LSTM for time series forcasting

Question

I manipulate the time series using the different structures of the neural networks in order to make a prediction, and I wonder if there is a way to choose the parameters of the networks intelligently? from the characteristics of the signal, namely (trend, seasonality ...) can we choose these parameters that will make learning better?

Answer 1

Indeed you can introduce some "unvariant" features to your LSTM network using Conditional RNN that use these features to create the initial hidden state:
https://github.com/philipperemy/cond_rnn

I hope this is what you are looking for.