Domain knowledge-enhanced graph reinforcement learning method for Volt/Var control in distribution networks

To address the limitations of existing deep learning-based Volt/Var Control (VVC) methods, such as inadequate modeling of inter-agent coupling and limited integration of power system domain knowledge, this paper proposes a graph attention-based multi-agent reinforcement learning approach that leverages domain-specific insights to enhance cooperative decision-making and improve voltage regulation performance. First, the VVC problem is modeled as decentralized partially observable markov decision process. Secondly, the influence of network topology, electrical characteristics, and source-load distribution on the coupling strength between agents is considered. Reactive voltage sensitivity among multiple inverters and the distribution network topology are embedded into a spatiotemporal graph with non-Euclidean structure as domain knowledge. Then, to capture dynamic correlations arising from fluctuations in source-load power, the graph attention mechanism layer is introduced into the critic network, optimizing the edge weights to suit the VVC task and generating spatially aggregated features for each agent. Finally, the critic network, leveraging the domain knowledge embedded in the model, provides a more accurate value function for policy evaluation, effectively guides the actor network in generating high-quality action policies, and facilitates domain-informed feature enhancement learning. The experimental results demonstrate significant improvements in voltage control performance compared to other advanced multi-agent reinforcement learning methods.