Application of the CNN-GRU-MHA Model in Partial Discharge Recognition

Partial discharge (PD) is one of the most important factors affecting the stable operation of high-voltage equipment and poses a great threat to the reliability and safety of power systems. Although some deep learning-based methods, such as Convolutional Neural Networks (CNN), have been proposed for PD detection, there are still some shortcomings in terms of global dependencies. In this paper, it is proposed to develop a model of Convolutional Neural Networks-Gated Recurrent Unit-Multi-Head Attention (CNN-GRU-MHA). Wavelet transformation is adopted for noise suppression. reduction, together with statistical feature downsampling for reducing data dimensionality. In the model, CNN extract localized temporal features, GRU capture sequential temporal patterns, and MHA focuses on modeling global dependencies while highlighting critical features. Experimental evaluations reveal that the model reaches a Matthews correlation coefficient of 74.78% and a recall rate of 74.61%, performing better than some conventional methods in terms of both accuracy and robustness. This model securely and effectively provided a method of enhancing power system safety and stability.