Reactive Power Optimization for Distribution Networks Based on Parameter-Adaptive Reinforcement Learning

High photovoltaic (PV) penetration makes voltage violations in distribution networks prominent. Rule-based IEEE 1547 droop control is stable but suboptimal, whereas reinforcement learning (RL) policies suffer from unsafe exploration and oscillatory actions. This paper proposes PAR-Volt-VAR, a hierarchical RL framework whose upper-level agent, instead of controlling inverters directly, outputs a seven-dimensional parameter vector every two hours to tune a lower-level Volt-VAR controller with filtering, hysteresis, and capacitor lockout. The policy is pretrained by behavior cloning (BC) on grid-search demonstrations and fine-tuned with Soft Actor-Critic under an annealed Kullback–Leibler (KL) anchor. On modified IEEE 33-bus and 69-bus systems with five training seeds, PAR-Volt-VAR achieves the best daily reward, network-loss energy, voltage-violation magnitude, and minimum voltage among five methods, including two direct-control multi-agent RL baselines. Relative to Volt-VAR droop control, it improves the daily reward by 16.1% and 16.8% and reduces the violation magnitude by 24.7% and 13.8% without curtailing PV, but its reactive power output varies more. It remains the best under unseen load levels and PV intensities. In the ablation, BC initialization lets training start above droop control, and the KL anchor limits subsequent reward drops. Voltage violations are reduced but not eliminated: the minimum voltage remains below 0.97 p.u.