Article Preview
TopIntroduction
Efficient and accurate fault identification of distribution network can support the stability and controllability of smart grid (Montakhab, Adams, 1998). With the access of distributed power source, the traditional distribution network has changed from the original radial network to the complex active distribution network with interconnected power sources and users (Tajdinian, et al, 2020). At the same time, the applicability and accuracy of traditional distribution network fault location methods are reduced, also bringing difficulties to relay protection (Le, et al, 2020; Chen, et al, 2020).
The distribution network fault is hidden. If the fault line is not cut off in time, it will cause great potential safety hazards. For example, in case of single-phase grounding fault, the line voltage remains symmetrical, but the phase voltage of non-fault phase will increase by three times. At this time, there is a risk of insulation breakdown of power equipment, which may cause two-point or multi-point grounding short-circuit fault, leading to further development of the fault (Liang, et al, 2020; Hagh, et al, 2019; Miguel, 2022). Therefore, the distribution network fault must be identified in time and handled quickly.
The fault treatment of distribution system mainly includes distance fault identification method and matrix algorithm based on Distribution Automation Terminal (Xie, et al, 2020; Gao, et al. 2021; Yao, et al, 2021). Function approximation is used as the basic goal to establish the 0-1 integer optimization model, and then faults are located through optimization algorithm, but there are some problems such as slow calculation speed, difficult construction of function model and poor convergence of optimization algorithm (Jia, et al, 2019; Xu, et al, 2019). It is not enough to cope with the increasingly complex active distribution network.
The emergence of deep learning network provides a new solution to fault identification of active distribution network. Deep neural network is constructed by using massive data training, the characteristics of input data are automatically extracted, and induction and classification are implemented accordingly (Ganjkhani, et al, 2021; Hou, et al, 2022; Zhao, Barati, 2021). At present, many researchers have carried out power grid fault identification and research based on deep network. Luo et al (2019) introduced the automatic encoder into the deep learning network model to realize the fault identification and analysis of radial distribution network; (Sun et al, 2021) proposed an adaptive long short memory network regression model to realize the state detection and fault identification of power transmission network by establishing the corresponding relationship between similar time factors and long and short-term memory network (LSTM). Based on the high-voltage direct current high voltage direct current system (HVDC), Wang, He and Li (2021) optimized convolutional neural network (CNN) and LSTM network models to realize fault identification and judgment of transmission lines; Rai, Londhe and Raj (2020) focused on the scene of active distribution network, and used CNN to build a fault identification model to support its safe operation. Based on the convolutional neural network (CNN), Zhang et al (2022) constructed a network structure fully suitable for power grid fault diagnosis, and took the minimum cross entropy as the goal to mine the deep fault features to achieve the fault diagnosis analysis of AC/DC transmission system. Wei et al (2021) used two bidirectional short-term and short-term memory networks as the basic classifiers, and applied the cross-entropy loss function and cost-sensitive loss function to the two classifiers respectively, effectively reducing the impact of sample category imbalance in fault event recognition. Yan et al (2022) input the word vector into the CNN deep learning model for training, and introduce the DSA mechanism to improve the CNN model according to the characteristics of the power grid alarm information.