The Laurentian Great Lakes contain 20% of the world's surface freshwater, making them a vital resource for both the United States and Canada. Accurately predicting the amount of water flowing from rivers and streams into the lakes is critical for effective water management. However, forecasting streamflow across such a vast, international region with inconsistent data has been a persistent challenge.

To solve this problem, we developed a new predictive tool using a deep learning approach. We created the most comprehensive dataset for the region to date, combining 43 years of daily weather and streamflow information from 975 basins. We used this dataset to train a single Entity-Aware-LSTM model to predict streamflow across the entire system.

Our results show the deep learning model is significantly more accurate than both a vanilla LSTM approach and the conceptual based operational model. This work demonstrates that a single, data-driven tool can provide a more robust and unified way to manage our shared water resources without needing expensive, location-specific adjustments, paving the way for better management of large, complex water systems worldwide.