An electrical model-free three-phase OPF for PV-rich LV networks using smart meter and transformer data

• An entirely model-free OPF approach that uses a trained NN instead of power flows. • It does not use any LV network models at any stage, nor past OPF results. • It maximizes PV generation whilst catering for voltage and thermal limits. • It shows similar accuracy as the conventional benchmark model-based AC OPF. • An alternative for distribution companies for operation without LV network models. The growing uptake of residential solar photovoltaic (PV) systems is prompting distribution companies to actively calculate PV inverters’ setpoints to manage voltage and thermal issues in low-voltage (LV) distribution networks. Although the model-driven Optimal Power Flow (OPF) has shown great potential for this purpose, the unavailability of electrical models (e.g., topology) for three-phase LV networks hinders its adoption by distribution companies. To address this, we propose an electrical model-free (independent from LV network models) three-phase OPF approach to calculate the PV inverters’ setpoints. No electrical model is used at any stage, instead a Neural Network is trained solely with historical measurement data (that calculates voltages instead of using power flows) to capture the nonlinear relationships of the LV network. The performance of the proposed approach is assessed using synthetic data considering a realistic Australian residential LV network with 146 customers, part of an integrated MV-LV network with 3,400+ customers. Time-series results demonstrate that the proposed approach has similar accuracy as the state-of-art model-driven OPF (that fully captures the physics) in calculating PV setpoints and almost same overall curtailment. Thus, the proposed approach presents an alternative for distribution companies to manage PV without any LV network models, only historical measurement data.