Deep reinforcement learning-based model-free voltage regulation in an active distribution network using demand-side flexibility aggregation

The integration of distributed energy resources is challenging the operational stability of low-voltage distribution networks, leading to frequent voltage fluctuations that conventional devices, such as on-load tap changers (OLTCs) and capacitor banks, struggle to manage under dynamic and varied operating conditions. Voltage regulation in active distribution networks typically relies on accurate network models for optimization or on reactive-power support from inverter-based resources, which can be difficult to deploy in community-scale settings with limited observability. This paper presents a real-time, model-free voltage regulation framework that uses only residential active-power demand-side flexibility and a single-point voltage measurement at the point of common coupling, without requiring the network topology, line parameters, or reactive-power actuation for online execution. The proposed adaptive droop learning framework follows a centralized-training distributed-execution architecture: a Soft Actor-Critic (SAC) coordinator maps PCC voltage deviations to aggregate flexibility requests and learns to dynamically allocate them across load buses by learning implicit voltage–power sensitivity structure from data without explicit sensitivity calculations. Distributed SAC bus-agents disaggregate allocated requests into appliance-level setpoints using XGBoost-based flexibility bounds and a quadratic-programming feasibility layer to enforce device constraints and satisfy dynamic grid response requirements. Experimental validation on an RTDS with an IEEE 13-bus radial distribution system demonstrates 88.5% voltage compliance within a ± 10 V band, a 73% reduction in OLTC operations, and 96.9% flexibility-tracking accuracy, outperforming TD3, DDPG, and PPO baselines. The results indicate that low-observability, model-free voltage regulation is feasible for practical distribution feeders.