Hybrid modeling for daily streamflow forecasting: A study over the contiguous United States

This study as lead by Dr. Dingbao Wang's group at University of Central Florida. It examines how machine-learning–based surrogate models can improve streamflow forecasting compared to traditional hydrologic models. Using data from 600 watersheds across the contiguous United States, the authors apply two hybrid modeling strategies that combine physical hydrologic modeling with deep learning. One approach focuses on learning the difference between observed streamflow and hydrologic model predictions, while the other augments the training data of the machine-learning model using outputs from the hydrologic model.

Both hybrid approaches consistently outperform the standalone hydrologic model in forecasting streamflow up to 30 days ahead across different regions and seasons. The improvements are particularly strong in the eastern United States, the Pacific Northwest, and the Rocky Mountains. While the hydrologic model’s performance varies with forecast lead time and tends to improve at longer horizons, the hybrid models deliver stable and superior accuracy across all lead times.

The analysis also shows that forecast improvements depend on geographic and hydrologic characteristics such as latitude, seasonality, precipitation patterns, vegetation, and baseflow conditions. Overall, the findings demonstrate that integrating machine learning with physical hydrologic models leads to more accurate and robust streamflow forecasts and provides insights for improving hydrologic model representations at the watershed scale.