Multi‐Timescale Voltage Optimization and Control for Rural Distribution Networks with High‐ <scp>PV</scp> ‐Penetration

Abstract The high penetration of distributed photovoltaic is easy to cause voltage over‐limit, and the centralized control method of traditional linear model is easy to fail in this scenario. To this end, this paper proposes a model‐free voltage control method based on multi‐agent reinforcement learning. Firstly, the influence mechanism of photovoltaic access on the voltage distribution of low‐voltage distribution network is analyzed by theoretical modeling, and the coupling relationship between active power fluctuation and voltage deviation at the end of feeder is revealed. On this basis, the objective function is constructed to drive the agent to learn the global optimal strategy autonomously. Then, aiming at the problem of collaborative optimization of heterogeneous device control instructions, a multi‐agent deep reinforcement learning collaborative framework is designed based on Markov decision process, and a multi‐agent deep deterministic strategy gradient algorithm is adopted. Finally, through the centralized training‐decentralized execution mechanism, the photovoltaic reactive power output and energy storage charge and discharge continuous instructions are jointly optimized to break through the bottleneck of discrete‐continuous hybrid action coordination. The simulation results show that the voltage qualification rate of the proposed method can be increased to 98.7% compared with the traditional method. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.