Multi-Agent Graph-Attention Deep Reinforcement Learning for Post-Contingency Grid Emergency Voltage Control.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Control Systems

Background:

• Grid emergency voltage control (GEVC) is critical for power system stability, preventing cascading outages.
• Existing deep reinforcement learning (DRL) methods struggle with the dynamic, multi-agent nature and uncertainties of real-world power grids.
• Data efficiency and control performance are key challenges for current DRL-based GEVC.

Purpose of the Study:

• To propose a novel multi-agent graph-attention (GATT)-based DRL algorithm for GEVC in multi-area power systems.
• To enhance decision accuracy and data efficiency through graph convolutional network (GCN) agents.
• To improve cooperative learning and scalability in dynamic grid environments.

Main Methods:

• Developed GCN-based agents for feature representation of graph-structured voltages.
• Implemented a graph-attention mechanism for effective information sharing among multiple agents.
• Tested the multi-agent GATT-DRL algorithm in a benchmark IEEE system.

Main Results:

• The proposed GATT-DRL algorithm demonstrates improved data efficiency and decision accuracy.
• Effective information sharing via attention mechanisms enhances cooperative learning among agents.
• The method shows superior performance and stability compared to existing multi-agent DRL algorithms in complex grid scenarios.

Conclusions:

• The multi-agent GATT-DRL approach offers a scalable and stable solution for GEVC in dynamic power systems.
• This method effectively addresses challenges posed by high uncertainties and complex grid operations.
• The findings highlight the potential of advanced DRL techniques for robust power system control.