Anomaly detection in power grids by means of graph convolutional neural networks

Abstract

In recent years, great successes have been achieved through the application of artificial neural networks in engineering and science. In particular, convolutional neural networks (CNNs), i.e. those based on the convolution operation, proved to be useful. The application of CNNs, however, has so far been limited to data acquired in Euclidean domains, such as pixels in a square grid or time series data with constant sampling frequency. Strict neighborhood relationships between individual data points are a requirement for the applicability of CNNs. For numerous technical and scientific applications, exactly this requirement cannot be met. An example, of a problem that is poorly handled by current machine learning technologies is anomaly detection in power grids. Graphs that can be used to model arbitrary neighborhood relationships between data points are often the most natural approach when processing such data. In the field of graph signal processing there is a generalization of the convolution operation for graph signals. Based on this, the new concept of graph convolutional neural networks developed in the last years. The applicability of graph convolution neural networks to concrete engineering problems has been insufficiently studied. This thesis presents, how an anomaly detection model based on a graph convolutional neural network can be implemented. The performance of the model for the detection of anomalies in voltage data from a power grid is examined. An adaption of the popular GCN framework is proposed, such that the problem of over-smoothing in this network is overcome.