Work

Power quality disturbance classification via a time-frequency feature-fused transformer model with cross-attention mechanism

Year: 2025

Type: article

Abstract: With the increasing penetration of renewable energy into power systems, power quality disturbances (PQDs) exhibit composite patterns and dynamic time-varying characteristics. To address the challenges of dimensional redundancy and alignment bias in conventional multimodal feature fusion, this study proposes a novel classification framework integrating time-frequency domain analysis with a Transformer-Cross Attention architecture. The core innovation lies in a dual-channel parallel processing design: temporal features capturing local details via convolutional neural networks (CNNs) are dynamically aligned with spectral features encoding global dependencies through Transformer encoders, followed by adaptive cross-domain fusion governed by learnable cross-attention mechanisms. A comprehensive dataset covering 20 composite PQD types under multi-noise scenarios (signal-to-noise ratio 0–60 dB) was constructed for validation. Experimental results demonstrate classification accuracies of 99.95% in noise-free conditions and 98.75% under 20 dB noise, outperforming state-of-the-art baseline models (CNN-BiLSTM, CNN-Transformer) by at least 3.2 percentage points, thereby demonstrating its robust superiority in complex grid environments.

Source: Electric Power Systems Research

DOI: 10.1016/j.epsr.2025.112330

Authors: Tingling Wang, Jiaxi Zhuo, Yonghui Hou, Zifan Lu, Yongqing Li

Institution: North China University of Water Resources and Electric Power

Language: English

Topic: Power Quality and Harmonics

Subfield: Electrical and Electronic Engineering

Field: Engineering

Domain: Physical Sciences

SDGs: Affordable and clean energy, Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,810