机器学习之MATLAB代码--IWOA_BILSTM(基于改进鲸鱼算法优化的BiLSTM预测算法)（十六）

代码

1、

%% 基于改进鲸鱼算法优化的BiLSTM预测算法 
clear;close all; 
clc
rng('default')
%% 读取负荷数据
load('QLD1.mat')
data = QLD1(1:2000);
%序列的前 90% 用于训练，后 10% 用于测试
numTimeStepsTrain = floor(0.9*numel(data));
dataTrain = data(1:numTimeStepsTrain+1)';
dataTest = data(numTimeStepsTrain+1:end)';
%数据预处理，将训练数据标准化为具有零均值和单位方差。
mu = mean(dataTrain);
sig = std(dataTrain);
dataTrainStandardized = (dataTrain - mu) / sig;
%输入BiLSTM的时间序列交替一个时间步
XTrain = dataTrainStandardized(1:end-1);
YTrain = dataTrainStandardized(2:end);
%数据预处理，将测试数据标准化为具有零均值和单位方差。
mu = mean(dataTest);
sig = std(dataTest);

dataTestStandardized = (dataTest - mu) / sig;
XTest = dataTestStandardized(1:end-1);
YTest = dataTestStandardized(2:end);
%%
%创建BiLSTM回归网络，指定BiLSTM层的隐含单元个数96*3
%序列预测，因此，输入一维，输出一维
numFeatures = 1;
numResponses = 1;
%% 定义改进鲸鱼算法优化参数
pop=5; %种群数量
Max_iteration=10; %  设定最大迭代次数
dim = 4;%维度，即BiLSTM网路包含的隐藏单元数目,最大训练周期,初始学习率,L2参数
lb = [20,50,10E-5,10E-6];%下边界
ub = [200,300,0.1,0.1];%上边界
fobj = @(x) fun(x,numFeatures,numResponses,XTrain,YTrain,XTest,YTest);

[Best_score,Best_pos,IWOA_curve,netIWOA,pos_curve]=IWOA(pop,Max_iteration,lb,ub,dim,fobj); %开始优化

figure
plot(IWOA_curve,'linewidth',1.5);
grid on
xlabel('迭代次数')
ylabel('适应度函数')
title('IWOA-BiLSTM适应度值曲线')

figure
subplot(221)
plot(pos_curve(:,1),'linewidth',1.5);
grid on
xlabel('迭代次数')
ylabel('隐藏单元数目')
title('隐藏单元数目迭代曲线')
subplot(222)
plot(pos_curve(:,2),'linewidth',1.5);
grid on
xlabel('迭代次数')
ylabel('训练周期')
title('训练周期迭代曲线')
subplot(223)
plot(pos_curve(:,3),'linewidth',1.5);
grid on
xlabel('迭代次数')
ylabel('学习率')
title('学习率迭代曲线')
subplot(224)
plot(pos_curve(:,4),'linewidth',1.5);
grid on
xlabel('迭代次数')
ylabel('L2参数')
title('L2参数迭代曲线')
%训练集测试
PredictTrainIWOA = predict(netIWOA,XTrain, 'ExecutionEnvironment','gpu');
%测试集测试
PredictTestIWOA = predict(netIWOA,XTest, 'ExecutionEnvironment','gpu');
%训练集mse
mseTrainIWOA= mse(YTrain-PredictTrainIWOA);
%测试集mse
mseTestIWOA = mse(YTest-PredictTestIWOA);
%% IWOA-BiLSTM优化参数
numHiddenUnits = round(Best_pos(1));%BiLSTM网路包含的隐藏单元数目
maxEpochs = round(Best_pos(2));%最大训练周期
InitialLearnRate = Best_pos(3);%初始学习率
L2Regularization = Best_pos(4);%L2参数
%设置网络
layers = [ ...
    sequenceInputLayer(numFeatures)
    bilstmLayer(numHiddenUnits)
    fullyConnectedLayer(numResponses)
    regressionLayer];
%指定训练选项
options = trainingOptions('adam', ...
    'MaxEpochs',maxEpochs, ...
    'ExecutionEnvironment' ,'gpu',...
    'InitialLearnRate',InitialLearnRate,...
    'GradientThreshold',1, ...
    'L2Regularization',L2Regularization, ...
    'Plots','training-progress',...
    'Verbose',0);
%训练BiLSTM
[net,info] = trainNetwork(XTrain,YTrain,layers,options);
%% 训练过程识别准确度曲线
figure;
plot(info.TrainingRMSE,'Color',[0 0.5 1] );
ylabel('TrainingRMSE')
xlabel('Training Step');
title(['训练集均方根值']);
%% 训练过程损失值曲线
figure;
plot(info.TrainingLoss,'Color',[1 0.5 0] );
ylabel('Training Loss')
xlabel('Training Step');
title(['损失函数值' ]);
%% 基础BiLSTM测试
numHiddenUnits = 50;
layers = [ ...
    sequenceInputLayer(numFeatures)
    bilstmLayer(numHiddenUnits)
    fullyConnectedLayer(numResponses)
    regressionLayer];
%指定训练选项
options = trainingOptions('adam', ...
    'MaxEpochs',50, ...
    'ExecutionEnvironment' ,'gpu',...
    'GradientThreshold',1, ...
    'InitialLearnRate',0.001, ...
    'L2Regularization',0.0001,...
    'Plots','training-progress',...
    'Verbose',1);
%训练BiLSTM
net = trainNetwork(XTrain,YTrain,layers,options);
%训练集测试
PredictTrain = predict(net,XTrain, 'ExecutionEnvironment','gpu');
%测试集测试
PredictTest = predict(net,XTest, 'ExecutionEnvironment','gpu');
%训练集mse
mseTrain = mse(YTrain-PredictTrain);
%测试集mse
mseTest = mse(YTest-PredictTest);

