I'm trying to traing an EncoderDecoder network for a multivariate time series input and a univariate time series output. In particular my dataset is composed of inputs of 32 features x 600 seconds and should produce 1 output x 300 seconds.
The MSE error both for training and validation goes down pretty quicly, but the result I'm obtaining it's not what I expect (see pictures below).
Do you have any clue why the model is learning this strange pattern? What can I do to prevent it?
