Power Flow Optimization and Voltage Stability Using DSTATCOM With Deep Learning Techniques in PV-Connected Grids Under Varying Load Conditions

As PV systems become more integrated into power grids, challenges arise in optimizing power flow and maintaining stability amid fluctuating load conditions. Variations in solar energy output and load lead to unbalanced reactive power and potential voltage instability. This study presents a Graph-Attention-Long Short-Term Reinforcement Network (GAL-RNet), which combines the innovations of the Reinforced Graph-Long Short-Term Control Network (RGLCN) and the Temporal Attention-Gated Recurrent Unit Convolutional Network (TAG-CNN) for power flow optimization and real-time voltage stabilization. RGLCN employs Deep Reinforcement Learning (DRL) for adaptive control, Long Short-Term Memory (LSTM) networks for disturbance prediction, and Graph Neural Networks (GNN) to model grid dynamics. TAG-CNN utilizes Temporal Convolutional Networks (TCNs) to forecast load changes accurately, employs attention mechanisms to focus on key factors, and incorporates Gated Recurrent Units (GRUs) to adjust parameters in real time. The MATLAB/Simulink code and related scripts for data generation for the IEEE-9-Bus PV-connected grid system are publicly available at https://github.com/obatolasamuel-ops/IEEE-9-BUS-SYSTEM-IN-PV-CONNECTED-GRID-SYSTEM. Simulation results on the IEEE 9-Bus system demonstrate that GAL-RNet achieves a total harmonic distortion of less than 1%, 98% accuracy, 95% system stability, and a computation time of 70 ms, outperforming traditional models such as RNNs, ANNs, and CNNs. However, the approach’s computational complexity limits its applicability in larger systems. Future research will focus on improving computational efficiency and enabling real-world deployment.