Abstract: With the growing integration of distributed energy resources, the phenomenon of three-phase imbalance in the distribution network (DN) becomes increasingly prominent, posing significant threats to the secure, stable and economic operation of DNs. To resolve this issue, this paper proposes an online three-phase imbalance mitigation method using distributed photovoltaic (PV) for DNs based on multi-agent deep reinforcement learning. First, the causes of three-phase imbalance in DNs are analyzed, and the collaborative goals for three-phase imbalance mitigation are proposed. Then, by dividing the DN into multiple regions according to geographical location and assigning a PV action strategy learning agent for each region, a multi-agent coordinated framework for three-phase imbalance mitigation is established. Subsequently, based on the multi-actor-attention-critic (MAAC) method, a centralized training algorithm for agent action strategy is proposed to achieve coordinated optimization of action strategies for a large number of geographically dispersed PV systems. Finally, the well-trained action network is deployed in each region, and the PV system action instructions are generated online based on the real-time regional observation information, realizing the distributed efficient coordinated mitigation of three-phase imbalance in DN. The proposed approach is tested in the modified IEEE 123-bus DN. Through comparison with four other benchmark approaches, the effectiveness and superiority of the proposed method in three-phase imbalances mitigation are verified.