Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts

This repository contains the code of the paper Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts.

Data

In this Study the EUPPBench dataset (Demaeyer et al., 2023) is used, which is publicly available on Zenodo or using climetlab-eumetnet-postprocessing-benchmark .

Files and Folders

🕸️ GNN Training and Evaluation

📄 sweep.py: Hyperparameter sweep for GNN models.
📄 train_ensemble.py: Train a GNN, given the parameters in trained_models/X_XXh/params.json.
📄 evaluate_ensemble.py: Load an ensemble of trained GNNs and evaluate their averaged prediction.

📁 Models
📄 drn.py: Reimplementatioin of DRN (Rasp & Lerch, 2018) in PyTorch
📄 loss.py: CRPS loss functions
📄 model_utils.py: Utility Functions for the Embedding or ensuring positivity of the predicted $\sigma$
📄 benchmark_models.py: Implementation of different GNN architectures used in graphensemble/multigraph.py

🌦️ DRN

📄 drn_sweep.py: Hyperparameter sweep for the DRN.
📄 drn_train.py: Train a DRN, given the parameters in trained_models/drn_XXh/params.json.
📄 drn_eval.py: Load an ensemble of trained DRNs and evaluate their averaged prediction.

🛠️ Others

📁 Trained Models: Folder where restults of trained models are stored. 📁 Utils: Utility functions for plotting, data loading and preparation.

📈 Further Results and Plotting

📄 evaluation_plots_etc.ipynb: Almost all plots in the paper are generated here.
📄 further_results.ipynb: Used to calculate PI length and coverage.
📄 permutation_imp.py: Calculate feature importance.
📄 stations.ipynb: Plot of map of stations.