Regional streamflow drought forecasting in the Colorado River Basin using Deep Neural Network models

Process-based, large-scale (e.g., conterminous United States [CONUS]) hydrologic models have struggled to achieve reliable streamflow drought performance in arid regions and for low-flow periods. Deep learning has recently seen broad implementation in streamflow prediction and forecasting research projects throughout the world with performance often equaling or exceeding that of process-based models. Deep learning models are a possible approach to increase the accuracy of streamflow drought predictions and to expand the spatial coverage of river locations with available streamflow drought forecasts.

As part of a multi-component Data-Driven Drought Prediction project, the U.S. Geological Survey is developing and testing deep learning models for streamflow drought forecasting. In this work, we present preliminary results of a deep learning model capable of predicting streamflow drought occurrence at ungaged locations for the Colorado River Basin (CRB). A long short-term memory (LSTM) neural network model was trained using 40 years (1980-2020) of daily streamflow data from 425 streamgages within and surrounding the CRB using static watershed attributes as well as meteorological and remotely sensed dynamic forcing inputs. Model tests were performed to evaluate model accuracy for now-casting streamflow drought conditions at ungaged locations and for forecasting drought conditions at lead times ranging from 0 to 14 days. Nearly all model configurations showed behavioral performance for predicting daily streamflow percentiles. Comparisons of LSTM model performance for predicting drought using fixed drought thresholds (calculated over all days and years) and variable drought thresholds (unique threshold calculated for each day of the year) identify differences in model skill between locations with implications for model design.