Dual-attention-enhanced Alexnet for fault diagnosis in multi-energy complementary power systems
TL;DR
According to the findings, the dual-attention network together with the hybrid dilation approach improves the performance of multi-energy complementary power system fault diagnosis by increasing the convergence speed, monitoring accuracy, and reducing the number of false alarms.
Abstract
As a result of the rapid expansion in the integration of renewable energy and the coupling of various aspects like generation, grid, load, and storage, the most adopted methods in power fault diagnosis today have been the use of convolutional neural networks. Despite these advances, the existing models suffer from problems related to poor feature selection and inadequate representation of cross-time-scale interactions in multi-source and heterogeneous input environments. This challenge is further compounded by the distinct operational cycles and characteristics of multi-energy complementary units. This study builds a dual-attention optimized AlexNet fault diagnosis model. It introduces channel attention, spatial attention, and hybrid dilated convolution into the AlexNet backbone. The results show that the final loss decreases to 0.20, and the diagnostic accuracy under standard testing conditions reaches 97.88%. In a 30 min short-term monitoring scenario, the macro-average F1 score increases from 91.82% to 97.18%, and the macro-average recall increases from 91.06% to 96.79%. According to the findings, the dual-attention network together with the hybrid dilation approach improves the performance of multi-energy complementary power system fault diagnosis by increasing the convergence speed, monitoring accuracy, and reducing the number of false alarms. The proposed model is an efficient tool in power system monitoring and fault detection.