EXPLAINABLE AI FOR FAULT DETECTION IN SMART POWER TRANSMISSION NETWORKS: ADDRESSING THE BLACK-BOX PROBLEM

Abstract

An effective and reliable operation of electrical power transmission systems is imperative for modern society. It is noteworthy that the performance of artificial intelligence and machine learning models in automated diagnosis tasks in power systems has been proven to be remarkable. Nevertheless, the deep learning models currently in use, Convolutional Neural Networks and Long Short-Term Memory (LSTM) work as opaque” black boxes,” which provide no information about the process of decision making within their functioning. Thus, the major challenge associated with deploying such technologies is a severe lack of explainability. In other words, there are no ways to understand the reasoning behind AI decision in the case of faults. This paper indicates the explainability challenge as a research problem by presenting a thorough overview of Explainable AI (XAI) methodologies, namely SHAP (SHapley Additive Explanations), LIME (Local Interpretable Model-Agnostic Explanations), GradCAM, and Explainable Boosting Machine (EBM). All these methodologies are considered within the framework of fault detection problem in power systems as solutions proposed by peer-reviewed sources. The paper includes a review published between 2022 and 2026 and indicates that fault detection models utilizing Explainable AI technologies reach 99% classification accuracy while providing fully interpretable and transparent decision-making processes. The findings demonstrate that explainability and predictive performance can coexist in modern power system protection applications.