What type of study is this?

September 10, 2025

Grid section overload control method based on reinforcement learning

Key Points

The GAT-KAN-D3QN model improved control accuracy, reducing relative error to 4.3% in managing overload.
Compared to DQN-DDPG, the new model showed a 1.4% improvement in managing power grid sections.
Control strategies dynamically address sections' overload issues while maintaining computational efficiency.
The research validates the effectiveness of reinforcement learning in balancing control accuracy and efficiency in power grids.

Abstract

With the expansion of the scale of power grids and the increase in uncertainty of renewable energy, the grid section overload problem is becoming increasingly prominent. The current grid section overload control method overlooks the coupling relationship between the control of each section, leading to the flow adjustment of one section causing overload in other sections. Additionally, the traditional strength of learning control suffers from problems of overestimation and high computational cost. This paper focuses on the power grid generator control as the object, utilizing the GAT model to dynamically extract node topology features and leveraging the advantages of the KAN model's lightweight nonlinear processing. It combines D3QN to control over-limit generators in the power grid section, thereby improving control accuracy. The study first uses GAT to calculate the embedding matrix of the power grid node, and uses KAN to calculate the embedding representation of the section task. Then, a multi-section task attribution graph is generated based on node embedding and task embedding, and a weighted pooling strategy is used to extract the final power grid diagram feature representation based on the embedding matrix and the multi-section task attribution graph. Finally, the power grid diagram feature representation is used as the input of the D3QN model to achieve accurate and dynamic over-limit control of generators in power grid sections. The experiment was conducted based on the IEEE 39-node system, and the results were refined into three dimensions: discrimination accuracy, dynamic response, and computational efficiency. The results showed that the relative error of the GAT-KAN-D3QN model was only 4.3%, which was 1.4% lower than the DQN-DDPG method, and the scheduling efficiency reached 92.7%. The results show that the reinforcement learning (RL) method in this paper can effectively improve the control accuracy and can better cope with the dynamic changes and uncertainties in the power grid, solving the key contradiction of “accuracy-efficiency trade-off”.

Mark Helpful

Bookmark

Relay