Comparative Study of Fault Detection Methods in Electrical Power Systems Using Hessian Locally Linear Embedding and Deep Autoencoders

Faults in power lines greatly impact the reliability of electrical distribution systems (EDS), causing instability and heightened risks of outages. Accurate detection of electrical faults is required for optimizing the operation and maintenance of the power system. In this paper, we conduct a comparative study of two advanced approaches for detecting faults in EDS: Hessian locally linear embedding (HLLE) and convolutional autoencoders (CAE). The first approach utilizes HLLE to convert data with high dimensionality into low dimensionality embedding coordinates. This process helps capture underlying variations and patterns, enabling robust feature extraction. The detection of faults is achieved by applying the Mann-Whitney U test. The second approach leverages deep learning techniques, specifically deep autoencoders for anomaly-based fault detection and CAEs for dimensionality reduction. Deep autoencoders offer powerful capabilities in pattern classification, while CAEs require less training time due to fewer parameters. Simulations conducted on an electrical distribution system showcase the effectiveness of both approaches. However, the deep autoencoder approach outperforms the HLLE-based method, achieving an accuracy of 97% on simulated datasets. This comparative study provides valuable insights into the developments of robust detection systems in EDS.