Authors: Savalan Neisary, Ryan Johnson, Steven Burian
Title: Accounting for the impacts of Reservoir Operations, Diversions, and Unaccounted for Water Losses on NWM flows with Machine Learning: Towards Season-to-Season Water Supply Forecasting with the NWM
Abstract: Accurate water prediction provides critical decision-making criteria to water managers considering the beneficial use nexus in water resources management – flood storage, drought mitigation, hydropower, recreation, etc. The National Water Model (NWM) suite of streamflow products provides a first ever national-scale management and warning tools but the is a) limited in the necessary forecasting horizon and b) does not account for anthropogenic water activities. Building on the demonstrated performance of machine learning in CAMELS catchments, we explore the use of ML for estimating streamflow in heavily managed or controlled catchments. We explore the utility of ML in accounting for the impacts of reservoirs, diversions, and unaccounted for water extractions on the naturalized streamflow estimates from the NWM v2.1 retrospective, using the Great Salt Lake basin as a demonstration domain with XGBoost and LSTM ML algorithms. Model training leverages USGS streamflow observations, NWM v2.1 retrospective estimates, upstream reservoir storage, AORC meteorological conditions, and NRCS SNOTEL snowpack observations, with preliminary results reducing error by up to 50%, a 20% reduction in bias, and a 33% improvement in Kling-Glutz efficiency. Future research efforts aim tol scale the modeling domain to the Upper Colorado River Basin and couple with water resources management systems models to test and demonstrate the utility of the ML-data-fusion techniques for water resources management decision-making.