PINN-Infused Hybrid ML Forecasting on Lake-Effect Precipitation

Abstract

Lake-effect snow poses severe risks to Great Lakes communities, yet accurate prediction remains elusive due to a fundamental challenge: critical satellite observations are unavailable during nighttime hours when these systems rapidly intensify. We present a paradigm shift in lake-effect snow forecasting by first solving the temporal data discontinuity problem, then leveraging complete observations for physics-informed prediction. Our two-stage framework employs PatchGAN to synthesize missing visible and near-infrared satellite imagery from continuous infrared data, achieving 59% improvement in forecast accuracy compared to models trained on incomplete observations. These synthesized sequences feed into a physics-informed neural network architecture that modifies MetNet-3 to enforce atmospheric conservation laws while processing high-density weather station data at adaptive resolutions. Most remarkably, our approach reveals that harsh lake-effect events become more predictable at extended horizons—improving from 27.1% accuracy at 24 hours to 77.6% at 72 hours—as large-scale precursor patterns emerge in the complete observational record. Evaluated on 11 years of Great Lakes data, our framework achieves 87.4% overall accuracy for 24-hour forecasts and maintains 81.3% at 72 hours, substantially outperforming both traditional NWP models (42.3%, 66.5%) and standard deep learning approaches (45.3%, 64.1%). By demonstrating that intelligent data synthesis can unlock the potential of physics-informed machine learning, this work establishes new foundations for predicting localized severe weather phenomena where observational gaps have historically limited forecast skill.