A power quality disturbance classification method using a hybrid transformer and discrete wavelet transform model

The growing development of distributed energy sources and the increasing reliance on advanced power electronic devices have significantly raised the complexity of modern power grids. Ensuring a stable and reliable power supply requires accurate identification of power quality disturbances. This study presents a disturbance classification method that integrates Discrete Wavelet Transform (DWT) with a customized Transformer-based model to improve recognition performance under complex conditions and to enhance robustness in noisy environments. The proposed approach employs DWT for multi-scale time-frequency decomposition of disturbance signals, while frequency-domain energy spectra are extracted through Fast Fourier Transform (FFT), forming a dual-channel input structure. The Transformer model is adapted by retaining only the encoder structure and integrating a HiLo Attention mechanism to enhance the extraction of disturbance patterns in high and low-frequency ranges. A gate attention mechanism is incorporated to integrate multi-channel features and dynamically merge the complementary data provided by DWT and FFT. In evaluations involving 29 distinct disturbance types, the method achieved a classification accuracy of 99.41 % under clean signal conditions and sustained performance levels exceeding 98 % in environments with signal-to-noise (SNR) ratios between 20 and 50 dB, as well as in simulated signal experiments, outperforming existing mainstream models. Results from ablation analyses verified the importance of incorporating multi-scale features, the HiLo Attention mechanism, and the encoder design in improving classification accuracy and noise robustness.