TCN-LSTM-Multihead-Attention实现风电多输入预测Matlab实现.rar

%% TCN-LSTM-Attention多变量时间序列预测，运行环境Matlab2023及以上 %% 清空环境变量 warning off % 关闭报警信息 close all % 关闭开启的图窗 clear % 清空变量 clc % 清空命令行 %% 导入数据 result = xlsread('data.xlsx'); %% 数据分析 num_samples = length(result); % 样本个数 or_dim = size(result, 2); % 原始特征+输出数目 kim = 4; % 延时步长（kim个历史数据作为自变量） zim = 1; % 跨zim个时间点进行预测 %% 划分数据集 for i = 1: num_samples - kim - zim + 1 res(i, :) = [reshape(result(i: i + kim - 1, :), 1, kim * or_dim), result(i + kim + zim - 1, :)]; end %% 数据集分析 outdim = 1; % 最后一列为输出 num_size = 0.7; % 训练集占数据集比例 num_train_s = round(num_size * num_samples); % 训练集样本个数 f_ = size(res, 2) - outdim; % 输入特征维度 %% 划分训练集和测试集 P_train = res(1: num_train_s, 1: f_)'; T_train = res(1: num_train_s, f_ + 1: end)'; M = size(P_train, 2); P_test = res(num_train_s + 1: end, 1: f_)'; T_test = res(num_train_s + 1: end, f_ + 1: end)'; N = size(P_test, 2); %% 数据归一化 [p_train, ps_input] = mapminmax(P_train, 0, 1); p_test = mapminmax('apply', P_test, ps_input); [t_train, ps_output] = mapminmax(T_train, 0, 1); t_test = mapminmax('apply', T_test, ps_output); %% 数据格式转换 pc_train{1, 1} = p_train; pc_test {1, 1} = p_test ; tc_train{1, 1} = t_train; tc_test {1, 1} = t_test ; %% 设置网络参数 numFilters = 16; % 卷积核个数 filterSize = 5; % 卷积核大小 dropoutFactor = 0.2; % 空间丢失因子 numBlocks = 1; % 残差块个数 numFeatures = f_; % 特征个数 %% 输入层结构 layer = sequenceInputLayer(numFeatures, Normalization = "rescale-symmetric", Name = "input"); %% 将输入层加入空白网络 lgraph = layerGraph(layer); outputName = layer.Name; %% 建立网络结构 -- 残差块 for i = 1 : numBlocks dilationFactor = 2 ^(i - 1); % 膨胀因子 % 建立网络结构 layers = [ convolution1dLayer(filterSize, numFilters, DilationFactor = dilationFactor, ... Padding = "causal", Name="conv1_" + i) % 一维卷积层 layerNormalizationLayer % 层归一化 spatialDropoutLayer(dropoutFactor) % 空间丢弃层 convolution1dLayer(filterSize, numFilters, ... DilationFactor = dilationFactor, Padding = "causal") % 一维卷积层 layerNormalizationLayer % 层归一化 reluLayer % 激活层 spatialDropoutLayer(dropoutFactor) % 空间丢弃层 additionLayer(2, Name = "add_" + i)]; % 合并层 lgraph = addLayers(lgraph, layers); % 将卷积层加入到网络 lgraph = connectLayers(lgraph, outputName, "conv1_" + i); % 将模块的卷积层首层和残差结构连接 % 残差连接 -- 首层 if i == 1 layer = convolution1dLayer(1, numFilters, Name = "convSkip"); % 建立残差卷积层 lgraph = addLayers(lgraph, layer); % 将残差卷积层加入到网络 lgraph = connectLayers(lgraph, outputName, "convSkip"); % 将残差卷积层 lgraph = connectLayers(lgraph, "convSkip", "add_" + i + "/in2"); % 将跳跃残差层连接到 addtion 层 输入口2 else lgraph = connectLayers(lgraph, outputName, "add_" + i + "/in2"); % 将残差层连接到 addtion 层 输入口2 end % 更新输出层名字 outputName = "add_" + i; end %% 网络输出层 % 添加 LSTM 