Article Details

Abstract

To address the issues of fixed feature fusion and high computational complexity in traditional Power Quality Disturbances (PQDs) signal recognition models, an Adaptive Gramian Time Frequency Enhancement Network (AGTFENet) is proposed. Firstly, a noise reduction strategy based on the Gram matrix is introduced to process one-dimensional input signals. A three-branch parallel architecture is adopted to handle the original signal, Gram noise-reduced signal, and frequency spectrum respectively. Secondly, multiple feature learning modules are stacked, and Depthwise Separable Convolution are used to extract features from each branch. Finally, adaptive average pooling and an adaptive weight mechanism are introduced to dynamically adjust the contributions of features from each branch, achieving weighted fusion of features and classification for disturbance signals. Simulation results show that AGTFENet achieved recognition accuracies of 98.9%, 98.7%, 98.5%, and 97.8% under different noise levels (no noise, 40 dB, 30 dB, 20 dB), respectively, outperforming other classification models. Moreover, benefiting from its lightweight design, it demonstrates excellent performance in terms of computational complexity.

Keywords

PQDs; Gramian Noise Reduction; adaptive mechanism; Depthwise Separable Convolution; disturbance recognition

References

1. [1] WANG S X，CHEN H W. A Novel Deep Learning Method for the Classification of Power Quality Disturbances Using Deep Convolutional Neural Network [J].Applied Energy，2019，235：1126-1140.
2. [2] Li Tiezhu, Wang Wei, Ding Chaoqian, etc Research on Fault Location in Distribution Networks Based on One Dimensional Dense Connected Convolutional Networks [J]. Manufacturing Automation, 2024, 46 (1): 41-44
3. [3] Zhang Beide, Qiu Jie, Lou Guangxin, etc A lightweight power quality disturbance identification model based on CNN and Transformer [J]. Electric Power Engineering Technology, 2025, 44 (1): 69-78
4. [4] Wang Gaofeng, Zhang Zhuoshi, Gao Man, etc A power quality disturbance classification method based on multi feature fusion convolutional neural network combined with Transformer [J]. Journal of Beihua University: Natural Science Edition, 2025, 26 (1): 115-124
5. [5] Qin Risheng, Xu Zhi, Kuang Hua, etc A power quality disturbance detection method based on dual resolution S-transform and improved multi-scale ResNet model [J]. Guangdong Electric Power, 2024, 37 (7): 68-77
6. [6] Chen Lucheng, Qin Chenggang, Lu Xiaoping, etc Fault diagnosis method for bearings under variable operating conditions based on joint distributed adaptive parallel convolutional neural network [J]. Mechanical Design and Research, 2025, 41 (1): 221-225
7. [7] JIA L S，CHOW T W S，YANG Y X. GTFE-Net: A Gramian Time Frequency Enhancement CNN for Bearing Fault Diagnosis [J/OL].Engineering Applications of Artificial Intelligence，2023，119：105794[2025-05-26].https://doi.org/10.1016/j.engappai.2022.105794.
8. [8] Fan Youmin Parameter adaptive setting and global optimization improvement of gradient descent method [D]. Lanzhou: Northwest Normal University, 2024
9. [9] MÜLLER R，KORNBLITH S，HINTON G. When Does Label Smoothing Help? [J/OL].arXiv:1906.02629 [cs.LG].[2025-06-03].https://doi.org/10.48550/arXiv. 1906.02629.
10. [10] SRIVASTAVA N，HINTON G E，KRIZHEVSKY A，et al. Dropout: A Simple Way to Prevent Neural Networks From Overfitting [J].Journal of Machine Learning Research，2014，15（1）：1929-1958.
11. [11] IEEE. IEEE Std 1159-1995(R2001), IEEE Recommended Practice for Monitoring Electric Power Quality [S].New York：IEEE，2001.

Funding

Natural Science Foundation of Anhui Province (2208085QE167)

About the authors

Zhang Xinyu (2004-), female, Han ethnicity, from Bozhou, Anhui, undergraduate student, research interests include deep learning and power quality disturbances.