disp('-------------------------------------------------------------')
disp('IWOA-BiLSTM优化得到的最优参数为：')
disp(['IWOA-BiLSTM优化得到的隐藏单元数目为：',num2str(round(Best_pos(1)))]);
disp(['IWOA-BiLSTM优化得到的最大训练周期为：',num2str(round(Best_pos(2)))]);
disp(['IWOA-BiLSTM优化得到的InitialLearnRate为：',num2str((Best_pos(3)))]);
disp(['IWOA-BiLSTM优化得到的L2Regularization为：',num2str((Best_pos(4)))]);
disp('-------------------------------------------------------------')
disp('IWOA-BiLSTM结果：')
disp(['IWOA-BiLSTM训练集MSE:',num2str(mseTrainIWOA)]);
disp(['IWOA-BiLSTM测试集MSE:',num2str(mseTestIWOA)]);
disp('BiLSTM结果：')
disp(['BiLSTM训练集MSE:',num2str(mseTrain)]);
disp(['BiLSTM测试集MSE:',num2str(mseTest)]);

%% 训练集结果绘图
errors=YTrain-PredictTrain;
errorsIWOA=YTrain-PredictTrainIWOA;

MSE=mean(errors.^2);
RMSE=sqrt(MSE);
MSEIWOA=mean(errorsIWOA.^2);
RMSEIWOA=sqrt(MSEIWOA);

error_mean=mean(errors);
error_std=std(errors);

error_meanIWOA=mean(errorsIWOA);
error_stdIWOA=std(errorsIWOA);

figure;
plot(YTrain,'k');
hold on;
plot(PredictTrain,'b');
plot(PredictTrainIWOA,'r');
legend('Target','BiLSTM','IWOA-BiLSTM');
title('训练集结果');
xlabel('Sample Index');
grid on;

figure;
plot(errors);
hold on
plot(errorsIWOA);
legend('BiLSTM-Error','IWOA-BiLSTM-Eoor');
title({
    
    '训练集预测误差对比';['MSE = ' num2str(MSE), ', IWOA-MSE = ' num2str(MSEIWOA)]});
grid on;

figure;
histfit(errorsIWOA, 50);
title(['Error Mean = ' num2str(error_mean), ', Error St.D. = ' num2str(error_std)]);
%% 测试集结果绘图
errors=YTest-PredictTest;
errorsIWOA=YTest-PredictTestIWOA;

MSE=mean(errors.^2);
RMSE=sqrt(MSE);
MSEIWOA=mean(errorsIWOA.^2);
RMSEIWOA=sqrt(MSEIWOA);

error_mean=mean(errors);
error_std=std(errors);

error_meanIWOA=mean(errorsIWOA);
error_stdIWOA=std(errorsIWOA);

figure;
plot(YTest,'k');
hold on;
plot(PredictTest,'b');
plot(PredictTestIWOA,'r');
legend('Target','BiLSTM','IWOA-BiLSTM');
title('测试集结果');
xlabel('Sample Index');
grid on;

figure;
plot(errors);
hold on
plot(errorsIWOA);
legend('BiLSTM-Error','IWOA-BiLSTM-Eoor');
title({
    
