Deep learning architecture which can forecast climatic variables like precipitation and temperature using univariate historical data of those variables. Model architecture is inspired from N-BEATS: Neural basis expansion analysis for interpretable time series forecasting by Oreshkin et. al. (2019).

Data for our experiments was obtained from Climate Change Knowledge Portal maintained by the World Bank Group.

To train the model, clone this repository locally and download the data from the above source. Make sure to name the data files rain_wb.csv and temp_wb.csv, or change the names of files to be loaded in data_utils.py. Then in the repo directory, run:

# For temperature 
python3 main.py --task 'temp' --config 'configs/temp.yaml' --root 'path/to/dir/containing/datafiles'

# For precipitation
python3 main.py --task 'rain' --config 'configs/rain.yaml' --root 'path/to/dir/containing/datafiles'

To load a trained model to perform any downstream tasks, use the --load CLI argument like so (model weight will be stored as best_model.ckpt):

python3 main.py --task 'temp' --config 'configs/temp.yaml' --root 'dir/with/data' --load 'dir/with/trained/model'