Numerical Weather Forecasting using Convolutional-LSTM with Attention and Context Matcher Mechanisms

Contains the datasets, models, and results for the paper https://arxiv.org/pdf/2102.00696.pdf

Datasets

Both of the datasets 'High Resolution' and 'WeatherBench' are available in here.
After downloading the datasets extract them under data directory.

• highres --> data/data_dump
• weatherbench --> data/weatherbench

Models

Download the trained models in here. Put the models under results directory. After the extraction the models should be in this hierarchy.

• results/
  • results/highres
    • models ...
  • results/weatherbench
    • results/sequential_results/
      • models ...
    • results/iterative_results/
      • models ...
    • results/direct/
      • models ...

Install & Run

For the installation, install the packages to a python=3.8 environment.

$ pip install -r requirements.txt

Arranging Configs

• For Higher Resolution dataset the arguments are taken from configs/higher_res/higher_res_config.py.
• For WeatherBench dataset the arguments are taken from configs/weatherbench/ according to forecast mode:
  • Sequential: configs/weatherbench/seq_model_confs.py
  • Iterative: configs/weatherbench/iter_model_confs.py
  • Direct: configs/weatherbench/direct_model_confs.py

these configuration scripts are overwriting the default parameters that are defined in config.py and config_generator.py

Training and Testing

The run.py script contains a main function for training and testing as shown in below:

if __name__ == '__main__':
    run(dataset="weatherbench",
        model_name="weather_model",
        exp_type="sequential",
        perform_training=True)

Reproducing the Paper's Outputs

We are also reproducing the outputs of the paper including table and figures in run.ipynb.