Improved Fault Classification and Localization in Power Transmission Networks Using VAE-Generated Synthetic Data and Machine Learning Algorithms

Abstract

The reliable operation of power transmission networks depends on the timely detection and localization of faults. Fault classification and localization in electricity transmission networks can be challenging because of the complicated and dynamic nature of the system. In recent years, a variety of machine learning (ML) and deep learning algorithms (DL) have found applications in the enhancement of fault identification and classification within power transmission networks. Yet, the efficacy of these ML architectures is profoundly dependent upon the abundance and quality of the training data. This intellectual explanation introduces an innovative strategy for the classification and pinpointing of faults within power transmission networks. This is achieved through the utilization of variational autoencoders (VAEs) to generate synthetic data, which in turn is harnessed in conjunction with ML algorithms. This approach encompasses the augmentation of the available dataset by infusing it with synthetically generated instances, contributing to a more robust and proficient fault recognition and categorization system. Specifically, we train the VAE on a set of real-world power transmission data and generate synthetic fault data that capture the statistical properties of real-world data. To overcome the difficulty of fault diagnosis methodology in three-phase high voltage transmission networks, a categorical boosting (Cat-Boost) algorithm is proposed in this work. The other standard machine learning algorithms recommended for this study, including Support Vector Machine (SVM), Decision Trees (DT), Random Forest (RF), and K-Nearest Neighbors (KNN), utilizing the customized version of forward feature selection (FFS), were trained using synthetic data generated by a VAE. The results indicate exceptional performance, surpassing current state-of-the-art techniques, in the tasks of fault classification and localization. Notably, our approach achieves a remarkable 99% accuracy in fault classification and an extremely low mean absolute error (MAE) of 0.2 in fault localization. These outcomes represent a notable advancement compared to the most effective existing baseline methods.

A Fault Location Method for Medium Voltage Distribution Network Based on Ground Fault Transfer Device

“…However, the analysis scale and wavelet basis function need to be selected according to the fault characteristics, and the extraction range is limited to a certain frequency band, so it is impossible to extract the TW signal with a single frequency. Based on the Morlet wavelet, the S transform solves the problem that the frequency of the STFT window cannot be adjusted, and it has the advantage of multi-resolution of the wavelet transform [23]. Therefore, this paper uses the S transform to identify the fault TW head.…”

IIR Shelving Filter, Support Vector Machine and k-Nearest Neighbors Algorithm Application for Voltage Transients and Short-Duration RMS Variations Analysis

“…Despite the remarkable contribution, this area presently is still open [2], not commercialized and not applied in practice. Various techniques are applied in the feature selection and optimization stage: for example, [19] uses a customized version of forward feature selection (an iterative method starting with no features), and [33] uses adaptive ABC-PSO for optimal feature selection. The frequency domain is often studied as well-authors must pay attention (and sometimes do) to various encountered (and not always widely known) challenges such as Heisenberg uncertainty, window function and size, spectral leakage, spectral aliasing, etc.…”

“…18 In this table, statistical parameters stand for well-known characteristics such as the mean, standard deviation, range, skewness, excess kurtosis, etc., including any manipulation (operation) with them. 19 The following parameters are required for voltage sag generation with the synthetic signal generator: signal duration, fault start time, fault duration, nominal RMS, signal frequency, sampling frequency, no-fault amplitude, phase angle offset, number of smoothing timepoints, fault severity, signal rotation, noise percentage, and the label (the type of the fault to be generated). 20 The ensemble utilizes discriminant, KNN and DT weak learners (three learners, nine methods).…”

A Multi-Input Convolutional Neural Network Model for Electric Motor Mechanical Fault Classification Using Multiple Image Transformation and Merging Methods

“…Multisensor approaches using 2D deep learning frameworks are also being explored for distributed bearing fault detection [19]. Lastly, researchers in the field are exploring the potential of synthetic data generation using variational autoencoders for enhanced fault classification and localization in transmission networks [20].…”