A physics-informed deep learning framework for real-time reactive power compensation in renewable-rich distribution networks

High penetrations of inverter-interfaced photovoltaics (PV) are rapidly transforming distribution feeders from passive networks into actively managed cyber–physical systems. While this transition enables decarbonization, it also amplifies operational challenges at the medium- and low-voltage levels: fast PV ramps, stochastic demand, and parameter drift can drive sustained voltage excursions, increased losses, and unfavorable power-factor behavior, particularly in high (Formula presented) radial feeders. Although smart inverters can provide reactive power support, practical deployment requires controllers that are (i) fast enough for real-time operation, (ii) robust to uncertainty and model mismatch, and (iii) credible against established baselines.This paper develops a physics-informed deep learning (PINN) framework for real-time reactive power compensation in renewable-rich distribution networks. The proposed controller learns the mapping from operating conditions (PV/load profiles and feeder states) to inverter reactive power setpoints while embedding distribution-network physics into the training objective through power-flow consistency and voltage-limit penalties. A complete benchmarking pipeline is implemented in MATLAB to generate realistic day-ahead scenarios (high PV penetration, cloud-induced ramps, stochastic loads, and parameter uncertainty with an aging factor) and to compare: (1) no compensation, (2) local Volt/VAR droop control, (3) loss-aware voltage-constrained OPF, (4) a purely data-driven MLP, and (5) the proposed PINN. Validation on two canonical radial feeders (IEEE 33-bus and IEEE 69-bus) is performed using voltage-profile statistics, energy losses, violation metrics, power factor, actuation patterns, robustness percentiles/boxplots, and computation time.