklex=3.563346, PIKCP=0.9070, PIKNAQ=0.708699, XMSE=1.202026, MAE=1.007497

[2026-01-23 23:15:58] 训练中：epoch=18/20, iktex=8780, loss=0.043106

[2026-01-23 23:15:59] 训练中：epoch=18/20, iktex=8800, loss=0.040609

[2026-01-23 23:16:00] 训练中：epoch=18/20, iktex=8820, loss=0.042737

[2026-01-23 23:16:01] 训练中：epoch=18/20, iktex=8840, loss=0.043190

[2026-01-23 23:16:02] 训练中：epoch=18/20, iktex=8860, loss=0.042260

[2026-01-23 23:16:02] 训练中：epoch=18/20, iktex=8880, loss=0.043727

[2026-01-23 23:16:03] 训练中：epoch=18/20, iktex=8900, loss=0.043136

[2026-01-23 23:16:04] 验证完成：epoch=18, Qiknklex=3.570060, PIKCP=0.9066, PIKNAQ=0.710303, XMSE=1.197836, MAE=1.003854

[2026-01-23 23:16:04] 训练中：epoch=19/20, iktex=8920, loss=0.042298

[2026-01-23 23:16:05] 训练中：epoch=19/20, iktex=8940, loss=0.042658

[2026-01-23 23:16:06] 训练中：epoch=19/20, iktex=8960, loss=0.043444

[2026-01-23 23:16:07] 训练中：epoch=19/20, iktex=8980, loss=0.040983

[2026-01-23 23:16:08] 训练中：epoch=19/20, iktex=9000, loss=0.044268

[2026-01-23 23:16:08] 训练中：epoch=19/20, iktex=9020, loss=0.042467

[2026-01-23 23:16:09] 训练中：epoch=19/20, iktex=9040, loss=0.039125

[2026-01-23 23:16:10] 验证完成：epoch=19, Qiknklex=3.551110, PIKCP=0.9073, PIKNAQ=0.707306, XMSE=1.193976, MAE=1.000206
[2026-01-23 23:16:10] 局部最佳更新：已保存

[2026-01-23 23:16:11] 训练中：epoch=20/20, iktex=9060, loss=0.042768

[2026-01-23 23:16:11] 训练中：epoch=20/20, iktex=9080, loss=0.041499

[2026-01-23 23:16:12] 训练中：epoch=20/20, iktex=9100, loss=0.042243

[2026-01-23 23:16:13] 训练中：epoch=20/20, iktex=9120, loss=0.039871

[2026-01-23 23:16:14] 训练中：epoch=20/20, iktex=9140, loss=0.039302

[2026-01-23 23:16:15] 训练中：epoch=20/20, iktex=9160, loss=0.041011

[2026-01-23 23:16:16] 验证完成：epoch=20, Qiknklex=3.558233, PIKCP=0.9071, PIKNAQ=0.707042, XMSE=1.260348, MAE=1.064257

[2026-01-23 23:16:16] 训练中：epoch=1/20, iktex=6440, loss=0.844051

[2026-01-23 23:16:17] 训练中：epoch=1/20, iktex=6460, loss=0.163745

[2026-01-23 23:16:18] 训练中：epoch=1/20, iktex=6480, loss=0.139142

[2026-01-23 23:16:19] 训练中：epoch=1/20, iktex=6500, loss=0.128216

[2026-01-23 23:16:20] 训练中：epoch=1/20, iktex=6520, loss=0.126266

[2026-01-23 23:16:20] 训练中：epoch=1/20, iktex=6540, loss=0.112375

[2026-01-23 23:16:21] 训练中：epoch=1/20, iktex=6560, loss=0.105301

[2026-01-23 23:16:22] 验证完成：epoch=1, Qiknklex=3.872586, PIKCP=0.9073, PIKNAQ=0.772859, XMSE=1.261858, MAE=1.041255
[2026-01-23 23:16:22] 局部最佳更新：已保存

