Causality Prediction in a Cyber-Physical Energy System using Knowledge Graph Embeddings

With the emergence of digitalisation to increase the capabilities and efficiency of systems, Cyber-Physical Systems have emerged at the intersection of physical and computational systems. Smart grids are an example of such a Cyber-Physical System as it is comprised of a physical layer containing the producers and consumers of a system as well as a computational layer, which makes the power grid "smart". An increasing number of smaller producers in a system due to a transition to more renewable energy sources sparked the need for a more capable and complex control mechanism of a power grid. Unfortunately, it is not inherently possible to measure the source of electrical energy in power grids as they exist today. Therefore, a need for new approaches exists in order to make power grids more explainable. In recent research, Aryan et al. introduced a Knowledge Graph which can represent a Cyber-Physical Energy System. As Knowledge Graphs are able to model complex and heterogeneous data, they seem ideal to represent the relations and features of a smart grid and to find causalities between events in the system. A rule-based approach currently aims to find relevant causalities in an existing Knowledge Graph.Based on this research, this thesis applies Knowledge Graph Embeddings (KGE) on the Knowledge Graph to investigate the possibilities of using machine learning approaches for causality link prediction between events in a smart grid. Upon literature research, four KGE models were chosen to train on the Knowledge Graph - TransE, TransH, ComplEx, TTransE. In a first step, the model performance was evaluated based on their ability to represent the Knowledge Graph and the relations between various entities in the graph. In the second step, each model was evaluated on predicting causality links between two events in the system. In this process, an evaluation workshop was held where model predictions were analysed by knowledge experts in order to determine whether there may be true causality links in the predictions which are not present in the current Knowledge Graph.While no certain conclusion can be drawn on which KGE model is best suited for causality link prediction in a Cyber-Physical Energy System, this thesis provides a framework to test the application of KGEs for causality link prediction on a Knowledge Graph representing a smart grid. Additionally, the exploration of using KGEs on this use case showed that there is potential on the use of hybrid AI for improving explainability in a smart grid.