FORECASTING STREAMFLOW DROUGHT IN THE COLORADO RIVER BASIN USING MACHINE LEARNING MODELS

Streamflow drought remains a persistent challenge for water resources management across the Western United States, and streamflow drought duration and severity are expected to increase over the coming decades. However, large-scale process-based hydrologic models often struggle to accurately predict streamflow in arid regions and during low-flow periods, limiting their skill at forecasting drought. Considering this deficiency, data-driven artificial intelligence and machine learning modeling techniques are increasingly being used to forecast streamflow drought in river basins worldwide, and often perform favorably to process-based hydrologic models. The U.S. Geological Survey has tested several data-driven machine and deep learning modeling methods, including random forest, gradient boosting, and long short-term memory models, for forecasting streamflow drought in the Colorado River Basin. The models were built using 40 years (1980–2020) of streamflow data from 425 U.S. Geological Survey streamgages within and around the Colorado River Basin, along with static watershed attributes and dynamic remotely-sensed and gridded climatic and land surface forcing data. Preliminary results from a gradient boosting model, a machine learning algorithm that applies an ensemble of decision trees, will be presented for a retrospective analysis of streamflow drought in the Colorado River Basin during 1980–2020, including predictions in ungaged watersheds. The gradient boosting model predicts percentiles of streamflow, with drought characterized by how much the percentiles fall below a drought threshold computed from a Weibull distribution of gaged streamflow. In addition, application of the gradient boosting model to operational forecasting of onset, severity, intensity, and duration of streamflow drought from one day to six months in advance will be discussed.