The widespread integration of distributed photovoltaic (PV) systems has exacerbated voltage violations and three-phase imbalance issues in distribution networks. While existing deep reinforcement learning (DRL)-based methods reduce model dependency and enable rapid regulation of flexible reactive power resources such as PV systems

they lack consideration for fair active power curtailment under reactive power-constrained scenarios and struggle to ensure the safety of trained policies when faced with imbalanced distribution network models with errors. To address these challenges

this paper proposes a fair and secure voltage optimization method for three-phase unbalanced distribution networks based on multi-agent deep reinforcement learning (MADRL). First

fairness in active power curtailment is defined and integrated into the multi-device reactive power optimization problem. Second

two optimization modes

namely "reactive priority" and "reactive-active coordination"

are established based on sequential control logic

and a partially observable Markov decision process (POMDP) is formulated. Then

an extended robust data-driven linear power flow (ERD-LPF) safety layer is proposed and embedded into the agent decision-making process to correct unsafe actions. Finally

a dual-mode fair-secure control strategy is trained and executed using a multi-agent soft actor-critic (MASAC) algorithm. Simulation results on a modified IEEE-123 node system demonstrate that the proposed method effectively coordinates regional devices and switches control modes to achieve fair active power curtailment during PV output peaks. Additionally

the continuous-discrete action prediction and minimal correction mechanisms in the safety layer exhibit strong robustness to model training errors.