Digital Twin Modelling of Transformer Aging Using Federated Learning

Abstract

Thermal degradation of cellulose insulation is responsible for about 40% of power transformer failures world-wide and aging is one of the most important challenges facing the reliability of the electrical transmission grid. In this paper, a novel framework based on digital twin (DT) technology and federated learning (FL) is proposed to realize a decentralized monitoring and prediction system for the transformer aging process without compromising on privacy in geographically distributed substations. The proposed architecture builds a high fidelity virtual representation of physical transformer assets, continually refreshed via secure federated aggregation of models, without needing to consolidate raw data. The framework is based on the FedAvg algorithm as the basic aggregation mechanism, and the accuracy of the model predicted the health of the transformer on the transformer health indices, with an accuracy rate of 93.5%, improving by 5.1% compared with traditional centralized baselines, and the communication overhead reduced by 37.5%. Experimental validation is performed using the data obtained from dissolved gas analysis (DGA) of the four simulated substations, covering a total operating time of 3,730 hours. The Arrhenius thermal aging model is built into the digital twin layer, allowing for real-time estimation of hotspot temperature and prediction of remaining life. The proposed DT-FL framework features the best scalability with up to 100 clients and round convergence of 75 rounds for communication. The quantitative results also show that the insulation degree of polymerization (DP) predictions have a root-mean-square error (RMSE) of 18.3 DP units, which confirms the applicability of the framework for use in next generation smart grid asset management.