Q-learning based event triggered reactive power sharing in AC microgrids

Reactive power-sharing (RPS) and voltage containment within grid-code limits are critical yet challenging objectives in islanded low-inertia AC microgrids (MG); because of unknown feeder impedances, droop-induced deviations, and the high communication burden of conventional distributed secondary control. To address these challenges, this paper proposes a distributed Q-learning-based linear quadratic Gaussian (LQG) policy scheme, where the value function (Q-values) adaptively determines the optimal control policy over time. Unlike conventional LQG methods that rely on deterministic feedback laws derived from system models, the Q-learning-based LQG utilizes a learned policy that adapts to observed performance. The proposed scheme is further enhanced by an event-triggered mechanism (ETM) to simultaneously achieve operational objectives in the physical layer and reduce data exchange in the cyber layer of MGs. Simulation and hardware-in-the-loop experiments demonstrate: (i) voltage restoration, (ii) exact proportional RPS, (iii) considerable reduction in communication events compared with continuous-time benchmarks, and (iv) robust performance under communication delays while preserving stability and power-sharing accuracy. These results enable cost-effective, resilient deployment of islanded MGs with low-bandwidth communication infrastructure, facilitating high renewable-energy penetration in remote communities while ensuring compliance with grid codes.