PNNL

Abstract

Accurate flood early warnings are critical to minimize damage and loss of life. Current large-scale operational forecasting systems, however, have limited accuracy, fail to adequately describe uncertainties, and are computationally expensive. Advancements in artificial intelligence (AI) methods have demonstrated potential, but their forecast accuracy also remains insufficient. Here we combine a novel AI-based supporting method, termed Errorcastnet (ECN) with the National Water Model (NWM) to showcase the significant improvement of medium-range ensemble flood predictions over the contiguous United States. ECN boosts NWM accuracy across lead times of 1-10 days, while providing uncertainty quantification. ECN is also computationally efficient, enabling nation-scale ensemble forecasts in minutes. ECN is successful across varying ecoregions and geology conditions, including human-impacted areas. Furthermore, we show that the use of ECN can gain superior economic value (over 380%) for decision-making, as compared to that of NWM, especially for extreme events above 20-year return period.