采用多层人工神经网络的能源消耗的时间序列预测（Matlab代码实现）

     目录

1 概述

2 运行结果

3 参考文献

‍4 Matlab代码

---

1 概述

该项目为能源消耗的时间序列预测，在Matlab中实现。该预测采用多层人工神经网络，基于Kaggle训练集预测未来能源消耗。

2 运行结果

 

 

 

 

 

3 参考文献

[1]程静,郑定成,吴继权.基于时间序列ARMA模型的广东省能源需求预测[J].能源工程,2010(01):1-5.DOI:10.16189/j.cnki.nygc.2010.01.012.

‍4 Matlab代码

seed = 52
    rng(seed);                                                                  % Seeds the random number generator using the nonnegative integer seed     

    % Load the data
    load lasertrain.dat 
    load laserpred.dat
    clear net
    close all

    p = 18;                                                                     % Lag    
    n = 4;                                                                      % Difference between the inputs and the number of neurons    

    [TrainData, TrainTarget] = getTimeSeriesTrainData(lasertrain, p);           % This creates TrainData and TrainTarget, such as divided in time  intervales of 5. TrainData(:,1) -> TrainTarget(1). R5 --> R1

    % Create the ANN
    numN = p - n;                                                               % Number of neurons in the hidden layer
    numE = 200;                                                                 % Number of epochs
    trainAlg = 'trainlm';                                                       % Training Algorithm 
    transFun = 'tansig';                                                        % Transfer Function   

    net = feedforwardnet(numN, trainAlg);                                       % Create general feedforward netowrk
    net.trainParam.epochs = numE;                                               % Set the number of epochs for the training 
    net.layers{1}.transferFcn = transFun;                                       % Set the Transfer Function
%     net.divideParam.trainRatio = 0.8;
%     net.divideParam.valRatio = 0.2;
%     net.divideParam.testRatio = 0;

    P = TrainData;       
    T = TrainTarget;     
    net = train(net,P,T);

    TotMat = [lasertrain ; zeros(size(laserpred))];

    for j = 1:length(laserpred) 

        r = length(lasertrain) - p + j;
        P = TotMat(r:r+p-1);

        a = sim(net,P);
        TotMat(length(lasertrain) + j) = a;
%         TotMat(length(lasertrain) + j) = (abs(round(a))+round(a))/2;

    end
    err = immse(TotMat(length(lasertrain)+1:end),laserpred);
    
    x = linspace(1,length(TotMat),length(TotMat));

figure
subplot(3,1,2)
plot(x(1:length(lasertrain)), TotMat(1:length(lasertrain)),'r', x(length(lasertrain)+1:end), TotMat(length(lasertrain)+1:end),'g', x(length(lasertrain)+1:end), laserpred','b');
xlim([1 1100])
title('Real set and forecast set');
legend('Training Data','Forecasted Data', 'Test data');

subplot(3,1,1)
plot(x(1:length(lasertrain)), TotMat(1:length(lasertrain)),'r', x(length(lasertrain):end), TotMat(length(lasertrain):end),'g');
xlim([1 1100])
title('Forecasted set');
legend('Training Data','Forecasted Data');

subplot(3,1,3)
plot(x(length(lasertrain)+1:end), TotMat(length(lasertrain)+1:end),'g', x(length(lasertrain)+1:end), laserpred','b');
xlim([1000 1100])
title('Detailed forcast data vs Test data');
legend('Forecasted Data', 'Test data');

formatSpec = 'The mean squared error is %4.2f \n';
fprintf(formatSpec, err)

主函数部分代码：

%% Exercise 2 - Recurrent Neural Networks

%% Initial stuff
clc
clearvars
close all

%% 1 - Hopfield Network

% Note: for this exercise, it will be interesting to check what happens if
% you add more neurons, i.e. modify T such as N is bigger.

T = [1 1; -1 -1; 1 -1]';                                                    % N x Q matrix containing Q vectors with components equal to +- 1.
                                                                            % 2-neuron network with 3 atactors
net = newhop(T);                                                            % Create a recurrent HOpfield network with stable points being the vectors from T

A1 = [0.3 0.6; -0.1 0.8; -1 0.5]';                                          % Example inputs
A2 = [-1 0.5 ; -0.5 0.1 ; -1 -1 ]';
A3 = [1 0.5  ; -0.3 -0.4 ; 0.8 -0.6]';
A4 = [0 -0.1 ; 0.1 0 ; -0.5 0.1]';
A5 = [0 0 ; 0 0.1  ; -0.1 0 ]';
A0 = [1 1; -1 -1; 1 -1]';
% Simulate a Hopfield network
% Y_1 = net([],[],Ai);                                                      % Single step iteration    

num_step = 20;                                                              % Number of steps                
Y_1 = net({num_step},{},A1);                                                % Multiple step iteration
Y_2 = net({num_step},{},A2);  
Y_3 = net({num_step},{},A3);  
Y_4 = net({num_step},{},A4);  
Y_5 = net({num_step},{},A5);

%% Now we try with 4 Neurons

T_ = [1 1 1 1; -1 -1 1 -1; 1 -1 1 1]';                                      % N x Q matrix containing Q vectors with components equal to +- 1.
                                                                            % 4-neuron network with 3 atactors
net_ = newhop(T_);                                                          % Create a recurrent HOpfield network with stable points being the vectors from T

A1_ = [0.3 0.6 0.3 0.6; -0.1 0.8 -0.1 0.8; -1 0.5 -1 0.5]';                 % Example inputs
A2_ = [-1 0.2 -1 0.2 ; -0.5 0.1  -0.5 0.1 ; -1 -1 -1 -1 ]';
A3_ = [1 0.5 1 0.5 ; -0.3 -0.4 -0.3 -0.4 ; 0.8 -0.6 0.8 -0.6]';
A4_ = [-0.5 -0.3 -0.5 -0.3 ; 0.1 0.8 0.1 0.8 ; -0.7 0.6 -0.7 0.6]';
% Simulate a Hopfield network
% Y_1 = net([],[],Ai);                                                      % Single step iteration    

num_step_ = 40;                                                             % Number of steps                
Y_1_ = net_({num_step_},{},A1_);                                            % Multiple step iteration
Y_2_ = net_({num_step_},{},A2_);  
Y_3_ = net_({num_step_},{},A3_);  
Y_4_ = net_({num_step_},{},A4_); 

%% Execute rep2.m || A script which generates n random initial points and visualises results of simulation of a 2d Hopfield network 'net'

% Note: The idea is to understand what is the relation between symmetry and
% attractors. It does not make sense that appears a fourth attractor when
% only 3 are in the Target T
close all