Deep Reinforcement Learning-Based Voltage-Var Control for Active Distribution Networks Using GAT-SimAM

High penetration of distributed photovoltaic (PV) generation can create spatially heterogeneous voltage excursions in active distribution networks, while conventional voltage-var control (VVC) methods struggle with uncertain operating conditions. This study proposes a soft actor-critic (SAC) VVC framework with a two-layer graph attention network (GAT) and a simple parameter-free attention module (SimAM). GAT aggregates topology-dependent information, and SimAM recalibrates the 32 feature channels within each node after graph aggregation without adding trainable parameters. The controller coordinates PV-inverter reactive power, static var compensators, and battery storage. Evaluation uses a frozen 96-day test set comprising 46,080 three-minute operating steps on a modified IEEE 123-bus feeder. Under a ±5% voltage band, the complete method reduces the whole-step violation rate from 15.651% without control and 0.701% with Q-V droop control to 0.106%; under the ±7% band specified by GB/T 12325-2008, it records no violations. A matched comparison between SAC with a GAT feature extractor and penalty-only GAT-SimAM reduces the violation rate from 0.697% to 0.109%, while average daily loss increases from 2.04 to 2.11 MWh. The results show that topology-aware feature extraction and parameter-free channel recalibration can improve voltage-risk sensitivity while maintaining millisecond-scale inference.