Using the Multi-head Self-attention Mechanism to Analyze and Extract Key Features from Power Grids with Complex Structures

As the proportion of distributed renewable energy generation rises, the structure of regional power grids is becoming more intricate and varied. Representing the spatial characteristics of the power grid accurately is crucial for dispatchers to gather essential feature information. Traditional methods, like graph convolutional neural networks, are limited to capturing information only from neighboring nodes. This paper, referencing Bidirectional Encoder Representations from Transformers (BERT), proposes a power grid representation learning method based on the Multi-head Self-Attention (MSA) for feature extraction of power grids. The proposed method’s effectiveness in addressing the challenge of power grid feature extraction is demonstrated through an improved IEEE 9-bus test system case study.