Soft open points scheduling in unbalanced active distribution networks based on multi-agent graph reinforcement learning

Abstract

This paper proposes an innovative unbalanced ADN operation strategy utilizing multi-agent graph reinforcement learning (MAGRL), where SOPs are scheduled to mitigate the three-phase unbalance and minimize system loss. The SOP scheduling problem in unbalanced ADN is modeled as a multi-agent partially observable Markov decision process (POMDP). Then, a direct approach based Backward/Forward Sweep (BFS) power flow model is proposed in our framework to provide precise power flow results within a few iterations to the training environment. The graph convolution networks (GCNs) are embedded in the policy network to further improve the agent capability of learning and capturing spatial correlations and topological linkages among nodes in complex unbalanced ADN, hence promoting the effectiveness of action strategy for the agents. This model has been tested on modified three-phase unbalanced IEEE 123-node system and IEEE 8500-node system. The results illustrate the notable regulation capability of the proposed method for unbalanced ADN.