problem in multi step dense time series prediction using LSTM network
3 views (last 30 days)
Show older comments
Hi everyone
For some time now, I have been trying to predict time series with multi features for input using the LSTM neural network. (multi input - one output)
The neural network is well trained and the error of training and test data is low. (About 1.5%) But when I want to predict the test data one by one, the error is very high and the end result is very wrong.
I put my code at the bottom. Please guide me in this regard.
clc
clear
%% Load Data
load('C:\Program Files\Polyspace\R2019b\bin\ontario_demand');
data=ontario_demand;
%% Create Train, Validation and test data
nData = numel(data);
TrainPercent = 0.995;
TestPercent = 0.005;
nTrain = round(TrainPercent*nData);
nTest = nData - (nTrain);
TrainData = data(1:nTrain,:);
TestData = data((nTrain + 1):end,:);
%% Standardize Data
mu = mean(TrainData);
sigma = std(TrainData);
StandardTrainData = (TrainData-mu)/sigma;
StandardTestData = (TestData-mu)/sigma;
%% Create X & Y for Train and Test Data
Lags = 1:5;
[XTrain, YTrain] = iLag(StandardTrainData, Lags);
[XTest, YTest] = iLag(StandardTestData, Lags);
%% Convert Matrix to Cell
XcTrain = iMakeCell(XTrain);
YcTrain = iMakeCell(YTrain);
XcTest = iMakeCell(XTest);
YcTest = iMakeCell(YTest);
%% Define LSTM Network Architecture
numFeatures = size(XTrain,1);
numResponses = size(YTrain,1);
numHiddenUnits = 100;
layers = [ ...
sequenceInputLayer(numFeatures)
lstmLayer(numHiddenUnits)
fullyConnectedLayer(numResponses)
regressionLayer];
options = trainingOptions('adam', ...
'MaxEpochs',100, ...
'GradientThreshold',1, ...
'InitialLearnRate',0.05, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropPeriod',50, ...
'LearnRateDropFactor',0.2, ...
'Verbose',1, ...
'VerboseFrequency',10, ...
'Plots','training-progress');
%% Train LSTM Network
net = trainNetwork(XcTrain,YcTrain,layers,options);
%% Calculate Training set error indicators
[net,YPredTrain] = predictAndUpdateState(net,XcTrain);
YPredTrain = cell2mat(YPredTrain);
YPredTrain = YPredTrain*sigma + mu;
YPredTrain = YPredTrain';
YTrain2 = YTrain*sigma + mu;
RMSE_Train = sqrt(mean((YPredTrain-YTrain2).^2));
MAE_Train = mean(abs(YPredTrain-YTrain2));
MAPE_Train = mean(abs((YPredTrain-YTrain2)./YTrain2))*100;
disp('Training set error indicators :')
disp(['RMSE = ', num2str(RMSE_Train)])
disp(['MAE = ', num2str(MAE_Train)])
disp(['MAPE = ', num2str(MAPE_Train)])
disp('===================================')
%% Calculate Test set error indicators
net = resetState(net);
[net,YPredTest] = predictAndUpdateState(net,XcTest);
YPredTest = cell2mat(YPredTest);
YPredTest = sigma*YPredTest + mu;
YPredTest = YPredTest';
YTest2 = YTest*sigma + mu;
RMSE_Test = sqrt(mean((YPredTest-YTest2).^2));
MAE_Test = mean(abs(YPredTest-YTest2));
MAPE_Test = mean(abs((YPredTest-YTest2)./YTest2))*100;
disp('Test set error indicators :')
disp(['RMSE = ', num2str(RMSE_Test)])
disp(['MAE = ', num2str(MAE_Test)])
disp(['MAPE = ', num2str(MAPE_Test)])
disp('===================================')
%% Predict Test Data
net = resetState(net);
[net,YPred(1)] = predictAndUpdateState(net,XTest(:,1));
[net,YPred(2)] = predictAndUpdateState(net,[YPred(1);XTest(2:end,2)]);
[net,YPred(3)] = predictAndUpdateState(net,[YPred(2);YPred(1);XTest(3:end,3)]);
[net,YPred(4)] = predictAndUpdateState(net,[YPred(3);YPred(2);YPred(1);XTest(4:end,4)]);
[net,YPred(5)] = predictAndUpdateState(net,[YPred(4);YPred(3);YPred(2);YPred(1);XTest(end,5)]);
for i=6:size(XTest,2)
[net,YPred(i)] = predictAndUpdateState(net,[YPred(i-1);YPred(i-2);YPred(i-3);YPred(i-4);YPred(i-5)]);
end
YPred = YPred*sigma + mu;
RMSE_Pred = sqrt(mean((YPred-YTest2).^2));
MAE_Pred = mean(abs(YPred-YTest2));
MAPE_Pred = mean(abs((YPred-YTest2)./YTest2))*100;
disp('Test set error indicators :')
disp(['RMSE = ', num2str(RMSE_Pred)])
disp(['MAE = ', num2str(MAE_Pred)])
disp(['MAPE = ', num2str(MAPE_Pred)])
disp('===================================')
%% Plot the results
figure()
plot(YTest2,'Color',[0.4 0.4 0.4])
hold on
plot(YPredTest,'Color',[0.6350 0.0780 0.1840],'LineStyle','--','LineWidth',1.5)
legend('YTest','YOutput','Location','northwest')
title(['MAE = ',num2str(MAE_Test),' , RMSE = ',num2str(RMSE_Test),' , MAPE = ',num2str(MAPE_Test),' %'])
figure()
plot(YTest2,'Color',[0.4 0.4 0.4])
hold on
plot(YPred,'Color',[0.6350 0.0780 0.1840],'LineStyle','--','LineWidth',1.5)
legend('YTest','YOutput','Location','northwest')
title(['MAE = ',num2str(MAE_Pred),' , RMSE = ',num2str(RMSE_Pred),' , MAPE = ',num2str(MAPE_Pred),' %'])
here is the test data results :
![](https://www.mathworks.com/matlabcentral/answers/uploaded_files/780628/image.jpeg)
and here is the prediction data results :
![](https://www.mathworks.com/matlabcentral/answers/uploaded_files/780633/image.jpeg)
also my data and functions are attached.
0 Comments
Answers (0)
See Also
Categories
Find more on Sequence and Numeric Feature Data Workflows in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!