Development of a machine learning based predictive maintenance algorithm for transformers in distribution substations.

Abstract

With inadequate maintenance being the third most cause of transformer failures in substations, adoption of suitable maintenance strategies for plant equipment has become an essential decision. The existing maintenance strategies currently employed in UMEME are not adequate to keep assets operational at all times. This affects reliability of power supply to consumers, and also results into loss of revenue. It has therefore become vital to introduce the concept of predictive maintenance which allows for safety compliance, pre-emptive actions and increased asset life. This helps in knowing what kind of faults are likely to occur hence prior investigations, maintenance scheduling adjustments and repairs are done before an asset fails. This report therefore presents the development of a machine learning based predictive maintenance algorithm that can forecast likely faults to happen in a transformer basing on faults that happened in the previous years. Information acquired from the maintenance department of UMEME and online research papers was used to generate datasets that were later used to train the LSTM model that was selected. The results show that the likely faults are forecasted up to one year (2021) with a high accuracy of 84%, MSE of 2 and RMSE OF 1.414.