"Analysis of 43 cases of MATLAB neural network": Chapter 40 research on prediction of dynamic neural network time series -- implementation of NARX based on MATLAB

1. Preface

《MATLAB neural network 43 A case study 》 yes MATLAB Technology Forum （www.matlabsky.com） planning , Led by teacher wangxiaochuan ,2013 Beijing University of Aeronautics and Astronautics Press MATLAB A book for tools MATLAB Example teaching books , Is in 《MATLAB neural network 30 A case study 》 On the basis of modification 、 Complementary , Adhering to “ Theoretical explanation — case analysis — Application extension ” This feature , Help readers to be more intuitive 、 Learn neural networks vividly .

《MATLAB neural network 43 A case study 》 share 43 Chapter , The content covers common neural networks （BP、RBF、SOM、Hopfield、Elman、LVQ、Kohonen、GRNN、NARX etc. ） And related intelligent algorithms （SVM、 Decision tree 、 Random forests 、 Extreme learning machine, etc ）. meanwhile , Some chapters also cover common optimization algorithms （ Genetic algorithm (ga) 、 Ant colony algorithm, etc ） And neural network . Besides ,《MATLAB neural network 43 A case study 》 It also introduces MATLAB R2012b New functions and features of neural network toolbox in , Such as neural network parallel computing 、 Custom neural networks 、 Efficient programming of neural network, etc .

In recent years, with the rise of artificial intelligence research , The related direction of neural network has also ushered in another upsurge of research , Because of its outstanding performance in the field of signal processing , The neural network method is also being applied to various applications in the direction of speech and image , This paper combines the cases in the book , It is simulated and realized , It's a relearning , I hope I can review the old and know the new , Strengthen and improve my understanding and practice of the application of neural network in various fields . I just started this book on catching more fish , Let's start the simulation example , Mainly to introduce the source code application examples in each chapter , This paper is mainly based on MATLAB2015b(32 position ) Platform simulation implementation , This is the research example of dynamic neural network time series prediction in Chapter 40 of this book , Don't talk much , Start ！

2. MATLAB Simulation example

open MATLAB, Click on “ Home page ”, Click on “ open ”, Find the sample file

Choose chapter40.m, Click on “ open ”

chapter40.m Source code is as follows ：

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% function ： Research on dynamic neural network time series prediction - be based on MATLAB Of NARX Realization 
% Environmental Science ：Win7,Matlab2015b
%Modi: C.S
% Time ：2022-06-21
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%% Matlab neural network 43 A case study 

%  Research on dynamic neural network time series prediction - be based on MATLAB Of NARX Realization 
% by  Wang Xiao Chuan (@ Wang Xiao Chuan _matlab)
% http://www.matlabsky.com
% Email:[email protected]163.com
% http://weibo.com/hgsz2003

%%  Clear environment variables 
clear
clc
tic
%%  Load data 
% load phdata
[phInputs,phTargets] = ph_dataset;
inputSeries = phInputs;
targetSeries = phTargets;

%%  Establish a nonlinear autoregressive model 
inputDelays = 1:2;
feedbackDelays = 1:2;
hiddenLayerSize = 10;
net = narxnet(inputDelays,feedbackDelays,hiddenLayerSize);

%%  Definition of network data preprocessing function 
net.inputs{
    1}.processFcns = {
    'removeconstantrows','mapminmax'};
net.inputs{
    2}.processFcns = {
    'removeconstantrows','mapminmax'};

%%  Time series data preparation 
[inputs,inputStates,layerStates,targets] = preparets(net,inputSeries,{
    },targetSeries);

%%  Training data 、 Validation data 、 Test data division 
net.divideFcn = 'dividerand';  
net.divideMode = 'value';  
net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

%%  Network training function setting 
net.trainFcn = 'trainlm';  % Levenberg-Marquardt

%%  Error function setting 
net.performFcn = 'mse';  % Mean squared error

%%  Drawing function settings 
net.plotFcns = {
    'plotperform','plottrainstate','plotresponse', ...
  'ploterrcorr', 'plotinerrcorr'};

%%  Network training 
[net,tr] = train(net,inputs,targets,inputStates,layerStates);

%%  Network testing 
outputs = net(inputs,inputStates,layerStates);
errors = gsubtract(targets,outputs);
performance = perform(net,targets,outputs)

%%  Calculate the training set 、 Verification set 、 Test set error 
trainTargets = gmultiply(targets,tr.trainMask);
valTargets = gmultiply(targets,tr.valMask);
testTargets = gmultiply(targets,tr.testMask);
trainPerformance = perform(net,trainTargets,outputs)
valPerformance = perform(net,valTargets,outputs)
testPerformance = perform(net,testTargets,outputs)

%%  Visualization of network training effect 
figure, plotperform(tr)
figure, plottrainstate(tr)
figure, plotregression(targets,outputs)
figure, plotresponse(targets,outputs)
figure, ploterrcorr(errors)
figure, plotinerrcorr(inputs,errors)


%% close loop The realization of pattern 
%  change NARX Neural network mode 
narx_net_closed = closeloop(net);
view(net)
view(narx_net_closed)

%  Calculation 1500-2000 Fitting effect of points 
phInputs_c=phInputs(1500:2000);
PhTargets_c=phTargets(1500:2000);

[p1,Pi1,Ai1,t1] = preparets(narx_net_closed,phInputs_c,{
    },PhTargets_c);
%  Network simulation 
yp1 = narx_net_closed(p1,Pi1,Ai1);
plot([cell2mat(yp1)' cell2mat(t1)'])
toc

Add completed , Click on “ function ”, Start emulating , The output simulation results are as follows ：

performance =

    0.0188


trainPerformance =

    0.0180


valPerformance =

    0.0164


testPerformance =

    0.0252

 Time has passed  7.422134  second .

（ In turn, click Performance,Training State,Time-Series Response,Error Autocorrelation,Input-Error Cross-correlation Corresponding figures can be obtained respectively ）

3. Summary

Dynamic neural network has become a new research topic of deep learning . Compared to the static model ( Fixed calculation chart 、 Fixed parameter ), Dynamic network can adjust its own structure or parameters according to different inputs , Formed accuracy 、 Computational efficiency 、 Significant advantages in self adaptation, etc . Interested in the content of this chapter or want to fully learn and understand , It is suggested to study the contents of chapter 40 in the book （ The learning link is attached at the end of the text ）. Some of these knowledge points will be supplemented on the basis of their own understanding in the later stage , Welcome to study and exchange together .

【 review 】 An overview of Dynamic Neural Networks