[开源] 采用多种智能优化算法优化CNN_BiGRU_AM的时间序列预测模型matlab代码

整理了采用多种智能优化算法优化CNN_BiGRU_AM的时间序列预测模型matlab代码，免费分享给大家，记得点赞。值得注意的是该代码适用于新版本的matlab2023a及以上。

废话不多数,直接上链接。

免费获取链接：【免费】多种智能优化算法来优化CNN-BiLSTM-AM混合时间序列预测模型资源-CSDN文库

%%  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  

%% 导入数据
addpath(genpath(pwd));
data =  readmatrix('风电场预测.xlsx');
data = data(5665:8640,12);  %选取3月份数据
[h1,l1]=data_process(data,24);   %步长为24，采用前24个时刻的温度预测第25个时刻的温度
res = [h1,l1];
num_samples = size(res,1);   %样本个数


% 训练集和测试集划分
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


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);

%  格式转换
for i = 1 : M 
    vp_train{i, 1} = p_train(:, i);
    vt_train{i, 1} = t_train(:, i);
end

for i = 1 : N 
    vp_test{i, 1} = p_test(:, i);
    vt_test{i, 1} = t_test(:, i);
end
numFeatures = size(p_train,1);

%% 网络搭建CNN-BiGRU-ATTENTION
lgraph = layerGraph();

% 添加层分支
% 将网络分支添加到层次图中。每个分支均为一个线性层组。
tempLayers = sequenceInputLayer([numFeatures,1,1],"Name","sequence");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    convolution2dLayer([3,1],16,"Name","conv","Padding","same")
    batchNormalizationLayer("Name","batchnorm")
    reluLayer("Name","relu")
    maxPooling2dLayer([2 2],"Name","maxpool","Padding","same")
    
    convolution2dLayer([3,1],16,"Name","conv2","Padding","same")
    batchNormalizationLayer("Name","batchnorm2")
    reluLayer("Name","relu2")
    maxPooling2dLayer([2 2],"Name","maxpool2","Padding","same")

    flattenLayer("Name","flatten_1")
    fullyConnectedLayer(25,"Name","fc_1")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = flattenLayer("Name","flatten");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = gruLayer(35,"Name","gru1");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    FlipLayer("flip3")
    gruLayer(35,"Name","gru2")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    concatenationLayer(1,3,"Name","concat")
    selfAttentionLayer(1,50,"Name","selfattention")   %Attention机制
    fullyConnectedLayer(outdim,"Name","fc")
    regressionLayer("Name","regressionoutput")];
lgraph = addLayers(lgraph,tempLayers);

% 清理辅助变量
clear tempLayers;

% 连接层分支
% 连接网络的所有分支以创建网络图。
lgraph = connectLayers(lgraph,"sequence","conv");
lgraph = connectLayers(lgraph,"sequence","flatten");
lgraph = connectLayers(lgraph,"flatten","gru1");
lgraph = connectLayers(lgraph,"flatten","flip3");
lgraph = connectLayers(lgraph,"gru1","concat/in1");
lgraph = connectLayers(lgraph,"gru2","concat/in2");
lgraph = connectLayers(lgraph,"fc_1","concat/in3");


%%  参数设置
options = trainingOptions('adam', ...                 % 优化算法Adam
    'MaxEpochs', 30, ...                            % 最大训练次数
    'GradientThreshold', 1, ...                       % 梯度阈值
    'InitialLearnRate', 0.01, ...         % 初始学习率
    'LearnRateSchedule', 'piecewise', ...             % 学习率调整
    'LearnRateDropPeriod', 22, ...                   % 训练60次后开始调整学习率
    'LearnRateDropFactor',0.1, ...                    % 学习率调整因子
    'ExecutionEnvironment', 'cpu',...                 % 训练环境
    'Verbose', 1, ...                                 % 关闭优化过程
    'Plots', 'none');                    % 画出曲线

%  训练
tic
net = trainNetwork(vp_train, vt_train, lgraph, options);
toc
analyzeNetwork(net);% 查看网络结构
%  预测
t_sim1 = predict(net, vp_train); 
t_sim2 = predict(net, vp_test); 

