Abstract:
To improve the identification accuracy of high-voltage direct current(HVDC) transmission line faults under conditions of limited sample size and high impedance, a fault identification method for high-voltage direct current transmission lines that combines the gram angle difference field(GADF) and transfer learning using Residual Network 18(ResNet18-TL) is proposed. First, one-dimensional time-domain signals were transformed into two-dimensional angle-difference field maps using GADF. Subsequently, the weight parameters of a ResNet18 model pre-trained on the source domain ImageNet-1K dataset were transferred to a ResNet18 model with angle-field maps as the target domain, enabling the adaptive extraction of fault-related features for fault-type recognition. Experimental results demonstrate that, compared with other deep learning methods, the proposed approach can correctly identify internal positive-polarity ground faults, internal negative-polarity ground faults, internal bipolar short-circuit faults, and external faults under small-sample conditions, achieving an accuracy of 99.67%. Additionally, it exhibits a strong tolerance to transient resistance, noise resistance, and generalization capabilities.