Authors: Jonathan M. Frame, Ryoko Araki, Soelem Aafnan Bhuiyan, Tadd Bindas, Jeremy Rapp, Lauren Bolotin, Emily Deardorff, Qiyue Liu, Francisco Haces-Garcia, Mochi Liao, Nels Frazier and Fred L. Ogden
Title: Machine learning for a heterogeneous water modeling framework
Abstract: We will present recent efforts from the National Water Center Innovators Program Summer Institute on the integration of machine learning (ML) models for the Next Generation Water Resources Modeling Framework (NextGen). Specifically Long Short-Term Memory (LSTM) networks and differentiable parameter learning conceptual hydrological models (δ conceptual models). We demonstrate that an LSTM model trained on CAMELS catchments makes large-scale predictions with NextGen across all catchments over the New England region and matches the average flow duration curve observed by stream gauges for high-frequency events. We demonstrate improvements in peak flow predictions when using δ conceptual model, but results also suggest performance increases may come at a cost of the representativeness of hydrologic states within the conceptual model. As a heterogeneous modeling framework, NextGen encourages selecting the right model/s for the right basin/s. We propose a novel approach that leverages ML to predict the best-performing model and ensemble weights for a heterogeneous modeling approach. Our findings advocate for the future development of ML capabilities within NextGen, underscoring the critical role of ML, pure deep learning, and δ conceptual models, in advancing hydrological modeling practices.