Power Transformers Thermal Modeling Based on the Modified Set-Membership Evolving Multivariable Gaussian and Variable Step-Size Evolving Multivariable Gaussian

Knowledge of temperature distribution in power transformers is essential for the management of electrical distribution systems. Monitoring the hot-spot temperature of a power transformer can extend its lifetime. This paper introduces two novel models called Modified Set-Membership evolving multivariable Gaussian (MSM-eMG) and variable step-size evolving multivariable Gaussian (VS-eMG) for time series forecasting. Both approaches are an enhanced version of the evolving multivariable Gaussian model that use adaptive filtering to update the learning rate parameter, which updates the centers of the clusters, aiming to achieve better performance of the models. To evaluate their performance were used two data sets from a real power transformer; the first data set of the transformer has no overload conditions, and the second one has it. A synthetic data set was also used, as a benchmark, in order to show the effectiveness of these models in different scenarios. The obtained results are compared with the performance of the original evolving multivariable Gaussian and with other classical evolving and non-evolving models suggested in the literature. Both proposed models obtained the lowest errors in all simulations and presented a competitive number of rules in the real data, suggesting these models are flexible and efficient approaches to forecast complex data with high accuracy.