Seasonal anomaly detection

[misterspook]

What would be the best practices for seasonal anomaly detection? I have the data every 10 days for 17 years. The data is seasonal, non-stationary, the signal (vegetation index) depends mainly on the weather data. The goal is to catch the anomalous signal as soon as possible, early in the season.

I have tried using seq2seq models, nbeats, deepAR, Nhits... But the results are not so good in the beginning of the season, they get better as the season progresses and as we provide more data to the model. Our approach can be considered as a long term forecast. max_encoder_length is now around 70 (2 years) and the prediction length is in range 15-20 (150-200 days)

Maybe this would be a better option: https://github.com/jdb78/pytorch-forecasting/issues/149

Another question is on the sliding windows: is the TimeSeries dataset internally using overlapping windows (and if so, what is the overlap)? Usually we provide overlapping sliding windows to the models for seq2seq models.

many thanks!