Machine-Learning Based Prediction of Generator Runtime: A predictive maintenance solution for a telecommunications site

Abstract

The telecommunications industry has experienced significant growth and trans- formation in recent years, driven by technological advancements and the increasing demand for seamless connectivity. Maintaining the infrastructure that supports them becomes a critical challenge as communication networks expand and become more complex. There is a growing trend towards designing and implementing machine learning-based predictive maintenance solutions for telecommunications infrastructure to address this challenge. This study aims to develop and implement predictive models that can accurately predict the run- time of generators on telecommunication sites. This information is crucial for efficient planning and cost estimation purposes. The research collected data over three months and split it into training and testing datasets, with the training data accounting for 70% and the testing data accounting for 30%. Three models were used for prediction: the Ordinary Least Squares (OLS) method, the log- linear method, and the Vector Autoregression (VAR) model. The models were evaluated, and their fit was calculated. In this paper, a new model was proposed, namely the OLS model with negative clipping, to address the limitations of the OLS model. The results showed that the OLS model with negative clipping pro- vided the best prediction accuracy, with a root mean square error of 0.09. On the other hand, the Vector Autoregression model exhibited the least accuracy, with a root mean square error of 9.3385. These findings highlight the effectiveness of the proposed predictive maintenance solution in accurately predicting generator runtime on telecommunication sites.