    '测试集预测误差对比';['MSE = ' num2str(MSE), ', IWOA-MSE = ' num2str(MSEIWOA)]});
grid on;

figure;
histfit(errorsIWOA, 50);
title(['Error Mean = ' num2str(error_mean) ', Error St.D. = ' num2str(error_std)]);

2、

%% [1]武泽权,牟永敏.一种改进的鲸鱼优化算法[J].计算机应用研究,2020,37(12):3618-3621.
function [Leader_score,Leader_pos,Convergence_curve,BestNet,pos_curve]=IWOA(SearchAgents_no,Max_iter,lb,ub,dim,fobj)
% initialize position vector and score for the leader
net = {
    
    };
Leader_pos=zeros(1,dim);
Leader_score=inf; %change this to -inf for maximization problems
%% 改进点：准反向初始化
Positions=initializationNew(SearchAgents_no,dim,ub,lb,fobj);
Convergence_curve=zeros(1,Max_iter);
t=0;% Loop counter
% Main loop
while t<Max_iter
    for i=1:size(Positions,1)
        % Return back the search agents that go beyond the boundaries of the search space
        Flag4ub=Positions(i,:)>ub;
        Flag4lb=Positions(i,:)<lb;
        Positions(i,:)=(Positions(i,:).*(~(Flag4ub+Flag4lb)))+ub.*Flag4ub+lb.*Flag4lb;
        % Calculate objective function for each search agent
        
%         fitness=fobj(Positions(i,:));
        [fitness,net] =  fobj(Positions(i,:));
        % Update the leader
        if fitness<Leader_score % Change this to > for maximization problem
            Leader_score=fitness; % Update alpha
            Leader_pos=Positions(i,:);
        end
    end
    BestNet = net;
    %% 改进点：非线性收敛因子
    a=2 - sin(t*pi/(2*Max_iter) + 0);
    % a2 linearly dicreases from -1 to -2 to calculate t in Eq. (3.12)
    a2=-1+t*((-1)/Max_iter);
    %% 改进点：自适应权重
    w = 1 - (exp(t/Max_iter) - 1)/(exp(1) -1);
    % Update the Position of search agents
    for i=1:size(Positions,1)
        r1=rand(); % r1 is a random number in [0,1]
        r2=rand(); % r2 is a random number in [0,1]
        
        A=2*a*r1-a;  % Eq. (2.3) in the paper
        C=2*r2;      % Eq. (2.4) in the paper
        
        b=1;               %  parameters in Eq. (2.5)
        l=(a2-1)*rand+1;   %  parameters in Eq. (2.5)
        
        p = rand();        % p in Eq. (2.6)
        for j=1:size(Positions,2)
            if p<0.5
                if abs(A)>=1
                    rand_leader_index = floor(SearchAgents_no*rand()+1);
                    X_rand = Positions(rand_leader_index, :);
                    D_X_rand=abs(C*X_rand(j)-Positions(i,j)); % Eq. (2.7)
                    Positions(i,j)=w*X_rand(j)-A*D_X_rand;      % 引入权重
                elseif abs(A)<1
                    D_Leader=abs(C*Leader_pos(j)-Positions(i,j)); % Eq. (2.1)
                    Positions(i,j)=w*Leader_pos(j)-A*D_Leader;       % 引入权重
                end
            elseif p>=0.5
                distance2Leader=abs(Leader_pos(j)-Positions(i,j));
                % Eq. (2.5)
                Positions(i,j)=distance2Leader*exp(b.*l).*cos(l.*2*pi)+w*Leader_pos(j);  % 引入权重
            end
        end
        %边界处理
        Flag4ub=Positions(i,:)>ub;
        Flag4lb=Positions(i,:)<lb;
        Positions(i,:)=(Positions(i,:).*(~(Flag4ub+Flag4lb)))+ub.*Flag4ub+lb.*Flag4lb;
        %% 改进点：随机差分变异
        Rindex = randi(SearchAgents_no);%随机选择一个个体
        r1 = rand; r2 = rand;
        Temp = r1.*(Leader_pos - Positions(i,:)) + r2.*(Positions(Rindex,:) -  Positions(i,:));
        Flag4ub=Temp>ub;
        Flag4lb=Temp<lb;
        Temp=(Temp.*(~(Flag4ub+Flag4lb)))+ub.*Flag4ub+lb.*Flag4lb;
        if fobj(Temp) < fobj(Positions(i,:))
            Positions(i,:) = Temp;
        end
    end
    t=t+1;
    Convergence_curve(t)=Leader_score;
    pos_curve(t,:)=Leader_pos;
    fprintf(1,'%g\n',t);
end

