基于时间卷积神经网络结合门控循环单元的(TCN-GRU)的回归预测预测，多变量输入模型（Matlab完整源码和数据)

%%加入正则化 加入注意力机制 更改验证集和测试集 GRU加入TCN后 clear;close all;clc %% 数据读取 data = xlsread('PV_1.xlsx'); %输入天数 Tian=100; num_Tian = Tian*48; train_ratio = 0.7; % 训练集比例 val_ratio = 0.15; % 验证集比例 test_ratio = 0.15; % 测试集比例 num_train =train_ratio * num_Tian; num_val = val_ratio * num_Tian; num_test = num_Tian - num_train - num_val; %划分训练集验证集测试集 Tian_data=data(1:num_Tian,:); Tian_train =Tian_data(1:num_train, :); Tian_valid = Tian_data(num_train+1:num_train+num_val, :); Tian_test= Tian_data(num_train+num_val+1:end, :); lag = 1:96; %读取影响因素和功率数据 shuru=Tian_data(:,1:4);%影响因素 shuchu=Tian_data(:,end);%功率 nn=1:size(shuru,1);%正常排序 shurugeshu = size(shuru,2)+max(lag);%输入特征个数 input_train =shuru(nn(1:num_train),:);input_train=input_train'; output_train=shuchu(nn(1:num_train),:);output_train=output_train'; input_valid =shuru(nn((num_train+1):num_train+num_val),:);input_valid=input_valid'; output_valid=shuchu(nn((num_train+1):num_train+num_val),:);output_valid=output_valid'; input_test=shuru(nn((num_train+num_val+1):end),:);input_test=input_test'; output_test=shuchu(nn((num_train+num_val+1):end),:);output_test=output_test'; %%训练集 [aa,bb]=mapminmax([input_train input_valid input_test]); [cc,dd]=mapminmax([output_train output_valid output_test]); [inputn,inputps]=mapminmax('apply',input_train,bb); [outputn,outputps]=mapminmax('apply',output_train,dd); tr_history= lagmatrix(outputn, lag);%错开历史数据 tr_history = tr_history(max(lag)+1:end,:)';%截取历史数据 tr_input = inputn(:,max(lag)+1:end); tr_inputn = [tr_history;tr_input]; %按列（纵向）合并 tr_outputn = outputn(:,max(lag)+1:end); %truetrain=mapminmax('reverse',tr_inputn,dd);%反归一化 Xxun{1} = tr_inputn; %训练集输入 Yxun{1} = tr_outputn; % 训练集输出 numChannel = 9; % 通道数量 KerSize = 3; % 卷积核大小 dropoutFactor = 0.005; % droupt numChan = 4; % TCN残差块数 %% layer = sequenceInputLayer(shurugeshu, 'Name', 'input'); % 输入层 lgraph = layerGraph(layer); LayN = layer.Name; % 每层名字 % 循环遍历 numChan for i = 1:numChan KD = 2^(i-1); % 空洞系数、扩张系数 layers = [ convolution1dLayer(KerSize, numChannel, 'DilationFactor', KD, 'Padding', 'causal', 'Name', 'conv1_' + string(i)) % 卷积层 layerNormalizationLayer % 归一化层 spatialDropoutLayer(dropoutFactor) % 空间失活层 convolution1dLayer(KerSize, numChannel, 'DilationFactor', KD, 'Padding', 'causal') % 卷积层 layerNormalizationLayer % 归一化层 reluLayer % ReLU 激活函数层 spatialDropoutLayer(dropoutFactor) % 空间失活层 additionLayer(2, 'Name', 'add_' + string(i))]; % 加法层 % 添加层到网络 lgraph = addLayers(lgraph, layers); % 连接层 lgraph = connectLayers(lgraph, LayN, 'conv1_' + string(i)); % 残差连接 if i == 1 layer = convolution1dLayer(1, numChannel, 'Name', 'convSkip'); lgraph = addLayers(lgraph, layer); lgraph = connectLayers(lgraph, LayN, 'convSkip'); lgraph = connectLayers(lgraph, 'convSkip', 'add_' + string(i) + '/in2'); else lgraph = connectLayers(lgraph, LayN, 'add_' + string(i) + '/in2'); end % 更新 LayN LayN = 'add_' + string(i); end % %% % outputLayerName = 'add_' + string(numChan); % % % 查找匹配的层 % outputLayer = []; % for i = 1:numel(lgraph.Layers) % if strcmp(lgraph.Layers(i).Name, outputLayerName) % outputLayer = lgraph.Layers(i); % break; % end % end % outputLayer.InputNames % outputLayer.OutputNames % 获取最后一层卷积层的输出 % outputLayerName =' add_' + string(numChan); % 最后一层卷积层的名称 % outputLayer = lgraph.Layers(find(strcmp({lgraph.Layers.Name}, outputLayerName))); % % % 读取最后一层卷积层的输出 % % output = activations(lgraph, inputData, outputLayer.Name); % output = activations(lgraph, outputLayer.Name); % outputLayer = convolution1dLayer(KerSize, numChannel, 'DilationFactor', KD, 'Padding', 'causal', 'Name', 'conv1_' + string(numChan)); % % outputSize =outputLayer.OutputSize; % outputIndex = outputLayer.Connections.EndIndex; % outputSize = lgraph.Layers(outputIndex).OutputSize; % % 获取输出对象 % output = outputLayer.OutputNames{1}; % output1 = outputLayer.OutputSize; % size(output) %% % 添加 GRU 层 % gruLayer = gruLayer(hiddenSize, 'Name', 'gru'); gruLayer = gruLayer(3, 'Name', 'gru'); lgraph = addLayers(lgraph, gruLayer); lgraph = connectLayers(lgraph, LayN, 'gru'); % 添加全连接层 layers = [ fullyConnectedLayer(1, 'Name', 'fc') regressionLayer]; lgraph = addLayers(lgraph, layers); lgraph = connectLayers(lgraph, 'gru', 'fc'); %设置权重衰减参数实现 L2 正则化 weightDecay = 0.001; % 正则化项权重 % 更新每个卷积层、GRU 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、GRU 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', ... 'MaxEpochs', 800, ... 'InitialLearnRate', 0.02, ... 'Plots', 'training-progress', ... 'Verbose', 0); % 进行训练 net = trainNetwork(Xxun, Yxun, lgraph, options); %% 画图 %网络图 analyzeNetwork(net) %% 验证集输入 %验证集样本输入输出数据归一化 val_input=mapminmax('apply',input_valid,bb); val_output=mapminmax('apply',output_valid,dd); val_history= lagmatrix(val_output, lag); %历史功率 val_history = val_history(max(lag)+1:end,:)'; val_input = val_input(:,max(lag)+1:end); %影响因素 val_inputn = [val_history;val_input]; %按列（纵向）合并 val_outputn = output_valid(:,max(lag)+1:end); xvalid{1} = val_inputn; an = predict(net,xvalid); anx=an{1}; val_pre=mapminmax('reverse',anx,dd);%反归一化 %% % 验证集结果 reay = val_outputn; % 真实数据 prey = val_pre; % 仿真数据 reay = reshape(reay,1,size(reay,1)*size(reay,2)); % 真实数据 prey = reshape(prey,1,size(prey,1)*size(prey,2)); % 仿真数据 Tian_num=length(reay);%统计样本总数 error=prey-reay; %计算误差 MAE=sum(abs(error))/Tian_num; %计算平均绝对误差 MSE=sum(error.*error)/Tian_num; %计算均方误差 RMSE=sqrt(MSE); %计算均方误差根 ME=sum((error))/Tian_num; %计算平均绝对误差 r=min(min(corrcoef(prey,reay))); R2=r*r; disp('整体验证集') disp('--------------------------------------------') disp(['平均绝对误差MAE为： ',num2str(MAE)]) disp(['均方误差根MSE为： ',num2str(MSE)]) disp(['均方误差根RMSE为： ',num2str(RMSE)]) disp(['平均误差ME为： ',num2str(ME)]) disp(['相关系数R2为： ',num2str(R2)]) %验证集画图 figure plot(val_outputn ) hold on plot(val_pre ) legend('真实值','预测值') title('验证集') %% 测试集输入 %验证集样本输入输出数据归一化 test_input=mapminmax('apply',input_test,bb); test_output=mapminmax('apply',output_test,dd); test_history= lagmatrix(test_output, lag); %历史功率 test_history = test_history(max(lag)+1:end,:)'; test_input = test_input(:,max(lag)+1:end); %影响因素 test_inputn = [test_history;test_input]; %按列（纵向）合并 test_outputn = output_test(:,max(lag)+1:end); xtest{1} = test_