代码

1、

clc;
clear;
close all;
%% 导入数据
load DATA; % 导入60天的负荷数据
figure(1);
plot(fuhe);
legend('电力负荷');
xlabel('时间(15min)');
ylabel('功率(Kw)');
title('60日总负荷数据');
%% 数据归一化处理
% 归一化到0-1之间
Temp = fuhe;
[~,PS] = mapminmax(fuhe,0,1); 
Temp = mapminmax('apply',Temp,PS);
%% 输入输出构建及滑动时间窗输入结构构建
%数据分组
Windows = 8; %时间窗长度
T = 96; %预测未来一天(15min一次)
%各个输入集合大小
X_Train = cell(size(Temp,2)-T-Windows,1); % 训练数据
Y_Train = cell(size(Temp,2)-T-Windows,1); % 测试数据 
X_Test = cell(T,1);
Y_Test = cell(T,1);
% 训练数据构建
for i = 1:size(Temp,2)-T-Windows
    X_Train{
    
    i} = Temp(:,i:i+Windows-1)';
    Y_Train{
    
    i} = Temp(:,i+Windows);
end
% 测试数据构建
for i = 1:T
    X_Test{
    
    i} = Temp(:,end-T+i-Windows:end-T+i-1)';
    Y_Test{
    
    i} = Temp(:,end-T+i);
end
%% 定义神经网络
numFeatures = Windows; % 输入特征个数(等于滑动窗口大小)
numResponses = 1; % 输出特征个数
numHiddenUnits = 80; % 隐含层神经元个数
Train_number = 120; % 训练次数
dorp_rate = 0.2; % 遗忘率(可以防止过拟合，0.1表示遗忘30%)
[layer_lstm,layer_bilstm,options] = Net_definition(numFeatures,numResponses,numHiddenUnits,Train_number,dorp_rate); %网络定义函数
%% 网络训练
net_lstm = trainNetwork(X_Train,Y_Train,layer_lstm,options);
net_bilstm = trainNetwork(X_Train,Y_Train,layer_bilstm,options);
%% 网络测试
YPred_lstm = predict(net_lstm,X_Test); %lstm预测
YPred_bilstm = predict(net_bilstm,X_Test); %bi-lstm预测
% 反归一化
True = []; Predict_lstm = []; Predict_bilstm = [];
for i = 1:T
    True = [True,mapminmax('reverse',Y_Test{
    
    i},PS)];
    Predict_lstm = [Predict_lstm,mapminmax('reverse',YPred_lstm{
    
    i},PS)];
    Predict_bilstm = [Predict_bilstm,mapminmax('reverse',YPred_bilstm{
    
    i},PS)];
end
Predict_lstm = double(Predict_lstm);
Predict_bilstm = double(Predict_bilstm);
%% 误差分析
RMSE_lstm = sqrt(mean((True-Predict_lstm).^2));
RMSE_bilstm = sqrt(mean((True-Predict_bilstm).^2));
fprintf('LSTM的预测误差为：');
disp(RMSE_lstm);
fprintf('Bi-LSTM的预测误差为：');
disp(RMSE_bilstm);
%% 结果绘制
figure(2);
plot(True,'r-');
hold on
plot(Predict_lstm,'b-');
plot(Predict_bilstm,'g-');
hold off
legend('真实值','lstm预测','bi-lstm预测');
xlabel('时间(15min)');
ylabel('功率(Kw)');
title('一日内负荷预测情况');

figure(3)
subplot(2,1,1)
plot(True)
hold on
plot(Predict_lstm,'.-')
hold off
legend(['真实值' 'LSTM预测值'])
ylabel('Cases')
title('LSTM预测误差图像')
subplot(2,1,2)
stem(True - Predict_lstm)
xlabel('时间')
ylabel('误差')
title('RMSE = ' + RMSE_lstm)

figure(4)
subplot(2,1,1)
plot(True)
hold on
plot(Predict_bilstm,'.-')
hold off
legend(['真实值' 'Bi-LSTM预测值'])
ylabel('Cases')
title('Bi-LSTM预测误差图像')
subplot(2,1,2)
stem(True - Predict_bilstm)
xlabel('时间')
ylabel('误差')
title('RMSE = ' + RMSE_bilstm)

2、

function [layer_lstm,layer_bilstm,options] = Net_definition(numFeatures,numResponses,numHiddenUnits,Train_number,dorp_rate)
%% LSTM
layer_lstm = [ ...
    sequenceInputLayer(numFeatures)
    lstmLayer(numHiddenUnits,'OutputMode','sequence')
    dropoutLayer(dorp_rate)
    fullyConnectedLayer(numResponses)
    regressionLayer];
%% BI-LSTM
layer_bilstm = [ ...
    sequenceInputLayer(numFeatures)
    bilstmLayer(numHiddenUnits,'OutputMode','sequence')
    dropoutLayer(dorp_rate)
    fullyConnectedLayer(numResponses)
    regressionLayer];

%% 
options = trainingOptions('adam', ...
    'MaxEpochs',Train_number, ...
    'GradientThreshold',1, ...
    'InitialLearnRate',0.005, ...
    'LearnRateSchedule','piecewise', ...
    'LearnRateDropPeriod',Train_number/2, ...
    'LearnRateDropFactor',0.2, ...
    'Verbose',0, ...
    'Plots','training-progress');
end