The new US Department of Agriculture snowmelt runoff and water supply forecast model for the American West: leveraging multi-model ensembles, evolutionary computing, and automated, theory-guided, interpretable artificial intelligence

Abstract:

The western US depends on water supplies provided by spring-summer river runoff. Accurate water supply forecasts (WSF) are crucial for optimal management of agriculture, hydroelectric generation, transboundary rivers, reservoir safety, ecosystem health, and other issues. The USDA Natural Resources Conservation Service (NRCS) monitors snowpack and climate with its SNOTEL network and produces operational WSFs. Capitalizing on improvements potentially enabled by data science advances, we developed an AI-based metasystem as the next generation of operational NRCS WSF model. The approach integrates proven concepts underlying the existing system with a careful selection of modern data-driven predictive analytics achieving specific design goals. The architecture is modular and multi-layered. Main elements are a multi-model ensemble of six statistical and machine learning regression methods, with associated probability models; and optimal feature extraction and selection using unsupervised statistical pattern recognition with a genetic algorithm. Key attributes include using a priori hydrologic process knowledge to constrain AI solutions; employing automated machine learning (AutoML) to facilitate efficient use by operational hydrologists; using parallel processing to accelerate training; and addressing known statistical complexities around WSF in the US West, including ability to seamlessly handle nonlinear relationships and asymmetric and time-varying prediction uncertainty intervals. Application to 20 test cases spanning diverse geophysical environments demonstrated performance and usability advantages over the existing system. Moreover, live operational testing in the 2020 forecast season demonstrated logistical feasibility of associated real-time workflows, and amenability to forming a hydrologic ‘storyline’ around prediction results consistent with modern directions in explainable AI.