TY  - EJOU
AU  - Yao, Shun 
AU  - Ji, Bin 
AU  - Li, Jiangcheng 
AU  - Cheng, Leiming 
AU  - Zhang, Yiming 

TI  - Optimal Configuration of Distributed Energy Storage in Distribution Networks Based on Graph Multi-Agent Reinforcement Learning
T2  - Energy Engineering

PY  - 
VL  - 
IS  - 
SN  - 1546-0118

AB  - To address the problems of insufficient topology awareness and the difficulty of high-dimensional discrete-continuous joint decision-making in the configuration of distributed energy storage in distribution networks with high penetration of renewable energy, this paper proposes a bi-level optimal configuration method based on graph-structure-enhanced multi-agent reinforcement learning. First, a bi-level framework of “upper-layer multi-agent cooperative planning and lower-layer operational verification” is constructed. In the upper layer, Laplacian positional encoding and topology-consistency self-supervised constraints are introduced to enhance the agents’ structural perception of coupling relationships among power network nodes. In addition, Gumbel-Top-K sampling, combined with action masking, is adopted to enable end-to-end joint decision-making for siting and sizing under mutual-exclusivity constraints. In the lower layer, a DistFlow model based on second-order cone programming (SOCP) relaxation is used to evaluate samples in the operational dataset, and the operating cost and security violation penalties are fed back to the planning layer. Results on the IEEE 33-bus, 69-bus, and 123-bus systems show that the proposed method can reduce the life-cycle cost, improve the operational feasibility on the test sample set, and enhance node voltage profiles and line loading levels.
KW  - Distributed energy storage; distribution network planning; graph multi-agent reinforcement learning; energy storage allocation; optimal power flow

DO  - 10.32604/ee.2026.085439