Using Entity-Aware LSTM to Enhance Streamflow Predictions in Transboundary and Large Lake Basins

Accurate hydrological simulation of the Laurentian Great Lakes-the world’s largest surface freshwater system-is critical for effective water resource management, yet remains challenging due to the basin’s vast size, international boundaries, and data heterogeneity. While deep learning models such as long-short-term memory (LSTM) networks have shown promise in hydrological forecasting, a unified, data-driven model for daily streamflow prediction across the entire Great Lakes basin has not been realized. This study addresses this gap by developing and evaluating an Entity-Aware LSTM (EA-LSTM) model, trained on a newly compiled dataset spanning 975 sub-basins across the U.S. and Canada from 1980 to 2023. The dataset integrates daily meteorological forcings, static catchment attributes, and observed streamflow. The EA-LSTM, using a temporal training split, achieved a median Nash-Sutcliffe Efficiency (NSE) of 0.685, significantly outperforming both a standard LSTM (median NSE 0.567) and the operational NOAA National Water Model (median NSE 0.208). Performance analysis reveals that the model’s predictive skill varies by catchment characteristics, with reduced performance observed in very small and highly urbanized basins. Our findings demonstrate that a single, data-driven model can effectively simulate streamflow across a vast, international watershed without basin-specific calibration, offering a robust and unified tool for transboundary water resource management. This work highlights the potential of advanced deep learning models to enhance hydrological prediction and support transboundary water resources management.