Hybrid Forecasting: ML Predictions of Lake-Effect Regional Extreme Precipitations

Lake-effect snow events can produce extreme snowfall rates exceeding 5 inches per hour, severely impacting Great Lakes communities. Current forecasting methods face critical limitations: operational NWP models provide only 12-24 hour advance warning with limited accuracy for severe events, while their computational demands often exceed the time available for deployment. We present a hybrid deep learning framework that extends lake-effect precipitation forecasting from 24 to 72 hours while achieving 77.6 percent accuracy for severe events-a 26.9 percentage point improvement over traditional methods. Our approach combines three key innovations: (1) 72-hour observation windows that capture complete atmospheric evolution cycles, (2) multimodal fusion of visible, near-IR, and IR satellite imagery with GAN-based nighttime synthesis for continuous monitoring, and (3) hybrid ConvLSTM architecture with transformer-based feature integration. Evaluated on 11 years of Lake Michigan data (2006-2017), our model achieves 81.3 percent overall accuracy while requiring only 3.5 hours of computation-a $20-60 \times$ speedup over NWP methods. This efficiency enables operational deployment on standard GPU hardware, democratizing advanced forecasting for resource-limited weather services. The extended forecast horizon provides emergency managers critical lead time, with potential annual savings of ${\$}$2.3 million per county based on improved emergency response efficiency and reduced stormrelated damages.