Temperature rise estimation of substation connectors using data-driven models: Case: Thermal conveccion response

Abstract

A wide study regarding the suitability of data-driven modelling applied to the prediction of thermal convection responses on substation connectors is presented in this paper. The study starts with the compilation of a database with thermal profiles obtained from a finite element method simulation (FEM). Afterwards, we applied partitioning methods in order to increase the number of data sets used for modelling and later evaluate the stability of the learning algorithms. After the modeling process, the accuracy of the model per each data set is measured and the statistics about the errors are analyzed. Normality test are applied to measure the error variance. They bring us information about the error distribution and the stability of the learning algorithms. The study finish when it probes that any data-driven model is computationally less time expensive than any FEM simulation running on this study. Experimental work also confirms that the accuracy of the data-driven models: cascade feed forward neural network and feed forward neural network, can replace the FEM simulations; providing a high accuracy and a low error variance while speeding up the simulation time.