Power quality disturbance classification based on dual-parallel 1D2D fusion of improved ResNet and attention mechanism

Abstract

The traditional power quality (PQ) disturbance recognition methods have limitations in feature extraction, making it challenging to handle nonlinear relationships and leading to restricted recognition accuracy. Therefore, this paper proposes a dual-parallel 1D2D fusion classification method based on an improved ResNet architecture and attention mechanism to improve feature extraction capability. First, the one-dimensional PQ disturbance signal is transformed into a two-channel two-dimensional picture using the Gram Angle Field (GAF) method. Furthermore, to address the potential information loss in deep feature extraction within grouped convolutional networks, we propose the Group Residual Convolutional Neural Network (GRCNN). The use of grouped convolutions significantly reduces parameter count, while the proposed parallel attention module (PAM) compensates for the lack of inter-group feature interactions during group convolution and captures key features from multiple perspectives. Meanwhile, to more effectively capture the temporal characteristics of the signal, a Long Short-Term Memory (LSTM) network is employed in parallel to process the one-dimensional PQ disturbance signals and extract temporal features. Finally, all extracted features are fed into a feature interaction fusion module (FIFM) for comprehensive integration, further enhancing feature representation. The experimental results demonstrate that accuracy can reach 98.67 % under a 40 dB noise level and remains as high as 91.85 % even in the presence of strong noise at 10 dB.