A Novel Data-Driven Multi-Agent Reinforcement Learning Approach for Voltage Control Under Weak Grid Support

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Power Systems

Background:

• Weak voltage support in distribution networks with high PV integration poses challenges.
• Active voltage control is crucial for grid stability and efficiency.

Purpose of the Study:

• To develop a multi-agent deep reinforcement learning (MADRL)-based coordinated control for PV clusters.
• To enhance voltage compliance and energy efficiency in weak grids.

Main Methods:

• Formulated voltage control as a decentralized partially observable Markov decision process (Dec-POMDP).
• Employed a centralized training with decentralized execution (CTDE) framework.
• Designed barrier functions for reward shaping to balance voltage and efficiency.

Main Results:

• The MADRL framework ensured voltage compliance and reduced network losses.
• The MADDPG algorithm achieved a 91.9% Controllability Ratio (CR) with low power loss (0.0695 p.u.).
• Demonstrated superior performance compared to optimal power flow (OPF) and droop control.

Conclusions:

• The proposed MADRL approach effectively improves voltage stability and energy efficiency.
• The method is robust under model-free and communication-constrained weak grid conditions.