Graph-Based Topological Embedding and Deep Reinforcement Learning for Autonomous Voltage Control in Power System.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Power Systems

Background:

• Increasing power system complexity and distributed energy resources challenge traditional voltage control methods.
• Existing deep reinforcement learning (DRL) approaches face limitations in utilizing topological data and computational efficiency for grid control.

Purpose of the Study:

• To develop an advanced DRL-based voltage control strategy for smart grids.
• To enhance state representation by incorporating power grid topology and optimize load shedding for improved voltage stability.

Main Methods:

• A hybrid DRL approach combining Graph Convolutional Networks (GCNs) for topological feature extraction and soft actor-critic (SAC) for continuous action space optimization.
• GCNs process higher-order grid topological information to enrich the state representation.
• SAC algorithm optimizes the load shedding strategy to balance economic costs and voltage stability.

Main Results:

• The proposed GCN-SAC method significantly reduced the amount of load shedding required.
• Improved voltage recovery levels were observed post-disturbance.
• Demonstrated strong control performance and robustness against complex disturbances and topological changes in the IEEE 39-bus system.

Conclusions:

• The integrated GCN-SAC method offers an innovative and effective solution for voltage control in modern smart grids.
• The approach enhances grid stability and operational efficiency by intelligently managing load shedding.
• Highlights the potential of combining graph neural networks with advanced DRL for complex power system management.