Power Outage Prediction at E.ON Using Protective Relay Data: A Nonparametric and Survival Analysis Approach

Sammanfattning

This thesis investigates the potential of using protective relay data for power outage prediction in medium-voltage distribution networks operated by E.ON Sverige AB. The study is based on retrospective relay recordings from seven substations in the Swedish distribution grid, covering the period 2018-2026. The analysis focuses on whether abnormal relay activity before outages contains systematic patterns that may support data-driven maintenance planning. The thesis consists of two main parts. First, nonparametric methods are used to analyse whether interarrival times between consecutive relay recordings decrease as an outage becomes imminent. The results do not support this hypothesis. Although short interarrival times occur before some outages, they are not sufficiently systematic to serve as a general warning indicator. Secondly, the outage prediction problem is formulated within a survival analysis framework. A Cox proportional hazards model with time-dependent covariates is fitted using characteristics derived from the relay recordings. The final model retains five statistically significant covariates and achieves a concordance of approximately 0.69, indicating moderate ability to rank lifespans according to outage risk. Overall, the results suggest that protective relay recordings contain information associated with outage occurrence, but that this information is better captured by signal characteristics than by recording frequency alone. The proposed framework should be interpreted as an exploratory proof of concept rather than a fully validated prediction tool. Further work should focus on improved outage classification, larger data sets, external validation and more comprehensive feature extraction from analogue relay signals.