Accelerating AC Optimal Power Flow with Deep Learning

As power grids integrate renewable energy sources and grow in complexity, efficiently solving AC Optimal Power Flow (AC-OPF) is essential for grid stability, operational efficiency, and market participation. This talk presents two complementary approaches to accelerate AC-OPF solutions while ensuring accuracy and reliability. First, we introduce two novel deep learning frameworks: (i) an unsupervised learning approach with dynamic Lagrange multiplier adaptation, and (ii) a physics-informed gradient estimation method augmented by semi-supervised learning. These methods achieve up to 35x speedup compared to conventional solvers, with optimality gaps below 1%. Second, we propose a constraint screening framework that exploits the mathematical structure of convex OPF formulations to eliminate non-binding constraints, significantly reducing computational complexity. Time permitting, we will also briefly discuss other AI-driven research in energy, including load forecasting and power system event detection.

Speaker: Yu Zhang, UC Santa Cruz