基于多输入的时间序列数据LSTM预测的MATLAB代码

%读取数据 clc clear close all clear global ver rng(2) %读取double格式数据 [num1,ax,ay ]= xlsread('1994-2020数据信息.xlsx',1); num= num1(:,2:5); m = round(0.8*length(num)); n = randperm(length(num)); n = 1:27; mmm=m; pan=3; input_train =num(n(1:m),1:pan)';%训练数据输入数据 output_train = num(n(1:m),pan+1)';%训练数据输出数据 input_test = num(m+1:end,1:pan)';%测试数据输入数据 output_test = num(m+1:end,pan+1)';%测试数据输出数据 [inputn,inputps]=mapminmax(input_train,-0.5,1);%训练数据的输入数据的归一化 [outputn,outputps]=mapminmax(output_train,-0.5,1);%训练数据的输出数据的归一化de inputn_test=mapminmax('apply',input_test,inputps); [lstm_pred,lstmerror1] = lstmfun(num,inputn,outputn,inputn_test,output_test,pan,outputps); [AllSamInn,minAllSamIn,maxAllSamIn,AllSamOutn,minAllSamOut,maxAllSamOut]=premnmx(input_train,output_train); EvaSamIn=input_test; EvaSamInn=tramnmx(EvaSamIn,minAllSamIn,maxAllSamIn); % preprocessing Ptrain = AllSamInn; Ttrain = AllSamOutn; % Initialize PSO vmax=0.0151; % Maximum velocity minerr=0.001; % Minimum error wmax=0.90; wmin=0.30; % global itmax; %Maximum iteration number itmax=10; c1=2; c2=2; for iter=1:itmax W(iter)=wmax-((wmax-wmin)/itmax)*iter; % weight declining linearly end %Between (m,n), (which can also be started from zero) m=-1; n=1; % global N; % number of particles N=2; % global D; % length of particle D=4; gbests = [100 100 100 0.2];% % particles are initialized between (a,b) randomly a=[200 200 200 0.98]; b=[10 10 10 0.05]; % Initialize positions of particles % rand('state',sum(100*clock)); X = []; for ii = N X =[X;a+(b-a).*rand(N,D,1)]; %取值范围[-1,1] rand * 2 - 1 ,rand 产生[0,1]之间的随机数 end %Initialize velocities of particles V=0.2*(m+(n-m)*rand(N,D,1)); % % global fvrec; MinFit=[]; BestFit=[]; fitness=fitcal(X,num,inputn,outputn,inputn_test,output_test,pan,outputps); fvrec(:,1,1)=fitness(:,1,1); [C,I]=min(fitness(:,1,1)); MinFit=[MinFit C]; BestFit=[BestFit C]; L(:,1,1)=fitness(:,1,1); %record the fitness of particle of every iterations B(1,1,1)=C; %record the minimum fitness of particle gbest(1,:,1)=X(I,:,1); %the global best x in population %Matrix composed of gbest vector for p=1:N G(p,:,1)=gbest(1,:); end for ii=1:N pbest(ii,:,1)=X(ii,:); end V(:,:,2)=W(1)*V(:,:,1)+c1*rand*(pbest(:,:,1)-X(:,:,1))+c2*rand*(G(:,:,1)-X(:,:,1)); for ni=1:N for di=1:D if V(ni,di,2)>vmax V(ni,di,2)=vmax; elseif V(ni,di,2)<-vmax V(ni,di,2)=-vmax; else V(ni,di,2)=V(ni,di,2); end end end X(:,:,2)=X(:,:,1)+V(:,:,2); for ni=1:N for di=1:D if X(ni,di,2)>1 X(ni,di,2)=1; elseif X(ni,di,2)<-1 X(ni,di,2)=-1; else X(ni,di,2)=X(ni,di,2); end end end %****************************************************** for jj=2:itmax disp('Iteration and Current Best Fitness') disp(jj-1) disp(B(1,1,jj-1)) reset =1; % reset = 1时设置为粒子群过分收敛时将其打散，如果＝1则不打散 if reset==1 bit = 1; for k=1:N bit = bit&(range(X(k,:))<0.02); end if bit==1 % bit=1时对粒子位置及速度进行随机重置 for ik = 1:N X(ik,:) = funx; % present 当前位置,随机初始化 X(ik,:) = [0.02*rand()-0.01 0.02*rand()-0.01]; % 速度初始化 end end end % Calculation of new positions fitness=fitcal(X,num,inputn,outputn,inputn_test,output_test,pan,outputps); [C,I]=min(fitness(:,1,jj)); MinFit=[MinFit