A machine learning approach to streamflow prediction in the upper Colorado river basin

Robert Hull1, Laura Condon1, Luis De La Fuente1
1Department of Hydrology and Atmospheric Sciences, University of Arizona

Due to their infrequency, the impacts of low-probability hydrologic conditions, such as extreme snowpack and drought, on watershed behavior can be difficult to assess in purely data-driven approaches. Large-scale process-based models can be used to simulate high-resolution data that capture hydrologic dynamics subject to low-probability events; however, large computational demands and complex model construction often limit the usability of these process-based simulations for stakeholders and decision makers. For that reason, this study presents a Machine Learning (ML) approach to learn streamflow dynamics from a large-scale process-based simulation in the Upper Colorado River Basin. Our goal is to create faster and more-interpretable tools that can still capture low frequency events. Long Short-Term Memory (LSTM) statistical emulators are used to predict daily, monthly, bi-annual, and long-term streamflow across a range of high- and low-frequency hydrologic conditions where observations may be too sparse to otherwise make accurate predictions. Here we explore first the ability of LSTM models to capture simulated streamflow from the process-based model and second the ability to transfer across the domain spatially. Our results will help bridge gaps between computationally intensive process-based simulations and purely data driven approaches.

Go to All Abstracts | 2021 El Día Home Page