Model predictive control (MPC) has long been valued in power systems for its ability to handle constraints and optimize multi-step performance. However, most MPC designs assume accurate system models—an assumption that rarely holds in practice. Researchers have now addressed this gap by combining MPC with online learning, creating a framework that adapts to unknown system dynamics while maintaining provable stability guarantees.
The proposed approach uses a certainty-equivalence strategy where the controller simultaneously learns the system parameters and applies control. A regularized least-squares estimator refines its understanding of the system in real time, feeding improved estimates into the MPC algorithm. Crucially, a switching logic blends the primary MPC control with a saturated deadbeat controller, ensuring stability even during early learning phases when parameter estimates are poor.
A key innovation is the derivation of non-asymptotic stability bounds—mathematical guarantees that hold from the very first control step, not just in the limit as time goes to infinity. Previous results often required waiting for the learning phase to converge. Here, the authors prove high-probability bounds on closed-loop performance under sub-Gaussian stochastic disturbances, a realistic noise model for power systems experiencing measurement errors and load fluctuations.
The hard input constraints are critical for power applications where generators, converters, and HVDC links have maximum power outputs, voltage limits, and rate-of-change restrictions. The switching MPC explicitly handles these saturations while maintaining stability proof.
Numerical simulations validate the theoretical analysis across multiple test cases. For grid operators and control engineers, this work opens a path toward adaptive controllers that require less pre-tuning and offer mathematical confidence in stability from deployment onward. Future extensions may address nonlinear systems and distributed control architectures common in modern microgrids and multi-area power systems.
