Power Transformers Thermal Modeling using an Enhanced Set-Membership Multivariable Gaussian Evolving Fuzzy System

Abstract

In this work we present two new models based on Evolving Multivariable Gaussian approach and on the Set-Membership/Enhanced Set-Membership adaptive filtering framework to model the thermal behavior of power transformers. In these new models, adaptive filtering approaches work to adjust the learning rates of the evolutionary model, while the use of multivariable Gaussian membership functions, instead of single-variable as evolving models in general, makes information about interactions between input variables preserved and used in the training process. In addition, the evolving structure of the proposed models make these models more adaptable to changes in the operational conditions of power transformers (like insulation aging, environmental changes, load profile changes, among others) than its non-evolving counterparts. To evaluate the performance of the proposed models they were applied to two benchmark problems and to the thermal modeling of real power transformers problem under two load conditions: with and without an overload condition. The obtained results are compared with the performance of the original evolving Multivariable Gaussian and with other classical models suggested in the literature. Both proposed models obtained significantly higher performances than all the other tested models, suggesting that the models are flexible and efficient approaches in these scenarios representing a promising approach in the modeling of power transformers, especially for real-time applications.