3、

%% 基于准反向策略的种群初始化
function Positions=initializationNew(SearchAgents_no,dim,ub,lb,fun)
Boundary_no= size(ub,2); % numnber of boundaries
BackPositions = zeros(SearchAgents_no,dim);
if Boundary_no==1
    PositionsF=rand(SearchAgents_no,dim).*(ub-lb)+lb;
    %求取反向种群
    BackPositions = ub + lb - PositionsF;
end

% If each variable has a different lb and ub
if Boundary_no>1
    for i=1:dim
        ub_i=ub(i);
        lb_i=lb(i);
        PositionsF(:,i)=rand(SearchAgents_no,1).*(ub_i-lb_i)+lb_i;
        %求取反向种群
        BackPositions(:,i) =  (ub_i+lb_i) - PositionsF(:,i);
    end
end
%% 准反向操作
for i = 1:SearchAgents_no
    for j = 1:dim
        if Boundary_no==1
            if (ub + lb)/2 <BackPositions(i,j)
                Lb = (ub + lb)/2;
                Ub = BackPositions(i,j);
                PBackPositions(i,j) = (Ub - Lb)*rand + Lb;
            else
                Lb = BackPositions(i,j);
                Ub =  (ub + lb)/2;
                PBackPositions(i,j) = (Ub - Lb)*rand + Lb;
            end
        else
            if (ub(j) + lb(j))/2 <BackPositions(i,j)
                Lb = (ub(j) + lb(j))/2;
                Ub = BackPositions(i,j);
                PBackPositions(i,j) = (Ub - Lb)*rand + Lb;
            else
                Lb = BackPositions(i,j);
                Ub = (ub(j) + lb(j))/2;
                PBackPositions(i,j) = (Ub - Lb)*rand + Lb;
            end
        end
    end
end
%合并种群
AllPositionsTemp = [PositionsF;PBackPositions];
AllPositions = AllPositionsTemp;
for i = 1:size(AllPositionsTemp,1)
    %    fitness(i) = fun(AllPositionsTemp(i,:));
    [fitness(i),net{
    
    i}] =  fun(AllPositionsTemp(i,:));
    fprintf(1,'%g\n',i);
end
[fitness, index]= sort(fitness);%排序
for i = 1:2*SearchAgents_no
    AllPositions(i,:) = AllPositionsTemp(index(i),:);
end
%取适应度排名靠前的作为种群的初始化
Positions = AllPositions(1:SearchAgents_no,:);

end

4、

%适应度函数
%mse作为适应度值
function [fitness,net] = fun(x,numFeatures,numResponses,XTrain,YTrain,XTest,YTest)

disp('进行一次训练中....')
%% 获取优化参数
numHiddenUnits = round(x(1));%BiLSTM网路包含的隐藏单元数目
maxEpochs = round(x(2));%最大训练周期
InitialLearnRate = x(3);%初始学习率
L2Regularization = x(4);%L2参数

%设置网络
layers = [ ...
    sequenceInputLayer(numFeatures)
    bilstmLayer(numHiddenUnits)
    fullyConnectedLayer(numResponses)
    regressionLayer];

%指定训练选项，采用cpu训练， 这里用cpu是为了保证能直接运行，如果需要gpu训练，改成gpu就行了，且保证cuda有安装
options = trainingOptions('adam', ...
    'MaxEpochs',maxEpochs, ...
    'ExecutionEnvironment' ,'gpu',...
    'InitialLearnRate',InitialLearnRate,...
    'GradientThreshold',1, ...
    'L2Regularization',L2Regularization, ...
    'Verbose',0);
%'Plots','training-progress'
%训练LSTM
net = trainNetwork(XTrain,YTrain,layers,options);
%训练集测试
PredictTrain = predict(net,XTrain, 'ExecutionEnvironment','gpu');
%测试集测试
PredictTest = predict(net,XTest, 'ExecutionEnvironment','gpu');

%训练集mse
mseTrain = mse(YTrain-PredictTrain);
%测试集mse
mseTest = mse(YTest-PredictTest);

%% 测试集准确率

fitness =mseTrain+mseTest;
disp('训练结束....')
end

数据

结果