AI model delivers river temperature insights, strengthening US energy security

Hydrology experts at the U.S. Department of Energy's (DOE's) Oak Ridge National Laboratory (ORNL) used artificial intelligence and a physics-based understanding of streamflow to create a model that provides highly accurate predictions of river temperatures, even in waterways that lack sensors.

The method is important to hydropower utilities and dam operators for avoiding non-compliance risks, mitigating damage to aquatic ecosystems, and understanding impacts to downstream water users. The predictions have broad potential to support nuclear and other power plant operations, strengthening the nation's energy and economic security.

More than 70% of the nation's electricity is generated by thermoelectric power plants that use water for cooling, such as nuclear, natural gas, and coal-fired facilities. Information about the availability and temperature of nearby water resources is crucial for reliable and efficient power generation, in addition to agriculture, data center siting, managing fish populations, and overall ecosystem health. Yet, most U.S. waterways do not contain gauges or sensors that monitor temperature.

To construct a model to accurately predict river temperatures, ORNL scientists used an AI/machine learning approach called a Long Short-Term Memory network that's well suited to analyzing patterns over time. The model learned how weather and landscape conditions influence river temperatures over days, seasons and years.

"The model can improve our understanding of both existing nuclear power plant operation and siting suitability for the nation's nuclear expansion," said Sean Turner, senior engineer in the Water Resources Science and Engineering Group at ORNL.

The model achieved an average absolute error of only 1.1 degrees Celsius between predicted and actual values. The error rate was comparable to conventional, data-intensive models that take more time and resources to build and maintain, as detailed in the Journal of Hydrology. The framework:

• Consistently produced seasonal warming and cooling patterns across diverse waterways.
• Maintained accuracy during very hot weather events, times that are critical for grid reliability and regulatory compliance for water withdrawal and release.
• Made better predictions as scientists focused on nearby, relevant upstream areas that resulted in cleaner signals for downstream temperature predictions, especially in large rivers.
• Was trained using inputs that are available for all 2.7 million river reaches across the continental United States, meaning the model can generate daily in-stream temperature estimates anywhere-even in completely ungauged watersheds.

"These deep-learning foundation models, trained on vast amounts of data to recognize and predict long-term patterns, are producing better and more transferable results than the models that people have been building and tinkering with for the last 50 years," Turner said.

The team used publicly available data sources including nine years of daily observations from some 300 selected U.S. Geological Survey river gauges; ORNL-developed waterway data reflecting precipitation, air temperature, solar radiation, humidity, snowpack and other phenomena; ORNL-simulated daily streamflow statistics; and federal data on watershed characteristics.

More information on the model, River Temperature Time Series for Hydrothermal Modeling and Analysis (RiTHyMs), is available on the DOE HydroSource platform maintained by ORNL.

"We wanted a system that could be applied anywhere in the nation, and that means we needed to train it with data that's available for every waterway," Turner said. "That's where ORNL and the datasets we've generated for HydroSource came in."

Researchers are now applying the model to the managed river systems and utility operations of the Tennessee Valley Authority. They are also refining the model to enhance predictions in mountainous regions, targeting western watersheds influenced by glacial runoff, where other utilities have shown interest in water temperature projections.

RiTHyMs leveraged ORNL's high-performance computing resources to rapidly train the continental-scale model on large datasets across hundreds of river basins. The resources are part of the Oak Ridge Leadership Computing Facility, a DOE Office of Science user facility at ORNL.

Other ORNL researchers on the project included model development lead Md Abu Bakar Siddik, as well as Shih-Chieh Kao, Ahad Tanim, and Jesus Gomez-Velez, formerly of ORNL and now at the University of Iowa.

The project was funded by the DOE Office of Critical Minerals and Energy Innovation's Hydropower and Hydrokinetic Office.

UT-Battelle manages ORNL for DOE's Office of Science, the single largest supporter of basic research in the physical sciences in the United States. The Office of Science is working to address some of the most pressing challenges of our time. For more information, visit energy.gov/science. - Stephanie Seay