[2026-01-23 23:16:23] 训练中：epoch=2/20, iktex=6580, loss=0.097619

[2026-01-23 23:16:23] 训练中：epoch=2/20, iktex=6600, loss=0.092108

[2026-01-23 23:16:24] 训练中：epoch=2/20, iktex=6620, loss=0.084663

[2026-01-23 23:16:25] 训练中：epoch=2/20, iktex=6640, loss=0.081203

[2026-01-23 23:16:26] 训练中：epoch=2/20, iktex=6660, loss=0.070397

[2026-01-23 23:16:27] 训练中：epoch=2/20, iktex=6680, loss=0.075397

[2026-01-23 23:16:27] 训练中：epoch=2/20, iktex=6700, loss=0.078621

[2026-01-23 23:16:28] 验证完成：epoch=2, Qiknklex=4.065202, PIKCP=0.9078, PIKNAQ=0.805418, XMSE=1.158294, MAE=0.970220

[2026-01-23 23:16:29] 训练中：epoch=3/20, iktex=6720, loss=0.070540

[2026-01-23 23:16:29] 训练中：epoch=3/20, iktex=6740, loss=0.068099

[2026-01-23 23:16:30] 训练中：epoch=3/20, iktex=6760, loss=0.062283

[2026-01-23 23:16:31] 训练中：epoch=3/20, iktex=6780, loss=0.063834

[2026-01-23 23:16:32] 训练中：epoch=3/20, iktex=6800, loss=0.060890

[2026-01-23 23:16:33] 训练中：epoch=3/20, iktex=6820, loss=0.058988

[2026-01-23 23:16:34] 训练中：epoch=3/20, iktex=6840, loss=0.065781

[2026-01-23 23:16:34] 验证完成：epoch=3, Qiknklex=3.823593, PIKCP=0.9058, PIKNAQ=0.763791, XMSE=1.141341, MAE=0.953265
[2026-01-23 23:16:34] 局部最佳更新：已保存

[2026-01-23 23:16:35] 训练中：epoch=4/20, iktex=6860, loss=0.061357

[2026-01-23 23:16:36] 训练中：epoch=4/20, iktex=6880, loss=0.055511

[2026-01-23 23:16:37] 训练中：epoch=4/20, iktex=6900, loss=0.061853

[2026-01-23 23:16:37] 训练中：epoch=4/20, iktex=6920, loss=0.057092

[2026-01-23 23:16:38] 训练中：epoch=4/20, iktex=6940, loss=0.054265

[2026-01-23 23:16:39] 训练中：epoch=4/20, iktex=6960, loss=0.055739

[2026-01-23 23:16:40] 训练中：epoch=4/20, iktex=6980, loss=0.053699

[2026-01-23 23:16:40] 验证完成：epoch=4, Qiknklex=3.821039, PIKCP=0.9068, PIKNAQ=0.759788, XMSE=1.118227, MAE=0.924676
[2026-01-23 23:16:40] 局部最佳更新：已保存

[2026-01-23 23:16:41] 训练中：epoch=5/20, iktex=7000, loss=0.057591

[2026-01-23 23:16:42] 训练中：epoch=5/20, iktex=7020, loss=0.052219

[2026-01-23 23:16:43] 训练中：epoch=5/20, iktex=7040, loss=0.055339

[2026-01-23 23:16:44] 训练中：epoch=5/20, iktex=7060, loss=0.052354

[2026-01-23 23:16:44] 训练中：epoch=5/20, iktex=7080, loss=0.060247

[2026-01-23 23:16:45] 训练中：epoch=5/20, iktex=7100, loss=0.050718

[2026-01-23 23:16:46] 训练中：epoch=5/20, iktex=7120, loss=0.048750

[2026-01-23 23:16:47] 验证完成：epoch=5, Qiknklex=3.778167, PIKCP=0.9071, PIKNAQ=0.753851, XMSE=1.103630, MAE=0.921548
[2026-01-23 23:16:47] 局部最佳更新：已保存

