AI could optimize the power grid.

MIT researcher Priya Donti says AI can improve grid operations by better predicting renewable output and speeding complex optimization, while noting that different AI models vary widely in energy use and must respect grid physics.

AI has been criticised for rising energy use, but some tools can also reduce consumption and support cleaner grids. MIT News spoke with Priya Donti, an EECS professor who applies machine learning to power‑grid optimization. She explains that operators must maintain an exact balance between supply and demand while managing uncertain demand and variable renewable supply. Donti describes several technical ways AI could help, and she also highlights limits tied to model design and energy cost. Key points: - The power grid requires continuous balance between power entering and leaving the system, and demand is uncertain because customers do not pre‑register usage. - Variable renewable sources like wind and solar add uncertainty that makes forecasting and real‑time decisions more important. - AI can use historical and real‑time data to make more precise renewable generation forecasts and enable cleaner use of those resources. - AI can produce faster, more accurate approximations for complex optimization tasks that determine generator schedules, battery charging/discharging, and load flexibility. - AI can also assist planning with large simulation models, support predictive maintenance by flagging anomalous behavior, and speed materials or battery research; however, model size and training choices affect energy consumption and suitability for grid use. Summary: AI tools could improve grid efficiency and resilience by enabling better renewable forecasting, faster optimization, and support for planning and maintenance. Donti warns that AI approaches differ in energy cost and that models need to be built to respect physical grid constraints to avoid harmful mistakes. Researchers are working on algorithms aligned with operational physics and on broader, more democratized AI development. Undetermined at this time.