Trustworthy machine learning in smart grids

More than a decade after its introduction, the concept of a smart grid remains essential to the industry's ongoing digital transformation. A smart grid (SG) is an electricity network that enables the bidirectional flow of electricity and data and can detect, react to, and proactively address changes in demand and a variety of other concerns, all through the use of digital communications technology. Modern SGs designed for the 21st century are required to have self-healing capabilities, which are characterized by the capacity to automatically restore and recover the interruption of energy in the grid and to shorten the interruption period for customers, thereby decreasing the likelihood of a more severe disaster, such as one caused by a cascading effect. A wide range of disciplines, including computer science, electrical engineering, signal processing, statistics, artificial intelligence, and machine learning, have been applied to the study of automatic fault prediction tasks over the past years. This dissertation focuses on the integration of machine learning-based techniques to improve the self-healing capabilities of SGs and examines these ML approaches not only from the standpoint of fault prediction accuracy but also their trustworthiness. Among the numerous facets of trust, this study focuses on the robustness (against faults and adversarial attacks) and interpretability of the proposed fault prediction systems. This doctoral research project was assigned by the e-distribution Smart Grid Lab in Milan, where the objective of the project was to develop Artificial Intelligence (AI) algorithms with the goal of enabling automatic self-healing characteristics for next-generation smart grids. Self-healing can be used in a distribution network, e.g., in the smart grid, to detect a fault, localize the fault and diagnose the fault type, to isolate and neutralize it. It should be noted that we followed a Human-Centred Design approach. initially visiting the e-distribution Smart Grid Lab in Milan to interview electrical engineers and study their work, systems, and artifacts. Then, due to the limits imposed by the COVID-19 epidemic, we conducted monthly video sessions with the Lab team in order to discuss the preliminary research findings. In the context of this doctoral dissertation, methods for predicting faults and identifying them by type and origin have been devised, implemented, and evaluated. In order to extract useful information from the electrical signal and incorporate it into a machine-learning fault prediction system, a number of novel techniques have been proposed in addition to existing ones being improved. These techniques include handcrafted temporal, frequency, and wavelet features, as well as 2D CNN-based visual spectrogram methods. We also examine the explainability of the various integrated technologies, including the use of visual explanation, in order to make the systems more transparent to a wider audience (operators, consumers). Furthermore, this work enlightens a crucial research area in the security of smart grids, namely what happens to fault prediction methods when they are targeted by malicious actors or adversarial attacks. It is demonstrated that state-of-the-art adversarial techniques like FGSM and BIM are capable of learning minor perturbations that can trick the ML models, for example, by misclassifying the fault type or location, hence prolonging or impeding the recovery time of the rescue team.