Forecasting is basicaly sequence-to-sequence regression, let suppos that your entire sequence is data,
1. You divide data into train and test parts, you can specify the proportion as you wish:
numTimeStepsTrain = floor(0.9*numel(data));
dataTrain = data(1:numTimeStepsTrain+1);
dataTest = data(numTimeStepsTrain+1:end);
2. Preparing training data and response sequences by shifting data by one time step, such as for data(t) the response will be data(t+1)
XTrain = dataTrain(1:end-1);
YTrain = dataTrain(2:end);
3. Preparing the network and training hyperparameters, then train the network using training data and training responses
numFeatures = 1;
numResponses = 1;
numHiddenUnits = 200;
layers = [ ...
sequenceInputLayer(numFeatures)
lstmLayer(numHiddenUnits)
fullyConnectedLayer(numResponses)
regressionLayer];
options = trainingOptions('adam', ...
'MaxEpochs',250, ...
'GradientThreshold',1, ...
'InitialLearnRate',0.005, ...
'Verbose',0, ...
'Plots','training-progress');
net = trainNetwork(XTrain,YTrain,layers,options);
3. Now you can forecast 1, 2, 3 or 4 steps ahead using predictAndUpdateState function, since you use predicted values to update the network state and you don’t use actual values contained in dataTest for this, you can make predictions on any time step number
net = predictAndUpdateState(net,XTrain);
[net,YPred] = predictAndUpdateState(net,YTrain(end));
stepsAhead = 4;
for i = 2:stepsAhead+1
[net,YPred(:,i)] = predictAndUpdateState(net,YPred(:,i-1));
end
Best Answer