Hybrid Fault Detection in Power Systems

This paper proposes a novel hybrid technique for machine-learning-based fault detection in power systems. Since typical power systems, such as the considered example of a buck converter, are complex and highly nonlinear, with parameters changing over time, it is unrealistic to derive a fully accurate mathematical model, covering every physical aspect of the system. Hence, fully deterministic model-based methods, a common approach for fault detection with buck and other DC-DC converters, suffer from modeling inaccuracies and consequential issues in precision and reliability. To overcome these problems, we propose a hybrid approach, consisting of two stages. The first stage carries out a model-based estimation of significant system parameters. However, instead of treating these parameters as deterministic ground-truth information, we treat them as a statistical quantity, with associated uncertainty. Hence, we add a second, statistical classification stage, using them as features to classify the converter condition as normal or faulty, either by a Bayesian or a deep-neural-network-based classifier. The proposed method was applied on a buck converter connected to a local distribution grid, for which we can demonstrate high accuracy in the detection of even small parameter changes.