Predictive Maintenance of Power Grids: Clustering Analysis with KMeans and Hierarchical Methods

Abstract

Maintenance observations are very important for monitoring the power grid's state, avoiding costly outages and planning preventive maintenance schedules. This thesis explores the use of unsupervised learning, in the forms of KMeans clustering and Hierarchical clustering, for labelling the maintenance records the power grid companies operate with. The objective is to group existing records in a way that will help companies to prioritise them in terms of the action that should be taken effectuating maintenance decision-making.

In this thesis, I gathered maintenance records with attributes like field evaluation of the situation, personal damage risks and the estimation of the eventual outage in the transformer circuit. The data was preprocessed and standardised before the clustering algorithms were applied.

I applied KMeans and Hierarchical clustering, experimenting with both feature reduction techniques (PCA and t-SNE) and normalisation/standardisation techniques (Standard Scaler, Robust Scaler, MinMax Scaler, and MaxAbs scaler) to obtain optimal clusters. Further, the results were evaluated with the Silhouette score and a selection of the results was manually evaluated by the domain users.

This work discusses challenges and limitations of the clustering methods in labelling maintenance records. It also discusses the ways of improving machine learning techniques and how it could be achieved in order to facilitate a more efficient maintenance strategy.

Overall, this thesis explores the use of unsupervised learning techniques for labelling maintenance records and how raw data can be transformed into a model to enhance the effectiveness of maintenance operations in power grids.