Leveraging historical climate data and machine learningfor extreme precipitation forecasting

This study investigates the prediction of extreme precipitation events using CMIP6 historical climate model data from 1850 to 2014, employing both unsupervised and supervised machine learning approaches to predict events one day in advance. We compared various algorithms, including Isolation Forest, K-means, DBSCAN, Logistic Regression, Naive Bayes, Decision Trees, Random Forests, XGBoost, LightGBM, and CatBoost. Our findings reveal that supervised models, particularly XGBoost with hyperparameter optimization, outperformed other approaches, achieving the highest F0.5 score of 0.4433 and ROC AUC of 0.9109. Boosting algorithms generally demonstrated superior performance across various metrics. We also analyzed the effect of hyperparameter tuning on the models, observing significant improvements in some cases. Feature importance analysis identified local and neighboring precipitation measurements as the most influential predictors. Our machine learning approach allows for quantification of various features' effects and geographical influences, providing valuable insights into the mechanisms behind extreme precipitation events. This research demonstrates the potential of machine learning techniques in advancing our ability to predict extreme precipitation events, contributing to improved forecasting capabilities and enhanced understanding of underlying atmospheric dynamics, ultimately aiding in the development of more effective early warning systems and adaptation strategies to mitigate the impacts of extreme precipitation in a changing climate.