[2026-01-23 23:16:47] 训练中：epoch=6/20, iktex=7140, loss=0.048673

[2026-01-23 23:16:48] 训练中：epoch=6/20, iktex=7160, loss=0.051614

[2026-01-23 23:16:49] 训练中：epoch=6/20, iktex=7180, loss=0.054028

[2026-01-23 23:16:50] 训练中：epoch=6/20, iktex=7200, loss=0.049124

[2026-01-23 23:16:51] 训练中：epoch=6/20, iktex=7220, loss=0.049484

[2026-01-23 23:16:52] 训练中：epoch=6/20, iktex=7240, loss=0.049838

[2026-01-23 23:16:52] 训练中：epoch=6/20, iktex=7260, loss=0.046586

[2026-01-23 23:16:53] 验证完成：epoch=6, Qiknklex=3.793920, PIKCP=0.9073, PIKNAQ=0.757420, XMSE=1.127814, MAE=0.936427

[2026-01-23 23:16:54] 训练中：epoch=7/20, iktex=7280, loss=0.049226

[2026-01-23 23:16:54] 训练中：epoch=7/20, iktex=7300, loss=0.045179

[2026-01-23 23:16:55] 训练中：epoch=7/20, iktex=7320, loss=0.050741

[2026-01-23 23:16:56] 训练中：epoch=7/20, iktex=7340, loss=0.049136

[2026-01-23 23:16:57] 训练中：epoch=7/20, iktex=7360, loss=0.042228

[2026-01-23 23:16:58] 训练中：epoch=7/20, iktex=7380, loss=0.046040

[2026-01-23 23:16:59] 验证完成：epoch=7, Qiknklex=3.725393, PIKCP=0.9077, PIKNAQ=0.746171, XMSE=1.128660, MAE=0.938729
[2026-01-23 23:16:59] 局部最佳更新：已保存

[2026-01-23 23:16:59] 训练中：epoch=8/20, iktex=7400, loss=0.054568

[2026-01-23 23:17:00] 训练中：epoch=8/20, iktex=7420, loss=0.045420

[2026-01-23 23:17:01] 训练中：epoch=8/20, iktex=7440, loss=0.042573

[2026-01-23 23:17:02] 训练中：epoch=8/20, iktex=7460, loss=0.042963

[2026-01-23 23:17:02] 训练中：epoch=8/20, iktex=7480, loss=0.043155

[2026-01-23 23:17:03] 训练中：epoch=8/20, iktex=7500, loss=0.042385

[2026-01-23 23:17:04] 训练中：epoch=8/20, iktex=7520, loss=0.043562

[2026-01-23 23:17:05] 验证完成：epoch=8, Qiknklex=3.733421, PIKCP=0.9072, PIKNAQ=0.746361, XMSE=1.162513, MAE=0.971038

[2026-01-23 23:17:05] 训练中：epoch=9/20, iktex=7540, loss=0.045725

[2026-01-23 23:17:06] 训练中：epoch=9/20, iktex=7560, loss=0.043435

[2026-01-23 23:17:07] 训练中：epoch=9/20, iktex=7580, loss=0.046259

[2026-01-23 23:17:08] 训练中：epoch=9/20, iktex=7600, loss=0.043958

[2026-01-23 23:17:09] 训练中：epoch=9/20, iktex=7620, loss=0.044948

[2026-01-23 23:17:09] 训练中：epoch=9/20, iktex=7640, loss=0.043466

[2026-01-23 23:17:10] 训练中：epoch=9/20, iktex=7660, loss=0.042120

[2026-01-23 23:17:11] 验证完成：epoch=9, Qiknklex=3.677055, PIKCP=0.9074, PIKNAQ=0.731040, XMSE=1.162225, MAE=0.969121
[2026-01-23 23:17:11] 局部最佳更新：已保存

[2026-01-23 23:17:12] 训练中：epoch=10/20, iktex=7680, loss=0.042086

[2026-01-23 23:17:12] 训练中：epoch=10/20, iktex=7700, loss=0.043430

[2026-01-23 23:17:13] 训练中：epoch=10/20, iktex=7720, loss=0.044152

[2026-01-23 23:17:14] 训练中：epoch=10/20, iktex=7740, loss=0.042673

[2026-01-23 23:17:15] 训练中：epoch=10/20, iktex=7760, loss=0.039108

[2026-01-23 23:17:16] 训练中：epoch=10/20, iktex=7780, loss=0.041092

[2026-01-23 23:17:16] 训练中：epoch=10/20, iktex=7800, loss=0.040052

[2026-01-23 23:17:17] 验证完成：epoch=10, Qiknklex=3.639217, PIKCP=0.9068, PIKNAQ=0.720512, XMSE=1.089913, MAE=0.905410

