Identification and localization of high impedance faults in distribution networks

Abstract

High impedance faults continue to pose significant challenges in ensuring reliable power systems due to their low fault currents and complex characteristics. HIFs can occur in various scenarios, such as when a live conductor breaks and either contacts the ground or nearby trees touching an energized overhead line conductor. These faults usually generate fault currents with lower amplitudes than the threshold values set for protection systems and are largely undetectable by traditional protection equipment. Due to their arcing nature and difficulty in detection and location, high impedance faults have the potential to cause fires and damage network equipment. Therefore, they pose serious risks not only for energy distribution companies but also for human and environmental safety. This situation has prompted researchers to focus on high impedance faults for over forty years. To date, numerous studies have been conducted to explore and mitigate HIF-related problems caused by high impedance faults. However, a universal solution for detecting or locating HIFs remains absent. Therefore, an economical and applicable solution, especially for the localization of HIFs, was sought within the thesis while enhancing the understanding of HIFs. In the early years, when the subject was studied, there was limited information available about the behaviors of HIFs. This led all early studies on the subject to use signals obtained from laboratory tests or staged faults in the field. Through these efforts, researchers were able to analyze actual fault signals and identify the key characteristics of HIFs. Once these unique behaviors were recognized, models that replicate one or more behaviors of HIFs were proposed. Most subsequent studies have utilized these proposed models to identify or locate HIFs. There are two main reasons for this. First, conducting field and laboratory tests can be quite challenging. The other is that many scenarios can be explored within a simulation environment using HIF models, which can be easily adapted to fit various distribution networks through simple modifications. However, it is essential to employ accurate and complete models to examine clearly the influences of such faults. Thus, the thesis presents a detailed explanation of HIF characteristics and proposed HIF models with their represented behaviors in chronological order after the introduction section. After examining and clearly understanding the fundamental concepts of HIF, the thesis shifts its focus to studies related to the detection and localization of HIFs within power systems. Comprehensive literature reviews were conducted for both topics. The proposed methods were described in detail, considering all key indices, such as system changes, noise conditions, test networks, and the HIF model. In addition, detection and localization techniques were categorized and compared within their classes, emphasizing their pros and cons. Finally, outcomes from the analyzed articles were addressed under the literature review section. In the third section of the thesis, the HIF experiments carried out in Istanbul Technical University Fuat Külünk High Voltage Laboratory are described in detail. While the main characteristics of HIF have been outlined in existing literature, it is crucial to assess whether these fault characteristics are observed consistently in real fault scenarios. Therefore, extensive laboratory experiments on HIF were designed and executed on various contact surfaces to capture realistic HIF signals. Due to the limitations of the experimental setup, the amplitudes of HIF currents obtained in the laboratory are typically lower than those of a fault observed in the field. However, by scaling the relevant signals as described in Section 4 and using them in simulations, the drawbacks of utilizing signals obtained from experimental tests have been eliminated. Through the analysis of the fault data collected from the laboratory measurements, a deeper understanding of the nature of HIF was achieved while the documented characteristics of HIF were validated. After performing a literature review and laboratory experiments, two different feature approaches for HIF detection were tested and compared by utilizing HIF signals captured in the laboratory. According to the literature, most published detection methods have yielded promising outcomes, while machine learning methods have demonstrated superior accuracy. Therefore, in addition to employing a different set of input features, two distinct machine-learning algorithms were analyzed in the thesis. Furthermore, events exhibiting similar characteristics to HIFs, along with HIF signals obtained in the laboratory and modeled, were included in the simulations to verify the effectiveness of the methods. The study also examined how different configurations of machine learning techniques, various mother waves, and varying noise levels affect the results. Findings indicated that the ANN algorithm produced more accurate detection results for noisy and clean signals, while the standard deviation approach used as input features outperformed the energy sum approach. However, it was observed that both using HIF signals obtained in a laboratory environment and adding noise to the signals caused a reduction in the overall performance of the methods. After studying distinguishing HIFs from other faults, which is the first step before localization, the thesis focuses on pinpointing the location of HIFs accurately. According to the conducted studies, there have been relatively few studies on locating HIFs compared to distinguishing them. Approximately 80% of studies have employed either impedance-based methods or machine learning techniques to pinpoint the exact location of HIFs. In recent years, machine learning methods that utilize features extracted from voltage and current signals have gained popularity due to their lower error rates. However, none of these studies have implemented any feature selection method or provided a rationale for their chosen features. Therefore, it is aimed to investigate the influence of feature selection techniques on HIF localization in the thesis. In this context, firstly, current waveforms of high impedance faults were obtained with the help of simulation studies. Then, a seven-level discrete wavelet transform was applied to the current signals, and five statistical parameters were calculated for each wavelet coefficient, resulting in a total of 40 features. These statistical parameters are standard deviation, variance, skewness, kurtosis, and the sum of energy content. Feature selection methods were then applied to the created feature pool, and the performance of ANNs trained with the selected feature subsets was evaluated. The features proposed from each feature selection method were tested 250 times on artificial neural networks, and the averages of all results were taken into account in the evaluations. Based on the comparisons, wrapper methods outperformed others when the ANN settings were the same. However, features chosen by the random forest method consistently achieved lower error distances across different ANN configurations and demonstrated higher accuracies in test data, even under noisy conditions. Overall, all feature selection methods successfully reduced computational complexity and enhanced accuracy, even in noisy conditions. Furthermore, it was observed that these methods could successfully identify significant and irrelevant features within the data. Following an assessment of available detection and localization methods by aiming to complete the missing areas in existing publications, this thesis proposes a new data-driven HIF location method that can identify the faulty branch associated with HIF and accurately pinpoint its location in a distribution network. This innovative method utilizes the traveling wave approach, Discrete Wavelet Transform (DWT), and machine learning technique. Since ANN is used commonly in localization studies, the proposed methodology incorporates this method as the chosen machine learning method. Normally, estimating the fault distance accurately using traveling wave methods requires identifying each reflected and refracted traveling wave. However, using this method directly can be challenging, especially in distribution networks with multiple short branches. Therefore, the proposed method utilizes information from a certain number of selected traveling waves that can be easily distinguished from the current signal in ANN. By employing this selection process alongside a combination of traveling wave and machine learning techniques, the aim was to overcome certain weaknesses of each method. The proposed approach has been evaluated across multiple scenarios, considering different inception angles, machine learning techniques, load variations, extreme conditions, and distribution networks to validate its robustness and efficiency. The findings indicate that the proposed method exhibits high accuracy in detecting the faulty network branch and has a minimal error rate in calculating the distance of the fault location.