Performance Analysis of Machine Learning Classifiers for Fault Type Classification in Multi-Bus Power Systems

Author Name(s): Janak Raj Chadha, Megha Pandey, Tejash Sharma, N.R. Patne, Piyush Khadke
Author Email: janakrajchadha@gmail.com

Abstract

In the past decade, multiple attempts have been made to classify faults on transmission lines using mathematical transformations and machine learning algorithms. The available literature primarily contains analysis of simple single line systems. This paper summarises the performance analysis of machine learning classification models on more complex power system models. Two different power system models are taken into consideration, a standard 3 bus 3 generator 2 load software model and a hardware simulator manufactured by SAMAR Instruments, Italy. The analysis of the first system captures the effects of multi-bus multi- generator systems on the performance of classifiers. The second system focuses on the results of testing the models on a hardware simulator. Discrete wavelet transform has been utilised for feature extraction and the performance of different classifiers has been analysed.

Introduction

There is some literature available on methods for fault detection, classification and location of faults in power systems using machine learning algorithms. K. Chen et al. [1] have comprehensively reviewed the methods currently used for these problems. However, in most cases, to the authors’ best knowledge, the systems under consideration consist of a single transmission line with a generator on one side and a load on another. While it makes intuitive sense to consider a single line model to train a model for a single relay, the presence of other lines, buses, generators and loads in a practical system affect the electrical quantities of the line under consideration. The first part of this paper tries to study the effects of this increase in system complexity on the performance of the algorithms. The second part of this paper analyses the performance on a hardware simulator. Data of real electrical faults created on the simulator is used as a test data set while the training data set is created using a software model replica of the hardware as it has limited ports to create faults.

Conclusion

Based on our analysis and observations, we present certain conclusions and some suggestions for future work. From the analysis of the three bus system, it can be said with confidence that multi-class classifiers perform better than a combination of individual binary classifiers. The effects of other transmission lines and components(sources, loads) in the system cannot be ignored (especially near buses) while creating models and need to be analysed properly. Through the analysis of the hardware simulator, it has been observed that there may be differences in the values in a software simulation and a real system. Values in a real system are often influenced by factors(environmental, weather etc.) which are mostly ignored in a simulation. Thus, software models either need to incorporate certain real life situations or work with data from real time simulators to create truly robust systems ready to work in parallel with current protection schemes. Based on the observations, we conclude that fault type classification models for complex power systems require: • More and randomised training data (not restricted to ranges of parameters) • Analysis of effects of other lines and components in the system on the line under consideration • Comprehensive test sets to ensure better generalisation of classifier models • Testing with data from hardware simulators and real time simulators.

177 total views, 1 views today

Download PDF File

About the author: admin