Review Jesper Dramsch
This solution uses an "opportunistic model mixing" approach.
5 individual contributions are fed into a Convolutional Neural Network (CNN) that weighs the inputs to generate a prediction. These 5 contributions consist of:
- 3 EMOS models from the 3 centres
- Climatology
- Another CNN that uses 18 input features
Some details here: The precipitation was normalized using the cube-root. The three EMOS models use a multiplexing approach (rolling window) to stabilize the solution over different weekly models. These EMOS models take the raw t2m and tp data and attempt a bias correction. The 18 inputs to the CNN use t2m and tp as well as orography, sst, and other data sources, both flat and at elevations.
I was not able to reproduce the results simply due to the fact that the data pre-processing into individual files takes a significant amount of compute. The code is well-structured and documented (thank you for that). It appears that special care has been taken to remove 2020 from the training data.
Out of interest, I would like to know if any sort of ablation study has been done on the individual parts of the model? (This would be highly unlikely given the pressures of a competition, but it would be interesting.)
Clearly, a lot of effort has been put into this contribution and the documentation thereof. I want to thank the authors for their contribution.