Author: Jonathan Frame, FloodBASE
Title: Machine Learning for the Next-Generation Water Resources Modeling Framework
Abstract: The Next-Generation Water Resources Modeling Framework (Nextgen) is model agnostic, simulates fluxes of interest across arbitrary scales and control volumes, and uses evidence-based evaluation of different approaches. This Nextgen recipe is perfect for making the best possible large-scale hydrologic simulations, and machine learning will undoubtedly be a key ingredient. We have developed a deep learning module for rainfall-runoff module for Nextgen, and we present results from a large-scale simulation across New England. We have also developed a machine learning-based module selection system specifically for Nextgen, which helps determine the best module configurations for rainfall-runoff predictions any basin across the United States, and we present results comparing four different modules. We will also present a plan for incorporating the state-of-the-art deep learning framework, NeuralHydrology, directly into Nextgen. NeuralHydrology provides robust and rigorous model training and a plethora of deep learning architectures. Additional NeuralHydrology features include probabilistic predictions and data assimilation. Finally, we will present a general forward looking strategy for best machine learning integration into Nextgen.