With the extensive integration of novel load demands and distributed PV generations

power quality issues are becoming more prevalent in low voltage distribution networks (LVDNs). Existing control strategies struggle to adapt to LVDNs

which features weak communication conditions

limited computational capacity

and complex power quality issues

thereby failing to meet the power quality needs in LVDNs. Therefore

this paper proposes a decentralized coordinated power quality control strategy for LVDNs based on multi-agent deep reinforcement learning (MADRL). First

considering the typical characteristics of LVDNs

the power quality coordinated control problem is transformed into a decentralized partially observable Markov decision process. Secondly

to address the low sample efficiency

the modified deep dense reinforcement learning (MD2RL) structure and the MATD3-MD2RL algorithm are designed. Then

a centralized training-decentralized execution framework is introduced to ensure decentralized observations while achieving multi-agent coordination. Finally

case studies are conducted on the IEEE 13-bus system and a practical LVDN. The results demonstrate that the proposed control strategy improves the power quality of LVDNs with additional advantages such as independence from communication

strong real-time performance

and no reliance on precise power flow models.