Real-Time Distribution Network Opeartion Optimization Using Mixed Safe Multi-Agent Deep Reinforcement Learning Method

The fluctuation in distributed generation and load significantly impacts the safety of system operation. This paper introduces a real-time distribution network operation optimization method based on multi-agent deep reinforcement learning to quickly determine the operation scheme for diverse controllable devices, such as switches, energy storage systems, and reactive compensators. To adapt to the mixed action settings within physical constraints, a novel mixed safe multi-agent soft Actor-Critic algorithm is introduced to train the policy. This approach entails constructing discrete neural networks and continuous neural networks, with two safety layers implemented to rectify unsafe discrete and continuous actions. a numerical analysis is conducted on the IEEE 33-bus system to confirm the efficacy of the proposed methods.