Machine Learning Enhancement of Spatial Lake-Effect Precipitation Forecasts

Lake-effect precipitation is often extremely localized and intense, making it impactful to society and difficult to predict. Because it is a convective, boundary layer process, even the highest resolution operational models struggle to adequately resolve lake-effect systems. In this study, we train machine learning models to improve lake-effect precipitation forecasts from the operational High-Resolution Rapid Refresh (HRRR) model. We focus on lake-effect events during a several year period over a domain downstream of Lake Ontario, but include other domains over the Great Lakes to increase and diversify our training dataset. The training data consists of HRRR output grids, such as lake-surface temperature, wind, temperature, moisture, and accumulated precipitation, labeled with precipitation totals from precipitation analyses. Both the National Severe Storms Laboratory Multi-Radar Multi-Sensor (NSSL MRMS) analysis and the National Centers for Environmental Prediction (NCEP) Stage IV analysis are tested. Machine learning models ranging in complexity from principal component analysis (PCA) to convolutional neural networks (CNNs) and Generative Adversarial Networks (GANs) are implemented and evaluated. Lastly, object-based verification techniques are used to evaluate and compare raw HRRR forecasts to predictions from the trained machine learning models. Results may improve forecasts of lake-effect precipitation and will serve as a guideline for using machine learning to enhance spatial forecasts from operational models.