PC-SINDYc AI stabilizes microgrids, beating PI and RL controllers

Integrating renewable energy sources like solar and wind makes microgrid frequency control harder because their dynamics are nonlinear, noisy, and often unknown. Traditional model-based controllers require accurate system equations, while pure data-driven methods such as reinforcement learning can be fragile and lack stability guarantees. Now, a team led by Mohan Du and Xiaozhe Wang from McGill University (based on the author affiliation context) introduces PC-SINDYc, a framework that combines sparse identification of nonlinear dynamics (SINDy) with physics-informed constraints and model predictive control (MPC) to learn and stabilize microgrids directly from phasor measurement unit (PMU) data.

PC-SINDYc's regression algorithm is built for real-world messiness: it uses physics-guided library construction to embed physical structure, total least squares to handle noise in both inputs and outputs, and random sample consensus to reject outliers caused by delays and constraint activations. The identified model is then embedded in a model predictive controller, which continuously computes optimal frequency setpoints. The authors prove that, under mild conditions, PC-SINDYc guarantees asymptotic stability—a key formal property that many black-box ML methods cannot offer.

In simulations on 4-bus and 13-bus microgrids, PC-SINDYc successfully controlled frequency across a range of disturbances not seen during offline identification. It outperformed traditional PI controllers, conventional SINDYc, and state-of-the-art reinforcement learning methods, demonstrating both robustness and generalization. This suggests that physics-aware sparse identification, paired with model predictive control, could be a practical, trustworthy alternative for autonomous grid management as renewable penetration increases.