A Neuro-Symbolic Approach for Fault Diagnosis in Smart Power Grids

Abstract

Power quality is a critical parameter of modern power electrical systems, the complexity and decentralization of which are rapidly increasing. Indeed, the highest possible quality is a requirement of all the stakeholders of a power grid. In response to this demand, we introduce, in this article, a novel neuro-symbolic approach for the diagnosis (i.e., detection and classification) of the typical faults that a smart power grid encounters during its operation (that is, voltage interruptions, voltage sags, voltage swells, transients and harmonics). Heart of the implemented system is an Artificial Neural Network (ANN) that identifies with high fidelity the patterns of voltage-waveforms — for the sake of comparison, two ANNs were evaluated, namely, a conventional Multilayer Perceptron (MLP) and a one-dimensional Convolutional Neural Network (CNN). The output of the ANN is passed through a symbolic reasoner, implemented by means of Answer Set Programming (ASP), which provides a final response on the condition of the power grid, taking into account the background knowledge of the domain, which is in turn encoded into appropriate symbolic rules. The proposed approach achieved very high classification-performance on the validation dataset ( the MLP and the CNN), and, thus, it constitutes a promising powerful tool that will contribute to the improved quality of future power grids.