24AI Deep Learning based Postprocessing for Freezing Level Forecasting over FIRO Watersheds

Forecast-Informed Reservoir Operations (FIRO) optimize reservoir management by integrating advanced weather and water forecasts. Accurate freezing level (FZL) forecasts are essential for predicting precipitation types. These are crucial for making informed water retention or release decisions, ultimately enhancing flood control and water supply reliability, particularly in the Western U.S. In this study, we propose both deterministic and probabilistic Deep Neural Network (DNN) postprocessing frameworks, utilizing the U-Net architecture, to improve dynamical model forecasts, such as those from the Center for Western Weather and Water Extremes (CW3E)’s West Weather Research and Forecasting (West-WRF) model. The deterministic postprocessing uses the hyperbolic cosine of the error, while the probabilistic approach employs a Gaussian Mixture Model (GMM) as a DNN loss function.

Using a decade’s (2010-2019) worth of West-WRF model reforecasts and observations from the California Nevada River Forecast Center (CNRFC) at the 6-hourly temporal resolution, we developed end-to-end and robust FZL-DNN models. Our findings demonstrate significant improvements in forecast skill with root-mean-square error (RMSE) reductions ranging from 5% to 25% and correlation increases in the order of 8% over the Yuba-Feather watershed, a representative study area in northern California, for both wet and dry water years. Additionally, the FZL-DNN models effectively capture the spatiotemporal variability of FZL and are capable of generating bias-corrected forecasts with improved accuracy over complex terrains as well as different elevational bands. In conclusion, this novel deep learning-based postprocessing approach effectively addresses the limitations of raw dynamical model forecasts and presents a promising pathway for integrating machine learning into hydrometeorological forecasting and decision support as part of the FIRO framework.