Multi-variate multi-step forecasting

[jSwords91]

Hi.
I have a dataset with columns: "ds", "y", "temperature", "rainfall".
Temperature and rainfall are included in the dataset history, and they also have forecasted values from t through to t+7 days. y is of course NaN for these forecasted dates.

I am having trouble training a multi-variate model and predicting in to the future up to t+7 days, using the forecasted weather values. I can predict the historic data fine (using this notebook as reference:
https://github.com/ourownstory/neural_prophet/blob/main/tutorials/feature-use/lagged_covariates_energy_ercot.ipynb

Do you have any advice how we can do this?

[jSwords91]

Self-solved...

This notebook is helpful:
https://github.com/ourownstory/neural_prophet/blob/main/docs/source/future-regressors.md

However the crucial element missing is that in the make_future_dataframe call there is a regressors_df option which is by default None. If you change this to be your dataframe of future regressors (ie. regressors_df=future_regressors_df), it will work.

In addition, I would advise setting periods=len(future_regressors_df) that way you'll be predicting only as far as your future regressors allow.

Great library - loving it so far.

[neomatrix369]

The above url does not work when we click on it, the URL to click on is https://github.com/ourownstory/neural_prophet/blob/main/docs/source/future-regressors.md

[jSwords91]

It must have changed. Thanks for finding the new location!