Projects

A research team from the University of Michigan’s School for Environment and Sustainability is working with the U.S. Bureau of Reclamation and operational forecasting partners to improve reservoir inflow forecasts across the Colorado River Basin. The project integrates high-resolution snow observations, operational hydrologic models, and explainable artificial intelligence methods to better understand how snowpack conditions affect seasonal water supply and streamflow forecasts.

The project supports Reclamation’s need for more reliable water supply prediction under changing climate and increasingly variable snow conditions. The research team is developing and evaluating machine learning and deep learning models, including gradient boosting, LSTM, state-space models, and mass-balance-informed AI models, alongside the benchmark Colorado Basin River Forecast Center (CBRFC)'s operational hydrologic models, such as SNOW-17 and SAC-SMA. A major goal is to identify when AI models can complement existing operational models and how snow observations can improve forecast skill and uncertainty characterization.

To date, the project has developed a reproducible AI/DL models integrating more than 2 TB of snow, meteorological, and hydrologic data from USBR RISE, USGS, NSIDC, NRCS SNOTEL, and PRISM archives. The dataset includes snow water equivalent, snow depth, snow-covered area, precipitation, temperature, and streamflow information aggregated across more than 200 Upper Colorado River Basin watersheds.

The research team has also reconstructed the operational SNOW-17 and SAC-SMA modeling framework used by CBRFC and compared its performance with AI and deep learning models. Recent work has expanded this comparison to include automatically calibrated SAC-SMA models from large-sample hydrology datasets such as CAMELS, helping evaluate the differences between expert-guided operational calibration and automated calibration methods.

The next phase of the project, now based at the University of Michigan, will focus on expanding model comparisons across basins, developing hybrid AI approaches with better water-routing capability, applying explainable AI methods to quantify snow influence on streamflow forecasts, and building visualization tools to help operational partners evaluate snowpack impacts on forecast uncertainty.

Expected outcomes include open and reproducible modeling workflows, improved understanding of snow–runoff predictability, peer-reviewed publications, technical reports, and decision-support tools that can support USBR, CBRFC, and other water management agencies working in snow-dominated river basins.