%  数据反归一化
T_sim1 = mapminmax('reverse', t_sim1, ps_output);
T_sim2 = mapminmax('reverse', t_sim2, ps_output);
T_train1 = T_train;
T_test2 = T_test;

%  数据格式转换
T_sim1 = cell2mat(T_sim1);% cell2mat将cell元胞数组转换为普通数组
T_sim2 = cell2mat(T_sim2);

T_sim1 = T_sim1';
T_sim2 = T_sim2';

CNN_BiGRU_ATTENTION_TSIM1 = T_sim1;
CNN_BiGRU_ATTENTION_TSIM2 = T_sim2;
save CNN_BiGRU_ATTENTION CNN_BiGRU_ATTENTION_TSIM1 CNN_BiGRU_ATTENTION_TSIM2


% 指标计算
disp('…………训练集误差指标…………')
[mae1,rmse1,mape1,error1]=calc_error(T_train1,T_sim1);
fprintf('\n')

figure('Position',[200,300,600,200])
plot(T_train1);
hold on
plot(T_sim1)
legend('真实值','预测值')
title('CNN-BiGRU-ATTENTION训练集预测效果对比')
xlabel('样本点')
ylabel('发电功率')

disp('…………测试集误差指标…………')
[mae2,rmse2,mape2,error2]=calc_error(T_test2,T_sim2);
fprintf('\n')


figure('Position',[200,300,600,200])
plot(T_test2);
hold on
plot(T_sim2)
legend('真实值','预测值')
title('CNN-BiGRU-ATTENTION预测集预测效果对比')
xlabel('样本点')
ylabel('发电功率')

figure('Position',[200,300,600,200])
plot(T_sim2-T_test2)
title('CNN-BiGRU-ATTENTION误差曲线图')
xlabel('样本点')
ylabel('发电功率')



%% 优化CNN-BiGRU-Attention

disp(' ')
disp('优化CNN_BiLSTM_attention神经网络：')

%% 初始化参数 
popsize=10;   %初始种群规模 
maxgen=8;   %最大进化代数
fobj = @(x)objectiveFunction(x,numFeatures,outdim,vp_train,vt_train,vp_test,T_test,ps_output);
% 优化参数设置
lb = [0.001 10 2  2]; %参数的下限。分别是学习率，biGRU的神经元个数，注意力机制的键值, 卷积核大小
ub = [0.01 50 50 10];    %参数的上限
dim = length(lb);%数量

% 可选：'DBO','GWO','OOA','PSO','SABO','SCSO','SSA','BWO','RIME','WOA','HHO','NGO';

[Best_score,Best_pos,curve]=NGO(popsize,maxgen,lb,ub,dim,fobj); %修改这里的函数名字即可
setdemorandstream(pi);

%% 绘制进化曲线 
figure
plot(curve,'r-','linewidth',2)
xlabel('进化代数')
ylabel('均方误差')
legend('最佳适应度')
title('进化曲线')

%% 把最佳参数Best_pos回带 
[~,optimize_T_sim] = objectiveFunction(Best_pos,numFeatures,outdim,vp_train,vt_train,vp_test,T_test,ps_output);
setdemorandstream(pi);

%% 比较算法预测值 
str={'真实值','CNN-BiGRU-Attention','优化后CNN-BiGRU-Attention'};
figure('Units', 'pixels', ...
    'Position', [300 300 860 370]);
plot(T_test,'-','Color',[0.8500 0.3250 0.0980]) 
hold on
plot(T_sim2,'-.','Color',[0.4940 0.1840 0.5560]) 
hold on
plot(optimize_T_sim,'-','Color',[0.4660 0.6740 0.1880])
legend(str)
set (gca,"FontSize",12,'LineWidth',1.2)
box off
legend Box off


