Distribution network fault identification method based on multimodal ResNet with recorded waveform-driven feature extraction
Jian Wang, Bo Zhang, Dong Yin, Jinxin Ouyang
Abstract
The existing distribution network fault identification research mainly focuses on the identification of single-cause faults or high impedance fault, and lacks of comprehensive identification of fault types and fault causes due to insufficient fault samples for reference. In this paper, a fault identification method for distribution networks based on recorded waveform-driven feature extraction and multimodal ResNet is proposed. First, the waveform characteristics are analyzed according to the typical fault recording data, and the faults of different grounding media are modeled with the fault mechanism, which are used to generate the dataset of unbalanced faults for fault waveform inversion and fault feature extraction. Second, the three-phase and zero sequence Volt-Ampere curves from the head end of a feeder are used as the feature inputs. Then, a multimodal ResNet model based on RGB normalization and attention mechanism is constructed to extract the fault features. Finally, experimental results show that the proposed model achieves better fault identification compared to other neural networks and feature extraction methods. The proposed method performs well by transfer learning without extensive re-training for different distribution systems, and can identify actual fault data for small samples. Moreover, the proposed model is properly adapted to both noise and sampling frequency. • The fault simulation method based on physical mechanism and fault records can invert the actual waveforms. • The three-phase and zero-sequence Volt-Ampere curves are used as multi-input features. • Multimodal inputs are more applicable for comprehensive identification of fault type and cause. • Proposed method can accurately extract fault features and adapt to sampling frequency and noise. • Multimodal ResNet trained with actual fault dataset through transfer learning has good adaptability.