Multi-agent Deep Reinforcement Learning Optimization Scheduling for Distribution Networks Incorporating Knowledge Distillation

Aiming at the challenges such as slow convergence speed caused by the continuous expansion of the scale of distributed power supply, this paper proposes a multi-agent deep reinforcement learning (MADRL) optimization scheduling method based on Knowledge Distillation (KD). This method accelerates the convergence process of the agent by applying KD technology, so that it can more effectively deal with the uncertainty caused by source-load fluctuations. Firstly, this paper expounds the basic principle of KD, and on this basis, an optimal scheduling framework based on KD is proposed. The framework aims at minimizing the daily operating cost and constructs a day-ahead optimal scheduling model considering the system operating constraints. Then, the PPO + KD method is used to train the model offline, and the differences in convergence efficiency and effect of different algorithms are compared. Finally, the online scheduling optimization is carried out by using the trained model, and the improved IEEE123 node example is used for simulation verification. The results show that the proposed method is closer to the global optimal solution in scheduling strategy, and has significant advantages in convergence and computational efficiency.