Learning Decentralized Frequency Controllers for Energy Storage Systems

This letter designs decentralized controllers for energy storage systems (ESSs) to provide active power control for frequency regulation. We propose a novel safety filter design to gracefully enforce the satisfaction of the limits on the state of charge during transients. Our technical analysis identifies conditions on the proposed design that guarantee the asymptotic stability of the closed-loop system with respect to the desired equilibria. We leverage these results to provide a controller parameterization in terms of a single-hidden-layer neural network that automatically satisfies the conditions. We then employ a reinforcement learning approach to train the controller to optimize transient performance in terms of the maximum frequency deviation and the control cost. Simulations in an IEEE 39-bus network validate the significant transient performance improvements of the proposed controller design.