A cascade perceptron and Kohonen network approach to fault location in rural distribution feeders
Graphical abstract
Introduction
One of the central premises of electric power grids is to minimize the period of interruptions for customers served by distribution utilities [1]. It is estimated that approximately 80% of all interruptions at energy utilities are caused by faults in the distribution system [2]. Thus, in the event of a fault, its location and isolation are significant for a quick restoration of the energy supply, bringing benefits to maintenance and safety, and minimizing economic losses [3]. In [4], the authors highlight that the location and isolation of the faulty point in a distribution management system is one of the most important features because, when it is precise and useful, it guarantees the reduction of the periods of interruption and increases the reliability of the utility’s system.
Fault location is a topic widely studied and discussed in the literature. However, the location of faults in the distribution feeders is still considered a significant challenge. Most of these feeders have a large mesh with many branches to serve their customers [5], [6], [7]. The challenge increases when there is a low level of network automation. In these cases, an approach widely applied by utilities is the so-called Trouble Call system [8]. This system is based on customer complaint calls to estimate in which equipment the fault can be found. However, this system has some limitations, mainly in rural or remote areas, due to communication difficulties [9]. For this scenario, in the literature, there are several approaches to the theme with several different solutions to the problem.
In the approach shown by [10], the authors propose the use of a fault locator based on Machine Learning (Learning-Based Fault Locator — LBFL) using the Support Vector Machine (SVM) algorithm as methodology. The methodology is validated using an unbalanced distribution feeder with a nominal voltage of 34.5 kV. In the tests, several fault situations are considered by varying the fault location and resistance. Also, the frequency and harmonics changes (up to 5%) in the current and voltage signals were measured at the substation. However, the location of the fault is defined as a classification problem, in which each class corresponds to a zone of the distribution system, and they can have different extensions, which can hinder the assertiveness of the method.
The authors of [11] present a review of the technologies currently applied to identify and locate the faults in single-phase distribution grids. The paper makes a comparative analysis of the main methods by showing the positive points and their limitations. The work shows the use of Artificial Neural Networks (ANNs) as one of the types of technologies used for this approach, although it mentions that its applicability in real systems is a complicated issue. However, this work aims to apply this methodology to a real system in rural areas in the interior of Brazil to locate and classify faults in distribution grids.
In [12] and [13], the authors propose to locate faults in distribution grids based on measurements made by sensors installed throughout the system. In [12], using Wavelet Transform to detect signals, parameters such as the magnitude, angle, and phase sequence of the voltage are measured by low voltage sensors not synchronized along with the feeder’s extension. Also, these parameters are concentrated and processed on a computer. The method’s performance is analyzed considering the imbalance of the system and the existence of distributed generators (DG). In [13], high impedance fault location is performed using measurements from sensors, and based on three fault detection algorithms, Principal Components Analysis (PCA), Fischer Discriminant Analysis (FDA), and Support Vector Machine (SVM). The method is validated using the IEEE 13-node test system. The main disadvantage of using sensors along the feeder is attributed to the complexity and cost of the project. In the proposed approach, this problem is minimized since only measurements coming from the feeder outputs in the substation are used.
Nevertheless, the works [14], [15] measure current and voltage at a single point of the feeder in the original substation, and in [16], the current signal is obtained at the secondary of the IEEE 13 bus transformer used. These papers focus their research on locating high impedance faults in distribution networks and use techniques from Artificial Neural Networks (ANN), analysis of signals in frequency and time such as the Stockwell transform, and the wavelet transform.
Based on the above, the next subsection presents the major contributions of this research.
Given the previously mentioned, the primary purpose of this work is to propose a methodology for fault location in distribution systems, using ANNs capable of accurately estimating the equipment that caused the interruption of the electricity supply, employing current and voltage signals from the substation. This methodology uses two cascade-connected ANNs, which the first is a Multilayer Perceptron for fault classification and, the second, four Kohonen Self-Organizing Maps to locate them.
The use of two types of ANN working consecutively, the Multilayer Perceptron and the Kohonen Self-Organizing Maps, one for classification and the other for location, guarantee a differential for the proposed technique. In this last stage, one can highlight the advantages of using self-organizing maps concerning other networks to produce a two-dimensional and visual representation of the data submitted to their inputs through data clustering and because of their non-supervised training [17], [18]. Thus, input x output standards are not necessary to adjust this network.
The contributions of this research consist of the following aspects: (i) The method presents a simple but efficient methodology for collecting and treating current and voltage data, with the feeder output as the only acquisition point; (ii) the methodology is based on real data from a real distribution system. In this way, the method’s effectiveness tends to remain high when applied to the considered real system. The method validation (iii) considers the noise in both current and voltage signals, and (iv) the proposed approach is applicable for the most common types of faults (phase–phase, phase–phase–ground, phase–ground, and three-phase). Also, the algorithm (v) uses self-organizing maps for the fault’s location, eliminating the presence of desired outputs to the algorithm, once it intelligently self-organizes based on the system’s data, it improves its efficiency over time.
Section snippets
System modeling
In a real distribution system, the distribution substation (SE) chosen to be a model in this study is the one that supplies a city in the northern state of Paraná, in Brazil. The circuit in consideration is supplied by a power transformer consisting of capacitor banks (from SE and the grid), conductors, and loads at a voltage of 13.8 kV.
