【LSTM数据预测】BP+ELM+LSTM+BiLSTM+SAELSTM数据预测【含Matlab源码 1825期】.zip

%% 多步预测 clc; clear; %% data.data=xlsread('BJPM10.xlsx'); option.len_windows=20; %滑动时间窗长度 option.p_train=0.8; %训练集样本比例 option.num_D=1; %% 滑动时间窗整理数据 %向后预测3步 Step=3; data.data_x=[]; data.data_y=[]; for i=1:option.len_windows data.data_x=[data.data_x,data.data(i:i+length(data.data(:,1))-option.len_windows-1,:)]; end i=i+Step; data.data_x(1:Step-1,:)=[]; data.data_y=(data.data(i:i+length(data.data(:,1))-option.len_windows-Step,:)); %% 归一化 [data.data_x1,data.data_x1_ps]=mapminmax(data.data_x'); data.data_x1=data.data_x1'; [data.data_y1,data.data_y1_ps]=mapminmax(data.data_y'); data.data_y1=data.data_y1'; %% 划分训练集和测试集 data.train_x=data.data_x1(1:ceil(length(data.data_x1(:,1))*option.p_train),:); data.test_x=data.data_x1(ceil(length(data.data_x1(:,1))*option.p_train)+1:end,:); data.train_y=data.data_y1(1:ceil(length(data.data_x1(:,1))*option.p_train),option.num_D); data.test_y=data.data_y1(ceil(length(data.data_x1(:,1))*option.p_train)+1:end,option.num_D); %% 实验参数 option.repeat_num=2; %重复实验10次，取平均值 %% 初始种群生成 for i=1:option.repeat_num %% 使用神经网络进行训练 tic net=newff(data.train_x',data.train_y',[20,20]); [net,tr]=train(net,data.train_x',data.train_y'); result.train_y_BP=sim(net,data.train_x'); result.test_y_BP=sim(net,data.test_x'); resulttrain.train_y_BP(i,:)=result.train_y_BP'; resulttest.test_y_BP(i,:)=result.test_y_BP'; time.time1(i,1)=toc; %% ELM神经网络预测 tic [IW,B,LW,TF,TYPE] = elmtrain2(data.train_x',data.train_y',100,'sig',0); result.train_y_ELM = elmpredict(data.train_x',IW,B,LW,TF,TYPE); result.test_y_ELM= elmpredict(data.test_x',IW,B,LW,TF,TYPE); resulttrain.train_y_ELM(i,:)=result.train_y_ELM'; resulttest.test_y_ELM(i,:)=result.test_y_ELM'; time.time2(i,:)=toc; %% 创建变量 %% 构建LSTM网络 %重新整理训练数据 tic for j=1:length(data.train_x(:,1)) data.train_x_LSTM{j}=data.train_x(j,:)'; data.train_y_LSTM{j}=data.train_y(j,:)'; end for k=1:length(data.test_x(:,1)) data.test_x_LSTM{k}=data.test_x(k,:)'; data.test_y_LSTM{k}=data.test_y(k,:)'; end numResponses = 1; numFeatures = length(data.train_x(1,:)); numHiddenUnits = 20; layers = [ ... sequenceInputLayer(numFeatures) lstmLayer(numHiddenUnits) fullyConnectedLayer(numResponses) regressionLayer]; options = trainingOptions('adam', ... 'MaxEpochs',250, ... 'GradientThreshold',1, ... 'InitialLearnRate',0.01, ... 'LearnRateSchedule','piecewise', ... 'LearnRateDropPeriod',125, ... 'LearnRateDropFactor',0.1, ... 'Verbose',0); %% net = trainNetwork(data.train_x_LSTM,data.train_y_LSTM,layers,options); [~,result.train_y_LSTM0 ] = predictAndUpdateState(net,data.train_x_LSTM); [~,result.test_y_LSTM0 ] = predictAndUpdateState(net,data.test_x_LSTM); result.train_y_LSTM=zeros(1,length(result.train_y_LSTM0)); for l=1:length(result.train_y_LSTM0) result.train_y_LSTM(l)=result.train_y_LSTM0{l}; end result.test_y_LSTM=zeros(1,length(result.test_y_LSTM0)); for m=1:length(result.test_y_LSTM0) result.test_y_LSTM(m)=result.test_y_LSTM0{m}; end resulttrain.train_y_LSTM(i,:)=result.train_y_LSTM'; resulttest.test_y_LSTM(i,:)=result.test_y_LSTM'; time.time3(i,:)=toc; %% 基于SAE的LSTM %SAE tic hiddenSize1 = 40; autoenc1 = trainAutoencoder(data.train_x_LSTM,hiddenSize1, ... 'MaxEpochs',500, ... 'EncoderTransferFunction','satlin',... 'DecoderTransferFunction','purelin',... 'L2WeightRegularization',0.001,... 'SparsityRegularization',4,... 'SparsityProportion',0.6); xReconstructed = predict(autoenc1,data.train_x_LSTM); numResponses = 1; numFeatures = length(data.train_x(1,:)); numHiddenUnits = 20; layers = [ ... sequenceInputLayer(numFeatures) lstmLayer(numHiddenUnits) fullyConnectedLayer(numResponses) regressionLayer]; options = trainingOptions('adam', ... 'MaxEpochs',250, ... 'GradientThreshold',1, ... 'InitialLearnRate',0.01, ... 