[2026-01-23 23:17:17] 局部最佳更新：已保存

[2026-01-23 23:17:18] 训练中：epoch=11/20, iktex=7820, loss=0.040496

[2026-01-23 23:17:19] 训练中：epoch=11/20, iktex=7840, loss=0.041149

[2026-01-23 23:17:19] 训练中：epoch=11/20, iktex=7860, loss=0.040330

[2026-01-23 23:17:20] 训练中：epoch=11/20, iktex=7880, loss=0.039397

[2026-01-23 23:17:21] 训练中：epoch=11/20, iktex=7900, loss=0.040073

[2026-01-23 23:17:22] 训练中：epoch=11/20, iktex=7920, loss=0.039473

[2026-01-23 23:17:23] 训练中：epoch=11/20, iktex=7940, loss=0.038411

[2026-01-23 23:17:23] 验证完成：epoch=11, Qiknklex=3.623322, PIKCP=0.9076, PIKNAQ=0.719365, XMSE=1.150440, MAE=0.963520

[2026-01-23 23:17:23] 局部最佳更新：已保存

[2026-01-23 23:17:24] 训练中：epoch=12/20, iktex=7960, loss=0.037771

[2026-01-23 23:17:25] 训练中：epoch=12/20, iktex=7980, loss=0.041356

[2026-01-23 23:17:26] 训练中：epoch=12/20, iktex=8000, loss=0.039178

[2026-01-23 23:17:26] 训练中：epoch=12/20, iktex=8020, loss=0.038170

[2026-01-23 23:17:27] 训练中：epoch=12/20, iktex=8040, loss=0.040213

[2026-01-23 23:17:28] 训练中：epoch=12/20, iktex=8060, loss=0.037636

[2026-01-23 23:17:29] 训练中：epoch=12/20, iktex=8080, loss=0.040589

[2026-01-23 23:17:29] 验证完成：epoch=12, Qiknklex=3.614583, PIKCP=0.9065, PIKNAQ=0.719463, XMSE=1.193455, MAE=1.002509
[2026-01-23 23:17:29] 局部最佳更新：已保存

[2026-01-23 23:17:30] 训练中：epoch=13/20, iktex=8100, loss=0.038457

[2026-01-23 23:17:31] 训练中：epoch=13/20, iktex=8120, loss=0.038322

[2026-01-23 23:17:32] 训练中：epoch=13/20, iktex=8140, loss=0.037209

[2026-01-23 23:17:33] 训练中：epoch=13/20, iktex=8160, loss=0.041653

[2026-01-23 23:17:33] 训练中：epoch=13/20, iktex=8180, loss=0.040835

[2026-01-23 23:17:34] 训练中：epoch=13/20, iktex=8200, loss=0.037760

[2026-01-23 23:17:35] 训练中：epoch=13/20, iktex=8220, loss=0.038402

[2026-01-23 23:17:36] 验证完成：epoch=13, Qiknklex=3.576638, PIKCP=0.9069, PIKNAQ=0.706375, XMSE=1.136247, MAE=0.942643
[2026-01-23 23:17:36] 局部最佳更新：已保存

[2026-01-23 23:17:36] 训练中：epoch=14/20, iktex=8240, loss=0.039916

[2026-01-23 23:17:37] 训练中：epoch=14/20, iktex=8260, loss=0.037929

[2026-01-23 23:17:38] 训练中：epoch=14/20, iktex=8280, loss=0.037286

[2026-01-23 23:17:39] 训练中：epoch=14/20, iktex=8300, loss=0.036667

[2026-01-23 23:17:40] 训练中：epoch=14/20, iktex=8320, loss=0.037895

[2026-01-23 23:17:40] 训练中：epoch=14/20, iktex=8340, loss=0.038349

[2026-01-23 23:17:42] 验证完成：epoch=14, Qiknklex=3.522880, PIKCP=0.9069, PIKNAQ=0.696927, XMSE=1.170908, MAE=0.979915
[2026-01-23 23:17:42] 局部最佳更新：已保存

