1
$\begingroup$

I'm working on soil moisture forecasting using multivariate time series (MTS). More precisely, each time sample comes with multiple measurements regarding: the soil (e.g. soil temperature, and soil moisture at different depths) and the weather data (e.g. temperature, rain, wind, etc.). The current weather data does not provide all the information to forecast the soil moisture that depends on future weather (e.g. rain events, temperature....). In a real situation I would have also the weather forecast, but this would cause a kind of misalignment in my MTS.

In other words I have:

  • historical data in the range [-T_past, 0] for soil variables v1(t), v2(t), ...vn(t) and weather variables w1(t),w2(t), ..., wm(t)
  • forecast data in the range [0, T_future] for weather variables w1(t),w2(t), ..., wm(t)

I would like to use all these available data (soil and weather historical values + weather forecast values) to predict soil moisture in range [0, T_future].

I'm currently working with standard LSTM, GRU or most recent crossformer but these approaches does not manage this situation. Is there any other approach able to manage such a case of misalignment MTS?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

You can use the forecast data lagged T_future terms back to align with the historical data. You can use it in Linear Regression with dummy time or Prophet, not only NNs.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.