Physics-Guided Multi-Agent Adversarial Reinforcement Learning for Robust Active Voltage Control With Peer-to-Peer (P2P) Energy Trading

The utilization of peer-to-peer (P2P) energy trading in the active distribution network can facilitate profit sharing among numerous prosumers while effectively accommodating the integration of distributed renewable energy. However, the market-oriented P2P energy trading involved self-interested prosumers could unavoidably result in voltage violations of buses. To tackle this challenge, this paper proposes a physics-guided multi-agent deep reinforcement learning (MADRL) integrated with adversarial learning for both day-ahead trading and intra-day voltage regulation. In the day-ahead energy trading, an energy cost minimization framework is built and constrained with distributed generators and battery energy storage systems (BESSs) for the repeated rounds of multilateral negotiations among prosumers to reach an optimal trading solution without considering physical network limitations. Then, in the intra-day voltage regulation, the physics-guided multi-agent adversarial twin delayed deep deterministic (PG-MA2TD3) policy gradient algorithm is designed to overcome the voltage fluctuation problem and minimize the line loss via adjusting the active power from BESSs and reactive power from photovoltaic (PV) inverters. Moreover, the Jacobian matrix is exploited to measure the impact of neighbor bus active power variations on local voltage due to neighborhood trading in P2P transaction and a multi-agent adversarial learning (MAAL) approach is implemented to obtain an adaptive descend gradient corresponding to the action of the adjacent agents in Q function for increasing the robustness of trained policy. It is verified that the proposed method provides better robustness and the largest steady-state reward with comparison to various state-of-the-art methods on the IEEE 33-bus system with three-year data in Portuguese power system.