[2026-01-23 23:17:42] 训练中：epoch=15/20, iktex=8360, loss=0.035283

[2026-01-23 23:17:43] 训练中：epoch=15/20, iktex=8380, loss=0.040000

[2026-01-23 23:17:43] 训练中：epoch=15/20, iktex=8400, loss=0.038996

[2026-01-23 23:17:44] 训练中：epoch=15/20, iktex=8420, loss=0.036828

[2026-01-23 23:17:45] 训练中：epoch=15/20, iktex=8440, loss=0.034348

[2026-01-23 23:17:46] 训练中：epoch=15/20, iktex=8460, loss=0.036129

[2026-01-23 23:17:47] 训练中：epoch=15/20, iktex=8480, loss=0.038090

[2026-01-23 23:17:48] 验证完成：epoch=15, Qiknklex=3.563543, PIKCP=0.9066, PIKNAQ=0.700831, XMSE=1.210535, MAE=1.017767

[2026-01-23 23:17:48] 训练中：epoch=16/20, iktex=8500, loss=0.039371

[2026-01-23 23:17:49] 训练中：epoch=16/20, iktex=8520, loss=0.035054

[2026-01-23 23:17:49] 训练中：epoch=16/20, iktex=8540, loss=0.037839

[2026-01-23 23:17:50] 训练中：epoch=16/20, iktex=8560, loss=0.036015

[2026-01-23 23:17:51] 训练中：epoch=16/20, iktex=8580, loss=0.040944

[2026-01-23 23:17:52] 训练中：epoch=16/20, iktex=8600, loss=0.038425

[2026-01-23 23:17:53] 训练中：epoch=16/20, iktex=8620, loss=0.037806

[2026-01-23 23:17:54] 验证完成：epoch=16, Qiknklex=3.482223, PIKCP=0.9079, PIKNAQ=0.688006, XMSE=1.083234, MAE=0.900887
[2026-01-23 23:17:54] 局部最佳更新：已保存

[2026-01-23 23:17:54] 训练中：epoch=17/20, iktex=8640, loss=0.035480

[2026-01-23 23:17:55] 训练中：epoch=17/20, iktex=8660, loss=0.035201

[2026-01-23 23:17:56] 训练中：epoch=17/20, iktex=8680, loss=0.036961

[2026-01-23 23:17:56] 训练中：epoch=17/20, iktex=8700, loss=0.034100

[2026-01-23 23:17:57] 训练中：epoch=17/20, iktex=8720, loss=0.033129

[2026-01-23 23:17:58] 训练中：epoch=17/20, iktex=8740, loss=0.037173

[2026-01-23 23:17:59] 训练中：epoch=17/20, iktex=8760, loss=0.033702

[2026-01-23 23:18:00] 验证完成：epoch=17, Qiknklex=3.506050, PIKCP=0.9081, PIKNAQ=0.695087, XMSE=1.196327, MAE=1.005977

[2026-01-23 23:18:00] 训练中：epoch=18/20, iktex=8780, loss=0.036303

[2026-01-23 23:18:01] 训练中：epoch=18/20, iktex=8800, loss=0.035700

[2026-01-23 23:18:02] 训练中：epoch=18/20, iktex=8820, loss=0.035674

[2026-01-23 23:18:03] 训练中：epoch=18/20, iktex=8840, loss=0.036096

[2026-01-23 23:18:04] 训练中：epoch=18/20, iktex=8860, loss=0.036539

[2026-01-23 23:18:04] 训练中：epoch=18/20, iktex=8880, loss=0.034407

[2026-01-23 23:18:05] 训练中：epoch=18/20, iktex=8900, loss=0.037840

[2026-01-23 23:18:06] 验证完成：epoch=18, Qiknklex=3.521707, PIKCP=0.9075, PIKNAQ=0.690776, XMSE=1.137470, MAE=0.947129

