Next-Hour Lake-Effect Quantitative Precipitation Forecasts over the Great Lakes with Deep Learning

Abstract Lake-effect precipitation, which significantly impacts regions downwind of the Great Lakes, poses unique forecasting challenges due to its dependence on localized processes and intricate interactions, making it an ideal candidate for deep learning. This study introduces a novel approach toward next-hour quantitative precipitation forecasts (QPFs) for lake-effect precipitation by leveraging an improved generative adversarial network (GAN). Key innovations include testing multiple input variable groupings, modernizing the generator and discriminator, and expanding predictions to four Great Lakes: Superior, Michigan, Erie, and Ontario. A UNetFormer generator and spectral normalization in the discriminator are added to a pix2pix GAN framework to predict next-hour QPF for lake-effect events. A novel dataset is developed for the model with High-Resolution Rapid Refresh (HRRR) and Multi-Radar Multi-Sensor (MRMS) system fields for input and MRMS precipitation estimates for the target. Five input variable groupings are tested to identify the optimal predictors for each of the four lakes, and saliency analysis is applied to identify the most important input variables. Two models are developed for each lake: one with a weighted loss function [Lake-Effect Snow Network–Aggressive (LESNet-A)] and one without [Lake-Effect Snow Network–Base (LESNet-B)]. The developed models outperform the HRRR’s predictions across all six metrics, with gains of over 100% over the HRRR in Pearson correlation coefficient and fractions skill score indicating superior forecasts in all precipitation intensities. LESNet-A performs better in heavier events, while LESNet-B performs better in lighter events. Both are comparable in light events. Their outstanding performance notably displays the significant potential of modernized GANs to advance lake-effect precipitation prediction. Significance Statement Lake-effect precipitation significantly impacts many communities across the world. Its highly localized and variable nature poses a significant challenge to forecasters in predicting it accurately and precisely. This study leverages recent advances in deep learning to improve precipitation forecasts during lake-effect events in the next hour. Deep learning enables models to learn patterns inherent to lake-effect precipitation autonomously. The models developed demonstrate superior performance to the state-of-the-art High-Resolution Rapid Refresh model in precipitation amounts and locations, with percent gains in all metrics and over 100% gains in Pearson correlation coefficient and fractions skill score. Their ability to more accurately and precisely predict precipitation reflects their capability to provide improved, invaluable forecasts to the public during lake-effect events.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.