LSTM open loop testing problem

I have designed a LSTM network for electricity price forecasting. It has 14 input variables and one output (day ahead electricity price). Once I have trained the network, I want to test it with an open loop. My idea is to update the network with real output values "YTest" and save the predicted results in "Y". Obviously, prediction is done with "XTest" values, which are the inputs.

The example given by MathWorks (https://es.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html) does a prediciton with 1 input and 1 output. As it is a time series, the input is the delayed (t-1) value of the output array. Therefore, when it comes to updating the network "predictandUpdateState" is used. In my case, I want to upddate it with the real values which are saved in the array "YTest".

This is the code that I have tried to use, but the lines which contain [net]=UpdateState are not correct. I attach the used data.

Thank you in advance.

CODE:

XTrain=INPUTSP5LSTM;

YTrain=OUTPUTSP1TRAIN;

XTest=INPUTSP5LSTMTEST;

YTest=OUTPUTSP1TEST;

layers = [

sequenceInputLayer(14,"Name","sequence")

lstmLayer(128,"Name","lstm_1")

lstmLayer(128,"Name","lstm_2")

fullyConnectedLayer(1,"Name","fc")

regressionLayer("Name","regressionoutput")];

options = trainingOptions("adam", ...

'MaxEpochs',100, ...

'GradientThreshold',1,...

'InitialLearnRate',0.005,...

'LearnRateSchedule','piecewise',...

'LearnRateDropPeriod',125,...

'LearnRateDropFactor',0.2,...

'Verbose',0,...

'Plots','training-progress');

net = resetState(net);

offset = 50;

[net] = UpdateState(YTest(:,1:offset));

numTimeSteps = 687;

numPredictionTimeSteps = numTimeSteps - offset;

Y = zero(numPredictionTimeSteps);

for t = 1:numPredictionTimeSteps

Y(:,t) = predict(net,XTest(:,offset+t));

[net]=UpdateState(YTest(:,offset+t));

end

Answer 1

[Akshat]

Hi Gonzalo,

As per my understanding of the question, you are facing trouble in updating the model based on the outputs you get.

The function "UpdateState" you mentioned may not be suitable for achieving your specific objective. There is a function called “predictAndUpdateState” which serves the exact purpose you are trying to serve.

Refer to the documentation here for more information on the “predictAndUpdateState” function:

https://www.mathworks.com/help/deeplearning/ref/seriesnetwork.predictandupdatestate.html

I had put a line in the code (line 23) to check if the network and data have any dimensionality issues, but I have commented it now and the rest of the code is running.

I have reproduced the updated code on my end and you can find the code attached below.

XTrain = INPUTSP5LSTM;
YTrain = OUTPUTSP1TRAIN;
XTest = INPUTSP5LSTMTEST;
YTest = OUTPUTSP1TEST;

layers = [
    sequenceInputLayer(14,"Name","sequence")
    lstmLayer(128,"Name","lstm_1")
    lstmLayer(128,"Name","lstm_2")
    fullyConnectedLayer(1,"Name","fc")  
    regressionLayer("Name","regressionoutput")];

options = trainingOptions("adam", ...
    'MaxEpochs', 100, ...
    'GradientThreshold', 1, ...
    'InitialLearnRate', 0.005, ...
    'LearnRateSchedule', 'piecewise', ...
    'LearnRateDropPeriod', 125, ...
    'LearnRateDropFactor', 0.2, ...
    'Verbose', 0, ...
    'Plots', 'training-progress');

% net = trainNetwork(XTrain, YTrain, layers, options);


% You can uncomment the above line to try training the network, I did this 
% to ensure that the network and dataset are not having any dimensionality
% issues.

offset = 50;
numTimeSteps = 687;
numPredictionTimeSteps = numTimeSteps - offset;
Y = zeros(numPredictionTimeSteps, 1);

% Reset the network state
net = resetState(net);


for t = 1:numPredictionTimeSteps
    % Predict using the updated state
    [net,Y(:, t)] = predictAndUpdateState(net, XTest(:, offset+t));
end

Hope this helps!

Regards

Akshat Wadhwa