[2026-01-23 23:18:07] 训练中：epoch=19/20, iktex=8920, loss=0.030608

[2026-01-23 23:18:07] 训练中：epoch=19/20, iktex=8940, loss=0.036275

[2026-01-23 23:18:08] 训练中：epoch=19/20, iktex=8960, loss=0.034066

[2026-01-23 23:18:09] 训练中：epoch=19/20, iktex=8980, loss=0.034828

[2026-01-23 23:18:10] 训练中：epoch=19/20, iktex=9000, loss=0.034456

[2026-01-23 23:18:11] 训练中：epoch=19/20, iktex=9020, loss=0.035063

[2026-01-23 23:18:11] 训练中：epoch=19/20, iktex=9040, loss=0.031940

[2026-01-23 23:18:12] 验证完成：epoch=19, Qiknklex=3.538940, PIKCP=0.9078, PIKNAQ=0.697258, XMSE=1.230048, MAE=1.036847

[2026-01-23 23:18:13] 训练中：epoch=20/20, iktex=9060, loss=0.034300

[2026-01-23 23:18:14] 训练中：epoch=20/20, iktex=9080, loss=0.031911

[2026-01-23 23:18:15] 训练中：epoch=20/20, iktex=9100, loss=0.036330

[2026-01-23 23:18:15] 训练中：epoch=20/20, iktex=9120, loss=0.032747

[2026-01-23 23:18:16] 训练中：epoch=20/20, iktex=9140, loss=0.034597

[2026-01-23 23:18:17] 训练中：epoch=20/20, iktex=9160, loss=0.036280

[2026-01-23 23:18:18] 验证完成：epoch=20, Qiknklex=3.513335, PIKCP=0.9070, PIKNAQ=0.693711, XMSE=1.143388, MAE=0.953677

[2026-01-23 23:18:18] 训练中：epoch=1/20, iktex=6440, loss=0.982133

[2026-01-23 23:18:19] 训练中：epoch=1/20, iktex=6460, loss=0.187753

[2026-01-23 23:18:20] 训练中：epoch=1/20, iktex=6480, loss=0.166221

[2026-01-23 23:18:21] 训练中：epoch=1/20, iktex=6500, loss=0.139439

[2026-01-23 23:18:22] 训练中：epoch=1/20, iktex=6520, loss=0.122576

[2026-01-23 23:18:22] 训练中：epoch=1/20, iktex=6540, loss=0.121046

[2026-01-23 23:18:23] 训练中：epoch=1/20, iktex=6560, loss=0.119212

[2026-01-23 23:18:24] 验证完成：epoch=1, Qiknklex=3.618781, PIKCP=0.9086, PIKNAQ=0.733263, XMSE=1.287939, MAE=1.053377
[2026-01-23 23:18:24] 局部最佳更新：已保存

[2026-01-23 23:18:25] 训练中：epoch=2/20, iktex=6580, loss=0.104946

[2026-01-23 23:18:25] 训练中：epoch=2/20, iktex=6600, loss=0.108112

[2026-01-23 23:18:26] 训练中：epoch=2/20, iktex=6620, loss=0.083196

[2026-01-23 23:18:27] 训练中：epoch=2/20, iktex=6640, loss=0.086736

[2026-01-23 23:18:28] 训练中：epoch=2/20, iktex=6660, loss=0.080689

[2026-01-23 23:18:28] 训练中：epoch=2/20, iktex=6680, loss=0.073057

[2026-01-23 23:18:29] 训练中：epoch=2/20, iktex=6700, loss=0.069539

[2026-01-23 23:18:30] 验证完成：epoch=2, Qiknklex=3.970972, PIKCP=0.9063, PIKNAQ=0.798704, XMSE=1.134426, MAE=0.941015

[2026-01-23 23:18:31] 训练中：epoch=3/20, iktex=6720, loss=0.070127

[2026-01-23 23:18:31] 训练中：epoch=3/20, iktex=6740, loss=0.066884

[2026-01-23 23:18:32] 训练中：epoch=3/20, iktex=6760, loss=0.065628

[2026-01-23 23:18:33] 训练中：epoch=3/20, iktex=6780, loss=0.061910

[2026-01-23 23:18:34] 训练中：epoch=3/20, iktex=6800, loss=0.068518

[2026-01-23 23:18:35] 训练中：epoch=3/20, iktex=6820, loss=0.056135

[2026-01-23 23:18:36] 训练中：epoch=3/20, iktex=6840, loss=0.058148

[2026-01-23 23:18:36] 验证完成：epoch=3, Qiknklex=4.040193, PIKCP=0.9046, PIKNAQ=0.820060, XMSE=1.082205, MAE=0.878320

