Stochastic Dynamic Power Dispatch With Human Knowledge Transfer Using Graph-GAN Assisted Inverse Reinforcement Learning
Junbin Chen, Tao Yu, Zhenning Pan, Mengyue Zhang, Guanhua Lu, Kedong Zhu
Abstract
This paper proposes a novel approach for dynamic economic dispatch (DED) of distribution networks, based on graph-generative adversarial network (Graph-GAN) assisted inverse reinforcement learning (IRL) with human knowledge transfer via demonstration. Firstly, the proposed method utilizes graph convolutional network (GCN) to capture the complex and nonlinear relationships between dispatch decision and system state. Secondly, a GAN-based approach is proposed to imitate the reward function from expert demonstration data, which avoids the need for manually designed reward functions. The trained policy network is then used for decision-making in real-time optimal dispatch of distribution networks. Experimental results demonstrate that the proposed approach outperforms traditional IRL methods and achieves supply-demand balance. Computation efficiency of the proposed method is thoroughly analyzed and shows that it is practically scalable to large-scale distribution networks. Overall, the proposed approach presents a promising alternative by incorporating human knowledge into reinforcement learning for DED of distribution networks.