An online learning method for constructing self-update digital twin model of power transformer temperature prediction

This study was focused on predicting the temperature of power transformers, which is a critical factor affecting their reliability and efficiency. Existing methods typically use a static digital twin model for temperature prediction; however, this approach often leads to prediction failures owing to the dynamic nature of the transformer thermal process. To address this issue, an online extreme learning machine with a kernel method was proposed for constructing a digital twin model for power transformer temperature prediction. The constructed model can update itself by continuously learning the input-output relationship of new data to maintain accuracy. The experimental results show that the static digital twin model for temperature prediction gradually loses its predictive accuracy over time. In contrast, the digital twin model constructed using the proposed method had 99.8% and 98.8% prediction accuracies for two datasets. Furthermore, the proposed method learns from new samples at a speed of at least three orders of magnitude faster than existing methods for retraining the static model. Compared with the existing methods, the proposed method can effectively deal with the transformer temperature prediction under the dynamic thermal process. The results of this study can be applied to thermal management when thermal processes change dynamically.