Enabling Predictive Maintenance using Semi-supervised Learning with Reg-D Transformer Data

Abstract

Predictive maintenance has now become a possible avenue within the electricity distribution sector of Eskom. A recent roll-out of large-scale substation data acquisition projects have allowed this sector to use a predictive based maintenance scheduling plan instead of a previously used frequency based maintenance plan. This paper describes the design and implementation of a low-complexity anomaly detection algorithm, which is able to detect discrepancies, indicative of small electrical grid changes or substation equipment deterioration. The algorithm is based on projecting parametric multivariate Gaussian functions on to spatially distributed pre-selected substation data points. This method enables the utility to monitor critical variables, and their relationships, in an effort to foresee equipment or network distresses from high-value assets, particularly in the transmission and distribution sector. The results demonstrate an early positive detection of anomalous load behaviour from a live substation. The presented Semi-supervised learning methodology can form the underpinnings of an integrated approach, to aid with operational decision-making, and seems eminently suitable to reduce unscheduled asset downtime.