ERI: Climate-Invariant Operational Strategies for Hydropower Reservoirs via Generative and Domain-Adaptive Reinforcement Learning

This NSF Engineering Research Initiation (ERI) project aims to make the nation's hydropower fleet, a leading source of U.S. renewable electricity, more reliable under varying uncertainties in climate conditions. This project will advance a new paradigm for reservoirs operations by replacing rules built from historical records with artificial intelligence (AI) tools that adapt to droughts, floods, and other unprecedented conditions. This will be achieved by combining two complementary advances including a generative model that produces realistic, physically consistent climate scenarios spanning a wide range of futures, and an adaptive learning method that uses those scenarios to extract reservoir-operating policies that remain effective when the climate shifts. The intellectual merit of the project includes new methodology that reframes climate change as a learning problem requiring policies that transfer across conditions, rather than retraining for each new regime, advancing the science of reliable AI for critical energy infrastructure. The broader impacts of the project strengthen national energy resilience, workforce development, and public engagement in STEM through multiple coordinated efforts: publicly released datasets, software, and a hydropower visualization platform that lower barriers to climate-adaptive energy research nationwide; a new module in a renewable-energy course; equal research opportunities for graduate and undergraduate engineering students; and a student-led club that hosts laboratory visits and expanding STEM education outreach activities for K-12 students. The technical work is organized around two integrated thrusts. Thrust 1 develops a multi-head conditional generative adversarial network (GAN) that synthesizes multivariate hydroclimatic time series, including precipitation, temperature, snow water equivalent, inflow, and electricity demand across drought, flood, and stochastic-rainfall regimes, while preserving the physical relationships between variables that conventional statistical downscaling tends to lose. Thrust 2 develops a physics-informed deep reinforcement-learning (RL) framework for cascaded multi-reservoir operation that remains effective as hydrological conditions evolve under climate changes. The framework learns a shared state representation that captures features common to various climate conditions and regimes, so that a single operating policy can be deployed across climates and across reservoirs without retraining the full model. Reservoir mass-balance dynamics, storage and discharge limits, and downstream demand reliability are embedded directly in the simulator and reward function, ensuring that all operating decisions respect physical constraints. The framework will be evaluated on publicly available U.S. federal reservoir operating records coupled with the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS-US) basin dataset, with head-to-head benchmarks against established climate-scenario generation and reinforcement-learning approaches. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2550658 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Mohsen Saffari | Institution: Purdue University, WEST LAFAYETTE, IN | Award Amount: $199,888 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2550658 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2550658.html