Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks
Authors: Joel Oskarsson, Tomas Landelius, Marc Deisenroth, Fredrik Lindsten
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experiment with the model on both global and limited area forecasting. |
| Researcher Affiliation | Collaboration | Joel Oskarsson Linköping University joel.oskarsson@liu.se Tomas Landelius Swedish Meteorological and Hydrological Institute tomas.landelius@smhi.se Marc Peter Deisenroth University College London m.deisenroth@ucl.ac.uk Fredrik Lindsten Linköping University fredrik.lindsten@liu.se |
| Pseudocode | Yes | Algorithm 1 Single-step prediction f for graph-based MLWP |
| Open Source Code | Yes | Our code is available at https://github.com/mllam/neural-lam/tree/prob_model_global (global forecasting) and https://github.com/mllam/neural-lam/tree/prob_model_lam (LAM). |
| Open Datasets | Yes | The dataset used for training and evaluation is a 1.5 version of the global ERA5 reanalysis3 [17], provided through the Weather Bench 2 benchmark [40]. |
| Dataset Splits | Yes | We use the years 1959 2017 for training, 2018 2019 for validation and 2020 as a test set. |
| Hardware Specification | Yes | The models are implemented2 in Py Torch and trained on 8 A100 80 GB GPUs in a data-parallel configuration. |
| Software Dependencies | No | The paper mentions 'Py Torch' but does not specify a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | We train all models using the Adam W optimizer [57] and utilize BFloat16 mixed precision to save GPU memory. The training costs for the models makes extensive hyperparameter tuning unfeasible. We choose hyperparameters based on initial experimentation with smaller models. For Graph-EFM the important weightings λKL and λCRPS in L can be chosen based on monitoring the model behavior during training. The full training schedule for the deterministic models is given in table 7 and for the probabilistic models in table 8. |