Link to paper (under review currently)
Deep learning-based models have recently outperformed state-of-the-art seasonal forecasting models, such as for predicting
El Ni~no-Southern Oscillation (ENSO).
However, current deep learning models are based on convolutional neural networks which are difficult to interpret and can fail to model large-scale atmospheric patterns. In comparison, graph neural networks (GNNs) are capable of modeling large-scale spatial dependencies and are more interpretable due to the explicit modeling of information flow through edge connections.
We propose the first application of graph neural networks to seasonal forecasting.
We design a novel graph connectivity learning module that enables our GNN model to learn large-scale spatial interactions jointly with the actual ENSO forecasting task.
Our model, \graphino, outperforms state-of-the-art deep learning-based
models for forecasts up to six months ahead.
Additionally, we show that our model is more interpretable as it learns sensible connectivity structures that correlate with the ENSO anomaly pattern.
- Download the datasets from this link
- Place the downloaded data into this subfolder (which already has the correct substructure with subdirs SODA, GODAS, CMIP5_CNN).
Please follow the instructions in this file.
All reported models (4 per #lead months) are provided in the out directory. To reload them & ensemble them as in the paper (and get the reported all season correlation skills), you may run the eval_gcn script for a given number of lead months/horizon.
Please run the run_graphino script for the desired number of lead months h in {1,2, .., 23} (the horizon argument).
To produce Figure 2 in our paper, i.e. a heatmap of eigenvector centrality scores of the nodes for various of our GCN models for different lead times, please see this jupyter notebook.
Please consider citing the following paper if you find it, or the code, helpful. Thank you!
@article{cachay2021world,
title={The World as a Graph: Improving El Ni\~no Forecasts with Graph Neural Networks},
author={Salva Rühling Cachay and Emma Erickson and Arthur Fender C. Bucker and Ernest Pokropek and Willa Potosnak and Suyash Bire and Salomey Osei and Björn Lütjens},
year={2021},
eprint={2104.05089},
archivePrefix={arXiv},
primaryClass={cs.LG}
}