%% 比较算法误差
test_y = T_test;
Test_all = [];

y_test_predict = T_sim2;
[test_MAE,test_MAPE,test_MSE,test_RMSE,test_R2]=calc_error(y_test_predict,test_y);
Test_all=[Test_all;test_MAE test_MAPE test_MSE test_RMSE test_R2];


y_test_predict = optimize_T_sim;
[test_MAE,test_MAPE,test_MSE,test_RMSE,test_R2]=calc_error(y_test_predict,test_y);
Test_all=[Test_all;test_MAE test_MAPE test_MSE test_RMSE test_R2];
 	

str={'真实值','CNN-BiGRU-Attention','优化后CNN-BiGRU-Attention'};
str1=str(2:end);
str2={'MAE','MAPE','MSE','RMSE','R2'};
data_out=array2table(Test_all);
data_out.Properties.VariableNames=str2;
data_out.Properties.RowNames=str1;
disp(data_out)

%% 柱状图 MAE MAPE RMSE 柱状图适合量纲差别不大的
color=    [0.66669    0.1206    0.108
    0.1339    0.7882    0.8588
    0.1525    0.6645    0.1290
    0.8549    0.9373    0.8275   
    0.1551    0.2176    0.8627
    0.7843    0.1412    0.1373
    0.2000    0.9213    0.8176
      0.5569    0.8118    0.7882
       1.0000    0.5333    0.5176];
figure('Units', 'pixels', ...
    'Position', [300 300 660 375]);
plot_data_t=Test_all(:,[1,2,4])';
b=bar(plot_data_t,0.8);
hold on

for i = 1 : size(plot_data_t,2)
    x_data(:, i) = b(i).XEndPoints'; 
end

for i =1:size(plot_data_t,2)
    b(i).FaceColor = color(i,:);
    b(i).EdgeColor=[0.3353    0.3314    0.6431];
    b(i).LineWidth=1.2;
end

for i = 1 : size(plot_data_t,1)-1
    xilnk=(x_data(i, end)+ x_data(i+1, 1))/2;
    b1=xline(xilnk,'--','LineWidth',1.2);
    hold on
end 

ax=gca;
legend(b,str1,'Location','best')
ax.XTickLabels ={'MAE', 'MAPE', 'RMSE'};
set(gca,"FontSize",10,"LineWidth",1)
box off
legend box off

%% 二维图
figure
plot_data_t1=Test_all(:,[1,5])';
MarkerType={'s','o','pentagram','^','v'};
for i = 1 : size(plot_data_t1,2)
   scatter(plot_data_t1(1,i),plot_data_t1(2,i),120,MarkerType{i},"filled")
   hold on
end
set(gca,"FontSize",12,"LineWidth",2)
box off
legend box off
legend(str1,'Location','best')
xlabel('MAE')
ylabel('R2')
grid on


%% 雷达图
figure('Units', 'pixels', ...
    'Position', [150 150 520 500]);
Test_all1=Test_all./sum(Test_all);  %把各个指标归一化到一个量纲
Test_all1(:,end)=1-Test_all(:,end);
RC=radarChart(Test_all1);
str3={'MAE','MAPE','MSE','RMSE','R2'};
RC.PropName=str3;
RC.ClassName=str1;
RC=RC.draw(); 
RC.legend();
RC.setBkg('FaceColor',[1,1,1])
RC.setRLabel('Color','none')
colorList=[78 101 155;
          181 86 29;
          184 168 207;
          231 188 198;
          253 207 158;
          239 164 132;
          182 118 108]./255;

for n=1:RC.ClassNum
    RC.setPatchN(n,'Color',colorList(n,:),'MarkerFaceColor',colorList(n,:))
end



%%
figure('Units', 'pixels', ...
    'Position', [150 150 920 600]);
t = tiledlayout('flow','TileSpacing','compact');
for i=1:length(Test_all(:,1))
nexttile
th1 = linspace(2*pi/length(Test_all(:,1))/2,2*pi-2*pi/length(Test_all(:,1))/2,length(Test_all(:,1)));
r1 = Test_all(:,i)';
[u1,v1] = pol2cart(th1,r1);
M=compass(u1,v1);
for j=1:length(Test_all(:,1))
    M(j).LineWidth = 2;
    M(j).Color = colorList(j,:);

end   
title(str2{i})
set(gca,"FontSize",10,"LineWidth",1)
end
 legend(M,str1,"FontSize",10,"LineWidth",1,'Box','off','Location','southoutside')