[Paper] Thunderstorm nowcasting with deep learning:a multi-hazard data fusion model

[jacobbieker]

https://arxiv.org/pdf/2211.01001.pdf

Detailed Description

This paper describes ML models for predicting thunderstorms, taking in very similar data to MetNet, and dealing with certain modalities dropping out while still working, and gives probabilistic forecasts. Its forecasting 60min ahead at 5 minute intervals.

They did find the satellite data and radar were the most important inputs to the model:

Context

This seems like its trying to do a similar thing to MetNet/MetNet-2, with similar inputs, so might be helpful with expanding MetNet or adding it as a new option.

Possible Implementation

Code: https://github.com/MeteoSwiss/c4dl-multi
Data: https://zenodo.org/record/6802292
Pretrained models: https://zenodo.org/record/7157986

[jacobbieker]

@JackKelly @peterdudfield Some more support for satellite being quite important for short term forecasting at least. Not solar, but they did use EUMETSAT data

[jacobbieker]

https://arxiv.org/pdf/2203.10114.pdf has more details, and is the lightning-only original work