Unsupervised High Impedance Fault Detection in Distribution Networks With Autoencoder

Detection of high impedance faults (HIFs) in power distribution networks remains a challenging task due to the low current magnitudes and varying characteristics associated with these faults. Conventional overcurrent relays often fail to detect HIFs effectively. Machine learning methods have shown promise as a viable solution, but most previous studies rely on supervised learning techniques to detect HIF by performing classifications that require a significant amount of HIF data that is difficult to acquire in real-world scenarios. As a result, the dependability and applicability of these models are compromised when encountering load profiles and faults not present in the training dataset. This paper introduces an unsupervised fault detection framework that utilizes autoencoder and principal component analysis (PCA)-based monitoring techniques. This approach effectively detects abnormal correlation structures in current waveforms that deviate from normal load conditions. Since HIFs can induce minor random distortions in current waveforms, leading to inconsistencies in the correlation between current measurements across cycles compared to normal load conditions, the proposed method is well-suited for detecting these faults. We evaluate the effectiveness of the proposed method using both a simulated dataset from the IEEE 123-bus distribution system and real data from a 4.16 kV overhead distribution system. The performance is compared with other data-driven fault detection methods, including a commercially available solution. The results demonstrate the effectiveness of our proposed method.