The rapid growth of photovoltaic (PV) systems in distribution networks has increased the need for advanced voltage regulation strategies. Smart inverters (SIs), equipped with voltage control functions such as fixed power factor, Volt-Var, and Volt-Watt control, as well as communication capabilities with external systems, are expected to manage local voltage profiles by adjusting active and reactive power. However, a major challenge remains in determining, at each prosumer endpoint, which control logic—such as power factor or Volt-Var control—should be applied and how its parameters should be set according to the local grid conditions.

To tackle this issue, this study proposes a data-driven optimization framework for individually customized inverter control settings. Specifically, we focus on a personalized federated learning scheme that leverages locally measured voltage and PV generation time-series data, summarized into compact statistical descriptors (i.e., approximate representations of the joint distribution of voltage and PV output, modeled as a mixture of basis distributions capturing their cooccurrence patterns). This enables evaluation of various control settings without relying on detailed distribution system models. The objective function considers both the total curtailment and fairness in PV output reduction, using a generalized squared Hellinger distance to quantify imbalance. By coordinating parameter updates among prosumers with similar operational contexts, the proposed method achieves scalable, model-free optimization. Simulation results on a high-PV penetration distribution system confirm that our approach enables more effective voltage regulation and fairer utilization of renewable generation capacity across the network.