Machine Learning Forecasts of Regional Extreme Precipitation Under Lake-Effect Weather Regimes

Abstract

The prediction of regional extreme precipitation under lake-effect weather regimes represents a formidable challenge in modern hydrometeorology due to the complex, highly localized, and nonlinear atmospheric dynamics involved. Traditional numerical weather prediction models often struggle to accurately resolve the microphysical processes and localized thermodynamic instabilities that characterize lake-effect events, frequently resulting in spatial displacement and magnitude underestimation of precipitation bands. This paper presents a comprehensive investigation into the application of advanced machine learning techniques to forecast lake-effect extreme precipitation. By integrating high-resolution meteorological variables, including lake surface temperatures, atmospheric sounding parameters, and synoptic-scale wind profiles, we develop and evaluate multiple predictive architectures, ranging from ensemble tree-based methods to deep sequential neural networks. The study systematically addresses the entire analytical pipeline, encompassing data acquisition, rigorous preprocessing, spatiotemporal feature engineering, and model validation against observational radar and gauge datasets. Our findings indicate that machine learning models, particularly those capable of capturing temporal dependencies and nonlinear interactions, significantly outperform conventional baseline approaches in predicting the onset, intensity, and spatial distribution of lake-effect precipitation. Furthermore, feature importance analyses elucidate the critical physical drivers of these events, reinforcing the physical interpretability of the data-driven models. Ultimately, this research demonstrates the operational viability of machine learning frameworks in enhancing early warning systems and mitigating the socioeconomic impacts of extreme winter weather events in lacustrine regions.