The power transformer is a three-phase three-winding power transformer with 25 MVA, two of which star-connected. The primary operates at 138 kV, the secondary
Artificial neural networks
In the electric power system, more specifically in the field of location and classification of faults in medium-voltage power networks, the ANNs have played an important role [19].
In this work, two types of neural networks are used sequentially: Multilayer Perceptron (MLP) for the classification of faults in the system studied and Kohonen Self-Organizing Maps for the location of defects through the interpretation of the regions where there are activated neurons in their maps. Therefore, the
Methodology
The distribution network consists of a series of shunts and branches. This fact adds greater complexity to this type of system, considering the need to install several fault protection equipment along with its extension.
When there is a fault in the distribution network, it is crucial to quickly restore it and identify which protection equipment was in operation [29]. Besides, considering that currently some of the equipment in rural grids are not automated; the objective of this work is to
Results
In this section, we present the distribution system tested in this work and the results of the two-step experiments: the first refers to the fault classification, and the other refers to their location stage.
Conclusion
This article presents a methodology for locating faults in distribution feeders located in the rural area, by identifying the protective equipment that acted at the time of the fault using current and voltage data available at the substation and artificial neural networks: MLP networks and SOM networks. This technique is an alternative to Trouble Call, commonly used in the electrical power companies, which has limitations overcome by the proposed method. One of Trouble Call’s main limitations
CRediT authorship contribution statement
Fabrício Augusto de Souza: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing - original draft, Writing - review & editing, Visualization. Marcelo Favoretto Castoldi: Conceptualization, Validation, Resources, Writing - original draft, Writing - review & editing , Supervision, Project administration, Funding acquisition. Alessandro Goedtel: Conceptualization, Validation, Resources, Writing - original draft, Writing - review & editing,
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
Authors would like to thank the financial support of the National Council for Scientific and Technological Development (CNPq) under grant #474290/2008-3, #473576/2011-2, #552269/2011-5, #307220/2016-8 and COPEL Distribution S/A, Londrina-Brazil .
References (32)
- et al.
High-speed fault detection and location in DC microgrids systems using Multi-Criterion System and neural network
Appl. Soft Comput.
(2019) - et al.
Distribution network fault section identification and fault location using wavelet entropy and neural networks
Appl. Soft Comput.
(2016) - et al.
Single phase fault location in electrical distribution feeder using hybrid method
Energy
(2016) - et al.
Performance analysis of a learning structured fault locator for distribution systems in the case of polluted inputs
Electr. Power Syst. Res.
(2019) - et al.
Review of methodologies for earth fault indication and location in compensated and unearthed MV distribution networks
Electr. Power Syst. Res.
(2018) - et al.
Non-linear high impedance fault distance estimation in power distribution systems: A continually online-trained neural network approach
Electr. Power Syst. Res.
(2018) - et al.
High impedance fault detection in power distribution systems using wavelet transform and evolving neural network
Electr. Power Syst. Res.
(2018) - et al.
Introducing self-organized maps (SOM) as a visualization tool for materials research and education
Results Mater.
(2019) - et al.
Artificial neural network based fault detection and fault location in the dc microgrid
Energy Procedia
(2016) - et al.
SOMprocessor: A high throughput FPGA-based architecture for implementing Self-Organizing Maps and its application to video processing
Neural Netw.
(2020)
Denoising autoencoder self-organizing map (DASOM)
Neural Netw.
Clustering and visualization of bankruptcy trajectory using self-organizing map
Expert Syst. Appl.
NATIONAL ELECTRIC POWER AGENCY – Aneel: Procedures for distribution of electricity in national electric system – PRODIST
Fault diagnosis in low voltage smart distribution grids using gradient boosting trees
Electr. Power Syst. Res.
Integrated fault location and isolation strategy in distribution networks using Markov decision process
Electr. Power Syst. Res.
Time reversal applied to fault location in power networks: pilot test results and analyses
Int. J. Electr. Power Energy Syst.
Cited by (5)
Faulty feeder detection based on image recognition of current waveform superposition in distribution networks[Formula presented]
2022, Applied Soft ComputingCitation Excerpt :Minimizing the period of power interruption is essential for improving power quality and customer satisfaction in the routine operation of electrical power grids. Recent research has shown that most interruptions are caused by faults in distribution networks [1], where single line-to-ground (SLG) faults account for 80% of all faults [2]. Serious faults may occur, resulting in equipment damage, casualties, and power outages if an SLG faulty feeder cannot be reliably identified [3].
Faulty feeder detection based on fully convolutional network and fault trust degree estimation in distribution networks
2022, International Journal of Electrical Power and Energy SystemsCitation Excerpt :IMFs of zero-sequence currents obtained by VMD are utilized as the input of long short-term memory neural network (LSTM), and the faulty feeder can be identified according to the output of LSTM [19]. In addition, neural network [20–21], clustering methods [22–23], and ensemble learning algorithm [24–25], are also employed as classifiers for detecting the faulty feeder. However, these ML-based methods have not been applied in practical distribution systems, and this is mainly because: 1) the extracted high-level features lack clear physical meaning; 2) the detection process has poor interpretability; 3) the generalization capability needs to be further verified when the detection methods encounter new fault scenarios.
Quantum annealing algorithm for fault section location in distribution networks
2023, Applied Soft ComputingAnalysis of Neural Network Training Algorithms for Implementation of the Prescriptive Maintenance Strategy
2022, Materials Research Proceedings