'LearnRateSchedule','piecewise', ... 'LearnRateDropPeriod',125, ... 'LearnRateDropFactor',0.5, ... 'Verbose',0); net = trainNetwork(xReconstructed,[data.train_y_LSTM],layers,options); [~,result.train_y_SAE_LSTM0 ] = predictAndUpdateState(net,data.train_x_LSTM); [~,result.test_y_SAE_LSTM0 ] = predictAndUpdateState(net,data.test_x_LSTM); result.train_y_SAE_LSTM=zeros(1,length(result.train_y_SAE_LSTM0)); for l=1:length(result.train_y_SAE_LSTM0) result.train_y_SAE_LSTM(l)=result.train_y_SAE_LSTM0{l}; end result.test_y_SAE_LSTM=zeros(1,length(result.test_y_SAE_LSTM0)); for m=1:length(result.test_y_SAE_LSTM0) result.test_y_SAE_LSTM(m)=result.test_y_SAE_LSTM0{m}; end resulttrain.train_y_SAE_LSTM(i,:)=result.train_y_SAE_LSTM'; resulttest.test_y_SAE_LSTM(i,:)=result.test_y_SAE_LSTM'; time.time4(i,:)=toc; %% bilstm tic numResponses = 1; numFeatures = length(data.train_x(1,:)); numHiddenUnits = 20; layers = [ ... sequenceInputLayer(numFeatures) bilstmLayer(numHiddenUnits) fullyConnectedLayer(numResponses) regressionLayer]; options = trainingOptions('adam', ... 'MaxEpochs',250, ... 'GradientThreshold',1, ... 'InitialLearnRate',0.01, ... 'LearnRateSchedule','piecewise', ... 'LearnRateDropPeriod',125, ... 'LearnRateDropFactor',0.1, ... 'Verbose',0); net = trainNetwork(data.train_x_LSTM,data.train_y_LSTM,layers,options); [~,result.train_y_BiLSTM0 ] = predictAndUpdateState(net,data.train_x_LSTM); [~,result.test_y_BiLSTM0 ] = predictAndUpdateState(net,data.test_x_LSTM); result.train_y_BiLSTM=zeros(1,length(result.train_y_BiLSTM0)); for l=1:length(result.train_y_BiLSTM0) result.train_y_BiLSTM(l)=result.train_y_BiLSTM0{l}; end result.test_y_BiLSTM=zeros(1,length(result.test_y_BiLSTM0)); for m=1:length(result.test_y_BiLSTM0) result.test_y_BiLSTM(m)=result.test_y_BiLSTM0{m}; end resulttrain.train_y_BiLSTM(i,:)=result.train_y_BiLSTM'; resulttest.test_y_BiLSTM(i,:)=result.test_y_BiLSTM'; time.time5(i,:)=toc; end %% time.time1=mean(time.time1); time.time2=mean(time.time2); time.time3=mean(time.time3); time.time4=mean(time.time4); time.time5=mean(time.time5); resulttrain.train_y_BP=mean(resulttrain.train_y_BP); resulttrain.train_y_ELM=mean(resulttrain.train_y_ELM); resulttrain.train_y_LSTM=mean(resulttrain.train_y_LSTM); resulttrain.train_y_BiLSTM=mean(resulttrain.train_y_BiLSTM); resulttrain.train_y_SAE_LSTM=mean(resulttrain.train_y_SAE_LSTM); resulttest.test_y_BP=mean(resulttest.test_y_BP); resulttest.test_y_ELM=mean(resulttest.test_y_ELM); resulttest.test_y_LSTM=mean(resulttest.test_y_LSTM); resulttest.test_y_BiLSTM=mean(resulttest.test_y_BiLSTM); resulttest.test_y_SAE_LSTM=mean(resulttest.test_y_SAE_LSTM); data.train_y=mapminmax('reverse',data.train_y,data.data_y1_ps); data.test_y=mapminmax('reverse',data.test_y,data.data_y1_ps); resulttrain.train_y_BP=mapminmax('reverse',resulttrain.train_y_BP,data.data_y1_ps); resulttrain.train_y_ELM=mapminmax('reverse',resulttrain.train_y_ELM,data.data_y1_ps); resulttrain.train_y_LSTM=mapminmax('reverse',resulttrain.train_y_LSTM,data.data_y1_ps); resulttrain.train_y_BiLSTM=mapminmax('reverse',resulttrain.train_y_BiLSTM,data.data_y1_ps); resulttrain.train_y_SAE_LSTM=mapminmax('reverse',resulttrain.train_y_SAE_LSTM,data.data_y1_ps); resulttest.test_y_BP=mapminmax('reverse',resulttest.test_y_BP,data.data_y1_ps); resulttest.test_y_ELM=mapminmax('reverse',resulttest.test_y_ELM,data.data_y1_ps); resulttest.test_y_LSTM=mapminmax('reverse',resulttest.test_y_LSTM,data.data_y1_ps); resulttest.test_y_BiLSTM=mapminmax('reverse',resulttest.test_y_BiLSTM,data.data_y1_ps); resulttest.test_y_SAE_LSTM=mapminmax('reverse',resulttest.test_y_SAE_LSTM,data.data_y1_ps); %% [RMSE(1),MAE(1),U1(1),U2(1)]=draw(data.train_y',resulttrain.train_y_BP,'BP训练集'); [RMSE(2),MAE(2),U1(2),U2(2)]=draw(data.train_y',r