Multivariate Regression (in time and features) Using LSTM

Trying to feed a LSTM with different streamflow time series and their delayed sequences for gap filling. Let x be the initial matrix with selected predictors, one per line, considering size(x,2) as the number of samples. To introduce time dependence, the predictors are alternated with their delayed versions (from dt= [1:ndt], ndt being the maximum delay considered) as below:

for ii=1:size(x,2)
    for j=1:ndt
        x1(j:end,ndt*(ii-1)+j)=x(1:end-j+1,ii);
    end         
end

with the respective LSTM:

numFeatures = size(xTrain,1);
    numResponses = size(yTrain,1);
    numHiddenUnits = 300;

    layers = [ ...
        sequenceInputLayer(numFeatures)
        lstmLayer(numHiddenUnits)
        fullyConnectedLayer(numResponses)
        regressionLayer];

The target is a line vector y. Is there a more effective arrange to introduce time dependencies in LSTM? I mean, I have tried to associate every y instance with a 3D matrix x2 containning the values of x (not of x1) from (t-ndt) to (t):

for ii=ndt:size(x,1)
    x2(:,:,ii)=x(ii-ndt+1:ii,:);

end

But I don't know how to addapt the respectve LSTM.

I know the "Sequence-to-Sequence Using Deep-Learning example

• https://au.mathworks.com/help/deeplearning/ug/sequence-to-sequence-regression-using-deep-learning.html

I does not include explicit time dependencies.

Thanks.