A Novel Identification Model for Potential Electric Hazard Discharge in Transmission Lines Using Multiscale Convolutional Domain Adversarial Network

The gradual development of potential electric hazards (PEHs) in transmission lines offers opportunities for early fault detection and timely intervention. In order to achieve hazard warning, it is essential to accurately identify PEH discharge types from the current signals collected in complex real-world conditions. Therefore, this study proposes a novel model named multiscale convolutional domain adversarial network (MCDA-Net) for robust PEH discharge type identification. MCDA-Net employs multiple parallel convolutional branches to extract features at different scales. Concurrently, a context-aware feature fusion (CAFF) module is designed to fuse multiscale features while reducing redundancy. Furthermore, a domain adversarial neural network (DANN) framework is introduced to jointly train the field-measured and experimental datasets, thereby enhancing the model’s generalization capability. The experimental results show that MCDA-Net achieves an average accuracy of 96.3% in PEH discharge type identification, significantly outperforming state-of-the-art (SOTA) methods, and demonstrating excellent inference speed and a lightweight model design.