15 May 2020 in IEEE Transactions on Smart Grid
Incipient faults in power distribution systems potentially lead to catastrophic failures. Detection of incipient faults contributes to proactive fault management and predictive maintenance, which effectively improves power supply reliability. Since the faults are infrequent and transient, few samples can be procured in real applications. In this paper, a detection method based on human-level concept learning (HLCL) is proposed to address this problem. The method contains two steps: human-level waveform decomposition (HLWD) and hierarchical probabilistic learning (HPL). HLWD, inspired by human perception, decomposes waveform into primitives: segments of general shape and residuals, which identify a waveform. HPL learns waveform through a generative process based on primitives, where the probability for an abnormal event to be an incipient fault can be hierarchically calculated. Experiments on simulation data and field data, which contain subcycle incipient faults, multicycle incipient faults, permanent faults and transient disturbances, indicate that the proposed method outperforms other three generally used classifiers.
8 November 2021 in IEEE Transactions on Power Delivery
The power distribution system’s fault root cause classification is an important but challenging problem. Traditional classifiers fail to achieve high accuracy and good generalization performance due to data insufficiency. A large volume of unlabeled data is available, which can be utilized to improve classification performance. This paper proposes a novel classifier called Robust Semi-Supervised Prototypical Network (RSSPN) based on Prototypical Network architecture and semi-supervised learning to address this issue. The proposed method can mine information from unlabeled data to improve the generalization ability and classification accuracy. Furthermore, RSSPN adopts the idea of meta-learning to obtain the “few-shot learning” ability for identifying new fault classes using very few samples encountered during the operation and update online. Experiments have been conducted on a dataset consisting of 1152 labeled samples belonging to 12 different classes and 10000 unlabeled samples. The accuracy of the proposed method is significantly better than the traditional classifiers.