Deep Reinforcement Learning for Load Shedding Against Short-Term Voltage Instability in Large Power Systems.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Power Systems

Background:

• Dynamic load-shedding is critical for maintaining large power grid stability.
• Existing methods struggle with the complexity of real-time spatial and temporal power system dynamics.

Purpose of the Study:

• To introduce a novel deep Q-network for load-shedding (DQN-LS) for optimal power system stability.
• To enhance voltage recovery quality and probability through real-time, accurate load-shedding decisions.

Main Methods:

• Utilized a convolutional long-short-term memory (ConvLSTM) network to capture dynamic features and translation-invariant characteristics of short-term voltage instability.
• Developed a new reward function design for the deep Q-network.
• Tested the DQN-LS approach on the China Southern Grid (CSG) for scalability and performance evaluation.

Main Results:

• The proposed DQN-LS demonstrated superior voltage recovery performance compared to existing methods.
• The approach proved effective under diverse and uncertain power system fault conditions.
• Achieved real-time, fast, and accurate load-shedding decisions.

Conclusions:

• The DQN-LS offers an innovative and effective solution for dynamic load-shedding problems in large power grids.
• The study showcases unprecedented scale, performance, and methodology in load-shedding research.
• The approach significantly improves power system stability and voltage recovery probability.