Enhancing frequency stability with decentralized adaptive control using multi-agent deep reinforcement learning of multi-VSGs

In recent years, the global demand for sustainable energy and reduced carbon emissions has increased significantly. Consequently, the integration of distributed generation (DG) units based on renewable energy sources (RES) into power systems is accelerating. However, this integration reduces inertia and damping within power systems, leading to significant stability and performance challenges. To address these issues, virtual synchronous generator (VSG) technology has been deployed. VSG uses power electronics-based control technology to emulate the operational characteristics of conventional synchronous generators, providing inertia and damping properties to improve frequency and voltage stability. This paper proposes a novel decentralized adaptive control strategy for frequency stability in distribution systems with integrated multi-VSGs, requiring no model knowledge. This approach leverages multi-agent deep reinforcement learning (MADRL) to determine the optimal control policies for adaptive VSG parameters. The key contributions of this work include: (1) the development of a MADRL framework tailored for VSG control, (2) the implementation of a centralized reward sharing mechanism to enhance coordination among DGs, and (3) extensive simulation results on the IEEE 33-bus system demonstrating significant improvements in grid stability and performance. • Decentralized MADRL-based adaptive control for VSGs in active grids. • Model-free real-time tuning of virtual inertia and damping for frequency stability. • Reward-sharing coordination among VSGs for system-wide control and resilience. • Stable, constrained policy learning via PPO under varying network conditions. • PSCAD validation on IEEE 33-bus under load variation, islanding, and fault events.