Work

Research on Systematic Optimization and Improvement Methods for Voltage Quality in Distribution Networks with High-Penetration Distributed Generation

Year: 2026
Type: article
Abstract: This paper develops a multi-agent deep reinforcement learning (MADRL) controller for voltage quality in distribution networks with high-penetration distributed generation. Each photovoltaic inverter acts as an autonomous agent that controls its reactive power from local measurements, while a QMIX-style mixing network is used during centralized training to decompose the global reward. The reward combines voltage deviation, network loss, a security barrier and action smoothness. The method is evaluated on the IEEE 33-bus feeder with six PV units, 15-min resolution, 50–75% penetration, three load levels and four operating scenarios (normal, communication delay, sensor data loss and PV tripping). Compared with no control, the controller raises the voltage qualification rate from 91.0% to 100% and reduces the average voltage deviation by 12.7% and the daily network loss by 14.8%; compared with droop control it attains the same qualification rate with a 1.9% lower daily loss. Among VDN, IQL, MADDPG and the proposed method, the four learned controllers reach nearly the same operating point; QMIX-CTDE is competitive with the best baseline, which shows that the proposed reward design and the decentralized execution architecture, rather than a particular value-decomposition variant, determine the performance. Batch experiments over 36 operating conditions show that the decentralized policy degrades gracefully when measurements or actuators are lost. The results provide a reproducible benchmark and practical evidence for deploying decentralized voltage control in high-DG distribution networks.
Language: English
Open Access status: bronze
OpenAlex ID: W7225906439
FWCI:
Cites:
Cited by:
Related to: 10