Relative fault vulnerability prediction for energy distribution networks

Aging infrastructure in energy distribution networks, e.g., electrical distribution systems and district heating networks, increases components’ fault vulnerability. This decreases the security of supply and infers large costs on consumers and utilities alike. In this paper, we explore the application of geospatial data processing and machine learning as a means of grasping the complex relationship between pipes’ and cables’ physical working environment and their relative fault vulnerability. The paper presents the application of relative fault vulnerability prediction on a district heating network and an electrical distribution network. A large part of this work revolves around the treatment of data. In this regard, we demonstrate a method to combine geospatial data from different sources, namely environmental data, georeferenced fault observations, and the GIS of the case networks, and how to further augment the data with spatial fault clustering and expert knowledge, injected in the form of rule-based data reconstruction. Additionally, we propose a new data level imbalanced learning technique to handle the scarcity of fault observations. Our results confirm, that our method outperforms traditional data level imbalanced learning techniques i.e., methods that change the data as opposed to changing the model or how it is trained. Our results also show that our model for relative fault vulnerability prediction effectively identifies failure-prone pipes and cables and outperforms age-based vulnerability ranking, which is a current industry practice.