Title: Optimization Meets AI: Foundation Models for Power Grid Operations
Abstract: Power grid operations depend on repeatedly solving large-scale optimization problems, such as AC optimal power flow and unit commitment, under tight time limits and hard physical constraints. Learned surrogates promise large speedups, but making them respect those constraints remains an open challenge. In this talk, I will present LUMINA, our grid foundation model effort, with a focus on how classical optimization ideas can shape AI model design. Concepts from constrained optimization inform how we train the model, so that a single model can transfer across grid topologies and adapt to new problem settings. I will show how these ideas translate into measurable gains in model performances, and briefly describe ongoing work to integrate LUMINA into broader agentic workflows for real-world use.
Bio: Yijiang Li is an Assistant Computational Mathematician in the Mathematics and Computer Science Division at Argonne National Laboratory. His research lies at the intersection of optimization and AI for science, including foundation models for power grid operations, federated learning across HPC facilities, and efficient training and compression of large models. He is a core developer of LUMINA, a grid foundation model for power system optimization. His work has appeared in ICLR, ICML, NeurIPS, NAACL, and KDD. He received his PhD in Operations Research from Georgia Institute of Technology in 2023.