S2S_weather: Improving Subseasonal-to-Seasonal Weather Forecasting by Leveraging Statistical Post-Processing
End of studies interniship at INRIA Paris in ARCHES team, supervised by Claire Monteleoni and David Landry (May-November 2024)
Subseasonal-to-Seasonal (S2S) weather forecasts are global or local weather predictions (classically 2meter temperature and total precipitation) beyond 2 weeks but less than 3 months ahead. These types of forecasts are essential for climate change adaptation, with an increasing number of extreme meteorological events (floods, cold-waves) threatening infrastructures and human health. However, classical Numerical Weather Predictions (NWP) atmospheric models show low skill at these ranges, as they are too far from the initial conditions while being sensitive to slowly evolving coupled components with the ocean.
We investigated statistical post-processing methods, including deep learning, to improve the IFS forecasts from the European Centre for Medium-Range Weather Forecasts, for 2m temperature and 10m wind speed, at 7, 14, 21, 28, 35 and 39 days ahead. We built reusable pipelines for WeatherBench2 (Rasp et al. 2024) data, from downloading to pre-processing. We also discussed and implemented normalization strategies and metrics fit to our problem and weather post-processing in general. You can read our findings in the report.
Figure: CRPSS of 2m temperature at lead 14 days for different methods: top row: raw model and EMOS (Gneiting et al. 2005), bottom row: variants of DRUNet (Pic et al. 2024)
The parameter folder contains the WeatherBench2 paths to the files we used in this study.
The obs folder contains static data useful for our study, from ERA5.
The processings folder contains codes to download (download_weatherbench2.py), format (format_data.py) and pre-process this data (dataset.py). Follow the instructions in the README file in the folder before running any experiments. For DRUnet, the [Month Lead Agg] is performed by the WeatherYearEnsembleDataset, while the [General Agg] by the WeatherYearEnsembleDataset.
You can train the different architectures of this project by running the functions in main.py. The corresponding models are defined in model.py:
- post-processing models: SpatialEMOS (EMOS), DRUNetPrior (DRUnet+prior both), DRUnetPriorVar (DRUnet+prior single)
- forecasting models: DRUnet (DRUnet both), DRUnetVar (DRUnet single)
Work in progress: DRUnetAll to unify all these variants under a single structure.
In main.py, define your training_results folder, and create subfolders for EMOS and DRunet. For EMOS, also create sub-folders for each lead (lead7, lead14, etc). It should look like this:
Update the paths accordingly in the main loop (base_dir variable), as well as the data_folder, train_folder, obs_folder in the mainXXX routines.
In order to train EMOS, comment the DRUnet part in the main loop and adjust the hyper-parameters (lr, batch_size, nb_epoch) in main_spatial_ens. To launch the training, use sbatch train.sbatch, modifying the lead using -lead within the bash script.
In order to train a DRUnet, comment the EMOS part in the main loop, uncomment your variant main_Unet_ensXXX part and adjust the hyper-parameters (lr, batch_size, nb_epoch) in it. To launch the training, use sbatch train_drunet.sbatch, modifying (DRUnets single) or removing (DRUNets both) the variable using -var within the bash script.
Samely, you can run the inference for these models using inference.py, as well as compute the baseline results, that is to say the Raw model (IFS) we are trying to post-process. The results will be save in a crps.nc file within the model folder in training_results.
In order to evaluate EMOS, comment the DRUnet part in the main loop and uncomment the specialSpatialEMOS_inference_detrend line. Launch using sbatch inference.sbatch.
In order to evaluate RAW, comment the DRUnet part in the main loop and the uncomment RawIFS_inference line. Launch using sbatch inference.sbatch.
In order to evaluate a DRUnet, comment the EMOS and RAW part and uncomment the DRunet part along with the DRUnet_inferenceXXX you are interested in. Launch using sbatch inference_drunet.sbatch, modifying the -var (variable), -id (number id in training result folder) and -e (epoch checkpoint to evaluate) in it.
You will then be able to visualize the results with the figures_results.ipynb notebook. The figures_data.ipynb notebook provides some code to visualize our data.