Interpretation of the Machine Learning Fault Classification Model Results Using Explainable Features

Abstract

Electrical faults in the distribution network can lead to interruptions in the power supply of the customers. Therefore penalties are applied to the DSOs if they overcome the benchmark set based on all the DSOs reliability performance. Hence, the fast restoration of the power supply is crucial for the grid operator in order for the operational costs to be decreased. The traditional way of fault locating is performed with the help of so-called Fault-current Passage Indicators. This can be improved by automatic estimation of the fault location using analysis of the current and voltage signals during the fault, which also depends on the accurate fault classification. However, the existence of distortions and instabilities in some of the fault waveforms leads to an unreliable fault loop-impedance/-reactance. As a result, the location of the fault has to be performed in the traditional way which leads to a delay in the restoration of the power supply.

The objective of this thesis is to explore the potential of modern signal processing and machine learning (ML) techniques for the development of an explainable classification scheme of faults in the distribution network. ML explainability is a huge part of the research held for this thesis, as it is very important for the fault analysis department to understand the model and for the grid operators to trust the results of the model. The problem is divided into two parts. The first part concerns the construction of an explainable ML-based classification model that can accurately differentiate between types of stable faults. The second part concerns the identification of suitable criteria for single-phase and multi-phase fault stability. Combined with a fault classifier based on the first part research, this will lead to a classification and location scheme that can incorporate all fault types.

First, a literature study is done. The pre-modeling explainability stage is the most suitable one and based on that, the Short-time Fourier Transform (STFT) is chosen as the signal processing technique since it can lead to more explainable features than other techniques, e.g. the Wavelet Transform (WT). Also, the Support Vector Machines (SVMs) are selected as the supervised learning technique, as it performs well in classification problems and it is not a method as complicated as for example the Artificial Neural Network (ANN). Next, three different feature sets are shaped appropriately from the three-phase fault current and voltage signals, which are the inputs for the SVM classifier model, so then this model can be tuned and evaluated. The model with the symmetrical feature set is selected serving suitably both the performance and explainability requirements. Following that, a set of stability rules are developed to better identify the stable faults that previously were classified as unstable. These rules are tuned based on a set of initially labeled unstable faults. In the end, the ML classification and the stability model combined with section selection rules constitute the final classification scheme. This is tested on hidden data and the results indicate good performance in classifying faults, locating faults and identifying stable faults that initially would be classified as unstable. In particular, the performance of the proposed scheme on that test set has f1-score of 95.1% exceeding 93.3% of the current algorithm tested on the same data, errors regarding the fault loop-impedance/reactance are located within the acceptable limits and 4 stable faults correctly identified that previously were seen as unstable.