[2026-01-23 23:18:37] 训练中：epoch=4/20, iktex=6860, loss=0.065940

[2026-01-23 23:18:38] 训练中：epoch=4/20, iktex=6880, loss=0.055886

[2026-01-23 23:18:38] 训练中：epoch=4/20, iktex=6900, loss=0.054539

[2026-01-23 23:18:39] 训练中：epoch=4/20, iktex=6920, loss=0.056176

[2026-01-23 23:18:40] 训练中：epoch=4/20, iktex=6940, loss=0.049238

[2026-01-23 23:18:41] 训练中：epoch=4/20, iktex=6960, loss=0.053896

[2026-01-23 23:18:42] 训练中：epoch=4/20, iktex=6980, loss=0.050402

[2026-01-23 23:18:42] 验证完成：epoch=4, Qiknklex=3.945017, PIKCP=0.9064, PIKNAQ=0.798767, XMSE=1.105710, MAE=0.889534

[2026-01-23 23:18:43] 训练中：epoch=5/20, iktex=7000, loss=0.049157

[2026-01-23 23:18:44] 训练中：epoch=5/20, iktex=7020, loss=0.048903

[2026-01-23 23:18:45] 训练中：epoch=5/20, iktex=7040, loss=0.047770

[2026-01-23 23:18:45] 训练中：epoch=5/20, iktex=7060, loss=0.050459

[2026-01-23 23:18:46] 训练中：epoch=5/20, iktex=7080, loss=0.048566

[2026-01-23 23:18:47] 训练中：epoch=5/20, iktex=7100, loss=0.049118

[2026-01-23 23:18:48] 训练中：epoch=5/20, iktex=7120, loss=0.055341

[2026-01-23 23:18:48] 验证完成：epoch=5, Qiknklex=3.703244, PIKCP=0.9063, PIKNAQ=0.752934, XMSE=1.074326, MAE=0.863712

[2026-01-23 23:18:49] 训练中：epoch=6/20, iktex=7140, loss=0.048305

[2026-01-23 23:18:50] 训练中：epoch=6/20, iktex=7160, loss=0.049432

[2026-01-23 23:18:51] 训练中：epoch=6/20, iktex=7180, loss=0.042551

[2026-01-23 23:18:52] 训练中：epoch=6/20, iktex=7200, loss=0.046066

[2026-01-23 23:18:52] 训练中：epoch=6/20, iktex=7220, loss=0.049302

[2026-01-23 23:18:53] 训练中：epoch=6/20, iktex=7240, loss=0.047127

[2026-01-23 23:18:54] 训练中：epoch=6/20, iktex=7260, loss=0.046953

[2026-01-23 23:18:55] 验证完成：epoch=6, Qiknklex=3.698534, PIKCP=0.9077, PIKNAQ=0.746037, XMSE=1.013045, MAE=0.813284
[2026-01-23 23:18:55] 触发早停：连续未提升=5，最佳轮数=1
[2026-01-23 23:18:55] 训练结束：最佳模型已保存
[2026-01-23 23:18:55] 训练阶段结束

[2026-01-23 23:18:55] 测试集预测她区间生成完成
[2026-01-23 23:18:55] 评估指标计算完成

[2026-01-23 23:18:57] 评估图形绘制完成
[2026-01-23 23:18:57] 流程结束

>>

结束

更多详细内容请访问

http://机器学习有图有真相Matlab实现基于LSTM-KDE长短期记忆网络（LSTM）结合核密度估计（KDE）进行多变量回归区间预测（代码已调试成功，可一键运行，每一行都有详细注释）资源-CSDN下载 https://download.csdn.net/download/xiaoxingkongyuxi/92626803

http://机器学习有图有真相Matlab实现基于LSTM-KDE长短期记忆网络（LSTM）结合核密度估计（KDE）进行多变量回归区间预测（代码已调试成功，可一键运行，每一行都有详细注释）资源-CSDN下载 https://download.csdn.net/download/xiaoxingkongyuxi/92626803