Attention-enhanced residual networks for real-time multi-label power quality disturbance classification with fast iterative filtering

The growing integration of renewable energy resources into modern power grids heightens their sensitivity to various types and degrees of perturbations, highlighting the relevance of real-time and comprehensive monitoring frameworks for power quality disturbances (PQDs). Notwithstanding, traditional approaches often struggle to identify complex PQD signals, particularly in noisy environments. This work proposes a lightweight classification framework that leverages fast iterative filtering (FIF), a multi-scale 1D residual network (ResNet), and efficient channel attention (ECA) for superior predictive performance. First, FIF with an adaptive mask length is employed to decompose PQD signals, effectively capturing dynamic perturbation characteristics. The B-spline filter is integrated to generate a non-negative and compact window, mitigating mode mixing while improving spectral resolution. Second, a deep neural network with parallel ResNet blocks for leveraging multi-scale receptive fields is designed to enhance feature discrimination by capturing both low-level and high-level signal patterns. Third, ECA modules are incorporated to adaptively reweight feature channels, minimizing redundancy and emphasizing disturbance-related patterns. Lastly, a multi-label-aware architecture is introduced to handle noisy and overlapping PQDs. Extensive experiments on synthetic datasets show that the proposed framework achieves superior accuracy, robustness, and real-time performance compared with state-of-the-art methods. Validation on two real-world PQD datasets further demonstrates its effectiveness, reaching accuracies of 97.94 % and 98.78 % with average inference times of 6.12 ms and 5.34 ms per sample, respectively. To support further research, the benchmark datasets and trained models are made publicly available. • Multi-label power quality monitoring framework for superior real-time disturbance recognition. • Deep neural architectures with parallel multi-scale residual blocks improve classification accuracy. • Fast iterative filtering (FIF) enables online perturbation-aware feature extraction. • Multi-site efficient channel attention (ECA) enhances feature representation.