When a power plant's cooling towers lose their grip on summer heat, the consequences ripple far beyond a single facility. Researchers at the University of Tennessee have developed ReForM—the RetroSight and ForeSight Ensemble Model—a novel forecasting system that combines physics-based predictions with machine learning to foresee river temperature changes days in advance, helping power plants dodge million-dollar operational crises.
The challenge is deceptively simple but consequential. Rivers like the Tennessee provide the water that cools power generation facilities, but rising temperatures—whether from climate change or operational strain—reduce cooling efficiency and trigger costly shutdowns. Automated sensors along waterways offer real-time data, but they're unreliable: debris, lightning strikes, and communication failures create gaps that leave operators flying blind about what comes next. When a sensor fails, power plants have no choice but to scramble for emergency measures, sometimes purchasing replacement power at last-minute rates or reducing output, both expensive options that can cost millions during a single summer.
A collaborative team led by Industrial Systems Engineering Professor Anahita Khojandi and Environmental Engineering Professor Jon Hathaway at the University of Tennessee's Tickle College of Engineering took on the problem after the Tennessee Valley Authority approached them three years ago. What started as a pitch for real-world data for an applied data science class evolved into a research partnership born from TVA's urgent need. "There are various water temperature forecasting models out there," Khojandi explained. "However, they cannot necessarily forecast the temperature well over extended periods of time. The novelty of our work is in using 'future' data from physics-based models and combining them with 'historical' sensor data using machine learning models to build much more accurate forecasting models."
The system's brilliance lies in its economy of insight. Rather than rebuilding expensive physics-based models from scratch, ReForM leverages existing ones alongside historical sensor readings, teaching machine learning algorithms to see patterns far enough ahead to matter. During high-temperature periods, when cooling towers struggle most, this foresight becomes invaluable. Instead of reacting to crises at midnight, operators can arrange replacement power from external sources days earlier, when rates are lower. The difference between last-minute desperation and planned procurement is measured in millions.
The stakes extend beyond operating costs. Governments have tightened regulations on river temperatures to protect aquatic ecosystems from warming, putting power companies under dual pressure: maintain operations while safeguarding environmental thresholds. ReForM helps operators balance both by enabling precise planning rather than emergency response. As Jon Hathaway noted, the burden isn't just shutting down—it's the cascading damage of restarting complex systems. "It's something they really, really want to avoid."
The research proved so compelling that the applied data science class work transformed into a published journal article in the Journal of Hydrology, based on a case study of the Buffalo River in Middle Tennessee. Students, energized by solving real-world problems with genuine data, pushed themselves to publish findings that might otherwise have remained classroom work. The collaboration demonstrates how universities can fulfill their land-grant mission while training the next generation of engineers to see data not as abstract numbers but as solutions to problems their communities actually face.