Work

Topology-Aware Federated Multi-Agent Reinforcement Learning for Dual-Level Voltage Regulation in PV-Rich Distribution Grids

Year: 2026
Type: article
Abstract: Photovoltaic (PV) generation with high penetration in the distribution networks causes voltage rise, voltage deviation and three-phase imbalance between connected Medium-Voltage (MV) and Low-Voltage (LV) feeders. In this paper, a topology-conscious federated multi-agent reinforcement learning is introduced to subordinated dual-level voltage regulation in PV-rich distribution grids. The given approach represents the network as a graph and implements the state encoding with the help of a Graph Neural Network (GNN) to ensure that every agent is able to represent local electrical state and local topology. Both LV control nodes are considered independent reinforcement learning agents, and the federated aggregation provides the means of updating shared parameters of the model without sharing the raw local measurements. The structure is aimed at minimizing MV voltage violations and LV phase unbalance collaboratively when there are different conditions of PV penetration and communication. The technique is tested on a stochastic load and PV test system on the IEEE 33-bus distribution test system. The findings show that voltage regulation and phase-balancing performance are enhanced when using the centralized RL strategy instead of baseline centralized RL and traditional MARL strategies. The research identifies the promise of topology-conscious learning and federated coordination in the provision of scalable voltage control in privacy-sensitive smart grid systems.
Language: English
Open Access status: bronze
OpenAlex ID: W7225999042
FWCI:
Cites:
Cited by:
Related to: 10