Deep Learning Ensemble Predictions of Forcings for Hydrological Modeling

Abstract:

The US National Weather Service (NWS) River Forecast Centers (RFCs) rely on the Meteorological Ensemble Forecast Processor (MEFP) to generate calibrated ensemble forecasts of precipitation and temperatures that are used as inputs to the Hydrologic Ensemble Forecast Service (HEFS). The HEFS is currently used across the country to provide probabilistic streamflow predictions to a variety of sectors including reservoir management (for both water supply and flooding), hydropower, navigation, and recreation. A known limitation of the MEFP is the general underprediction of heavy-to-extreme precipitation events. This can result in large negative biases in subsequent runoff forecasts, which creates significant challenges for the NWS flood warning program and federal, state, and private water management and planning partners across the country. Recent flexible deep learning (DL)-based frameworks have demonstrated significant improvements with respect to traditional methods in addressing the current challenges and improving predictive skills particularly for the largest events that are directly related to flood forecasting. We propose to develop, test, and implement novel and efficient DL-based schemes to improve the MEFP forecasts of precipitation and temperatures, which will result in significantly more reliable and skillful streamflow predictions for the NWS. These capabilities will also be transferred to the NWS Office of Weather Prediction (OWP) and the National Water Center (NWC). To that end, the DL tools will be integrated and made compatible with the developing Nextgen forcing engine, and shared with Cooperative Institute for Research to Operations in Hydrology (CIROH) collaborators and projects.