I am working on a Sequence-to-Sequence + Attention model for some time-series data. Now I have a really long time series, basically 40 years of daily observations for multiple sensors. The data itself comes from environmental sensors, so there is very distinctive seasonal trend. Here is a picture below. Right now I have a window of 365 days in and then predicting 365 days out. I could changes this to take in 730 days and predict out 365, and I would still have a lot of windows.
My question was, how long can I generally extend out a Sequence-to-Sequence model? Given the fact there is such strong seasonality in this data, seems like the long term behavior is very predictable, within some range of variation. If anyone has any references on how long an LSTM or Recurrent net can be extended, I would appreciate it.
Indeed, I did try and model this data with some ARIMA and SARIMA type models, but because most of the year the sensors are zero and then in the winter is where all the action is, the classical time-series models have a hard time determining the coefficients for the lags.