A Machine Learning Perspective in an Effective Monitoring of Thermal Performance of Transformer

Abstract

As an integral part of the Smart Grid (SG), transformers’ thermal profile (Accurate Top-oil Temperature (TOT) and Hot-spot Temperature (HST)) predictions are essential for maximizing transformer utilization and deciding on the best remedial action in the case of transformer failures. However, for these predictions and estimates, the classical mathematical models of TOT lead to a mismatch between the estimated and the actual value because of assumptions, simplifications, and lack of sufficient data points. The online monitoring of transformers’ rate of ageing, capability to overload, and diagnosis are restricted by uncertainties in measurements and classical mathematical models. Therefore, a Machine Learning (ML) perspective is explored by using the Gaussian Process Regression (GPR)based TOT model to incorporate these model uncertainty and measurement noise. The transformer LoL (Loss-of-Life) and HST with uncertainties are evaluated using existing thermal (thermal-electrical-based) and GPR models.To authenticate the effectiveness of the proposed approach, MATLAB-based virtual data and data from an in-service transformer are utilized.