Explainable Artificial Intelligence Methods for Intelligent Fault Detection in Inverter-Based Distribution Systems

The use of artificial intelligence (AI) and machine learning (ML) for intelligent fault detection in electric distribution systems (DS) has recently gained significant interest owing to the impact of the ongoing integration of distributed generation (DG) resources including inverters. While these approaches demonstrate superior fault detection accuracy over traditional means, they are dominated by the use of complex, black-box models such as shallow and deep neural networks. The nature of these models limits the ability of protection engineers to understand the underlying mechanism behind the fault detection process and hinders the practical adoption of these models in the DS protection schemes. This research aims to close this gap and proposes model and data-based explainability via explainable AI (XAI) to enhance trust and understanding of the models used for fault detection in inverter-based distribution systems. Artificial neural networks (ANNs) are developed and trained to detect the presence of a fault and its location in a simulated distribution system with inverter-based generation. The model explainable AI method demonstrated the ability of Shapley values to measure the relative feature importance for the input features of the fault type and location ANN classifiers. The novel data-based explainable AI method developed showed the ability of decision trees to provide data-driven explanations behind each of the predictions for each ANN classifier. Holistically, this work highlights the benefits and limitations of two distinct XAI techniques that will enable a greater transparency and understanding for the next generation of AI driven DS fault detection schemes.