层 % lstmLayer = lstmLayer(hiddenSize, 'Name', 'lstm'); lstmLayer = lstmLayer(100, 'Name', 'lstm'); lgraph = addLayers(lgraph, lstmLayer); lgraph = connectLayers(lgraph, outputName, 'lstm'); % 添加全连接层 layers = [ selfAttentionLayer(numFeatures,numFeatures,"Name","multihead-attention") % 多头注意力机制+多通道 fullyConnectedLayer(1, 'Name', 'fc') regressionLayer]; lgraph = addLayers(lgraph, layers); lgraph = connectLayers(lgraph, 'lstm', 'multihead-attention'); %设置权重衰减参数实现 L2 正则化 weightDecay = 0.01; % 正则化项权重 % 更新每个卷积层、LSTM 层和全连接层的权重衰减参数 numLayers = numel(lgraph.Layers); %获取网络中层的数量 for i = 1:numLayers layer = lgraph.Layers(i); if isa(layer, 'CausalConvolution1DLayer') || isa(layer, 'nnet.cnn.layer.FullyConnectedLayer') %检查当前层是否为因果卷积层、LSTM 层或全连接层。如果是，则执行以下操作 layer.WeightLearnRateFactor = 1; layer.WeightL2Factor = weightDecay; end end %% 设置训练参数 options = trainingOptions('adam', ... % Adam 梯度下降算法 'MaxEpochs', 1000, ... % 最大迭代次数 'MiniBatchSize',16,... % 批处理 'InitialLearnRate', 1e-3, ... % 初始学习率为0.001 'LearnRateSchedule', 'piecewise', ... % 学习率下降 'LearnRateDropFactor', 0.1, ... % 学习率下降因子 0.1 'LearnRateDropPeriod', 500, ... % 经过训练后 学习率为 0.001 * 0.1 'Shuffle', 'every-epoch', ... % 每次训练打乱数据集 'Plots', 'training-progress', ... % 画出曲线 'Verbose', 1); % 打印 %% 训练网络 [net,traininfo]= trainNetwork(pc_train, tc_train, lgraph, options); %% 查看网络结构 analyzeNetwork(net) %% 仿真预测 t_sim1 = predict(net, pc_train); t_sim2 = predict(net, pc_test ); %% 格式转换 T_sim1 = double(t_sim1{1, 1}); T_sim2 = double(t_sim2{1, 1}); %% 数据反归一化 T_sim1 = mapminmax('reverse', T_sim1, ps_output); T_sim2 = mapminmax('reverse', T_sim2, ps_output); %% 均方根误差 error1 = sqrt(sum((T_sim1 - T_train).^2) ./ M); error2 = sqrt(sum((T_sim2 - T_test ).^2) ./ N); %% 相关指标计算 % R2 R1 = 1 - norm(T_train - T_sim1)^2 / norm(T_train - mean(T_train))^2; R2 = 1 - norm(T_test - T_sim2)^2 / norm(T_test - mean(T_test ))^2; disp(['训练集数据的R2为：', num2str(R1)]) disp(['测试集数据的R2为：', num2str(R2)]) % MAE mae1 = sum(abs(T_sim1 - T_train)) ./ M ; mae2 = sum(abs(T_sim2 - T_test )) ./ N ; disp(['训练集数据的MAE为：', num2str(mae1)]) disp(['测试集数据的MAE为：', num2str(mae2)]) %% 平均绝对百分比误差MAPE MAPE1 = mean(abs((T_train - T_sim1)./T_train)); MAPE2 = mean(abs((T_test - T_sim2)./T_test)); disp(['训练集数据的MAPE为：', num2str(MAPE1)]) disp(['测试集数据的MAPE为：', num2str(MAPE2)]) % MBE mbe1 = sum(T_sim1 - T_train) ./ M ; mbe2 = sum(T_sim2 - T_test ) ./ N ; disp(['训练集数据的MBE为：', num2str(mbe1)]) disp(['测试集数据的MBE为：', num2str(mbe2)]) %均方误差 MSE mse1 = sum((T_sim1 - T_train).^2)./M; mse2 = sum((T_sim2 - T_test).^2)./N; disp(['训练集数据的MSE为：', num2str(mse1)]) disp(['测试集数据的MSE为：', num2str(mse2)]) % RMSE RMSE1 = sqrt(sumsqr(T_sim1 -