Machine learning for sub-seasonal pattern-based precipitation prediction

The repository includes codes for a benchmarking sub-seasonal pattern-based precipitation prediction model as well as feature selection using Random Forest model, including two folders:

1. DataProcess: create the target data and all possible features used for prediction

• ClusterSelect.ipynb/Cluster_CESM.ipynb: Derive monthly and seasonal precipitation regional patterns as the prediction target based on K-means clustering analysis, using precipitation data from observation EOBS and the ensemble earth system model CESM respectively;
• EOF_slp.ipynb/'EOF_sst.ipynb'/'eof_u200.ipynb'/'eof_z500.ipynb': pre-process sea level pressure, sea surface temperature, U200, Z500 from CESM ensemble and conduct EOF analysis to create features;
• NAO.ipynb: calculate seas surface temperature anomalies in a North Atlantic and the annual NAO index

2. FeatureEngineering:

• FeatureEngineering.ipynb: The script collects all types of pre-processed features and outputs a feature set with all 235 features and corresponding values;
• feature_selection.ipynb: The script uses the selected features to run feature selection models -- 1) hyperpatameter selection for the random forest model; 2) feature selection for Recursive feature elimination-Random Forest (REF-RF) and Monte Carlo-Random Forest (MC-RF); 3) model performance evaluation for RFE-RF and MC-RF model; And visualization of the evaluation scores.

Dataset preparation: Both observation dataset and earth system model data were used for model feature engineering and prediction

• Precipitation: collected from both EOBS as observation and CESM earth system model simulation;
• Other climate variables: collected from CESM earth system model simulation.
• CESM dataset refers to https://www.cesm.ucar.edu/;
• EBOS dataset refers to https://www.ecad.eu/download/ensembles/download.php