C]; BestFit=[BestFit min(MinFit)]; L(:,1,jj)=fitness(:,1,jj); B(1,1,jj)=C; gbest(1,:,jj)=X(I,:,jj); [C,I]=min(B(1,1,:)); % keep gbest is the best particle of all have occured if B(1,1,jj)<=C gbest(1,:,jj)=gbest(1,:,jj); else gbest(1,:,jj)=gbest(1,:,I); end if C<=minerr break end %Matrix composed of gbest vector if jj>=itmax break end for p=1:N G(p,:,jj)=gbest(1,:,jj); end for ii=1:N [C,I]=min(L(ii,1,:)); if L(ii,1,jj)<=C pbest(ii,:,jj)=X(ii,:,jj); else pbest(ii,:,jj)=X(ii,:,I); end end V(:,:,jj+1)=W(jj)*V(:,:,jj)+c1*rand*(pbest(:,:,jj)-X(:,:,jj))+c2*rand*(G(:,:,jj)-X(:,:,jj)); for ni=1:N for di=1:D if V(ni,di,jj+1)>vmax V(ni,di,jj+1)=vmax; elseif V(ni,di,jj+1)<-vmax V(ni,di,jj+1)=-vmax; else V(ni,di,jj+1)=V(ni,di,jj+1); end end end X(:,:,jj+1)=X(:,:,jj)+V(:,:,jj+1); for ni=1:N for di=1:D if X(ni,di,jj+1)>1 X(ni,di,jj+1)=1; elseif X(ni,di,jj+1)<-1 X(ni,di,jj+1)=-1; else X(ni,di,jj+1)=X(ni,di,jj+1); end end end end disp('Iteration and Current Best Fitness') disp(jj) disp(B(1,1,jj)) disp('Global Best Fitness and Occurred Iteration') [C,I]=min(B(1,1,:)); numFeatures = size(num(:,1:pan),2);%输入层维度 numResponses = size(num(:,end),2);%输出维度 % fitval = zeros(nn,1); x= gbest; x(:,1:3) = round( gbest(:,1:3)); numHiddenUnits = x(1,1);%第一层维度 % a fully connected layer of size 50 & a dropout layer with dropout probability 0.5 layers = [ ... sequenceInputLayer(numFeatures)%输入层 lstmLayer(numHiddenUnits,'OutputMode','sequence')%第一层 fullyConnectedLayer(x(1,2))%链接层 dropoutLayer(x(1,4))%遗忘层 fullyConnectedLayer(numResponses)%链接层 regressionLayer];%回归层 % Specify the training options. % Train for 60 epochs with mini-batches of size 20 using the solver 'adam' maxEpochs =60;%最大迭代次数 miniBatchSize = x(1,3);%最小批量 % the learning rate == 0.01 % set the gradient threshold to 1 % set 'Shuffle' to 'never' options = trainingOptions('adam', ... %解算器 'MaxEpochs',maxEpochs, ... %最大迭代次数 'MiniBatchSize',miniBatchSize, ... %最小批次 'InitialLearnRate',0.01, ... %初始学习率 'GradientThreshold',inf, ... %梯度阈值 'Shuffle','never', ... %打乱顺序every-epoch 'Plots','none',... %画图 'Verbose',0); %不输出训练过程 %% Train the Network net = trainNetwork(inputn,outputn,layers,options);%开始训练 %% Test the Network plolstm_pred0 = predict(net,[inputn inputn_test],'MiniBatchSize',x(1,3))';%测试仿真输出 plolstm_pred=(mapminmax('reverse', plolstm_pred0(mmm+1:end),outputps))'; error1 = plolstm_pred-output_test;%误差 wucha = mean(abs(error1)); % toc figure(1) grid hold on plot((BestFit),'r'); title(['粒子群算法优化lstm ' '最优代数＝' I]); xlabel('进化代数');ylabel('误差'); disp('适应度变量'); figure(2) grid plot(plolstm_pred,'-^g') hold on plot(lstm_pred,'-ob') hold on plot(output_test,'-*r'); legend('粒子群优化lstm预测输出','lstm预测输出','期望输出') title('粒子群优化lstm网络预测输出','fontsize',12) ylabel('函数输出','fontsize',12) xlabel('样本','fontsize',12) [psolstmMSE, psolstmRMSE, psolstmMBE, psolstmMAE ] =MSE_RMSE_MBE_MAE(output_test,plolstm_pred) [lstmMSE, lstmRMSE, lstmMBE, lstmMAE ] =MSE_RMSE_MBE_MAE(output_test,lstm_pred) psolstmR = R_2(output_test,plolstm_pred) psolstmR = R_2(output_test,lstm_pred)