An Event-Adaptive Deep Learning Framework for Extreme Weather Prediction

Abstract

Accurate prediction of extreme weather events remains a significant challenge due to nonlinear atmospheric interactions, temporal variability, and severe class imbalance in observational datasets. This study proposes an event-adaptive deep learning framework for prediction of extreme weather events such as heat wave, cold wave, heavy rainfall, and thunderstorm using multivariate surface observatory data. Sequential models including Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) are compared with a tree-based boosting model (XGBoost). Class-weighted training is employed to address rare event imbalance. Performance is evaluated using operational meteorological metrics such as Probability of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI). Results indicate that persistent temperature-driven extremes are better captured by memory-based architectures, while short duration convective events favor boosting methods. The findings highlight the necessity of event specific model selection in operational extreme weather forecasting systems.