Physics-Shielded Multi-Agent Deep Reinforcement Learning for Safe Active Voltage Control With Photovoltaic/Battery Energy Storage Systems

While many multi-agent deep reinforcement learning (MADRL) algorithms have been implemented for active voltage control (AVC) in power distribution systems, the safety of electrical components involved in the operation of these algorithms are mostly ignored. In this work, a safe MADRL control scheme is proposed to regulate the reactive and active power control of photovoltaics (PVs) to alleviate power congestion and improve voltage quality by coordinating battery energy storage systems (BESSs) and static var compensators (SVCs). Uniquely, the learning algorithm designed in this paper can limit the action of the agent when approaching a dangerous state to ensure the safety of BESSs during the training process, which is realized by developing a multi-agent twin delayed deep deterministic (MATD3) policy gradient algorithm with a physics-based shielding mechanism. Specifically, actions that lead to dangerous states, the state-of-charge (SoC) of BESSs is fully loaded or drained, are replaced by the shielding mechanism with safe actions while maintaining system stability. Furthermore, each PVs node in the power distribution network is treated as an agent under the fact of reactive and active power sensitivities to voltage in the MATD3 algorithm, which is beneficial for improving scalability. Training, testing and comparative results on IEEE 33-bus and 141-bus with real-world data are provided to demonstrate the effectiveness and superiority of the proposed algorithm.