Trade-Off Between Accuracy and Computational Time for Magnetics Thermal Model Based on Artificial Neural Networks (ANN)

The thermal management of magnetic components for power electronics is crucial to ensure their reliability and for pushing the limits in terms of power density. In recent years, several studies have been carried out to obtain an Artificial Neural Network (ANN) that predicts the thermal behaviour of magnetic components removing the need to employ complex thermal models. These ANN were obtained from a dataset of multiple FEM simulations and, in order to speed up the process of building the dataset, different approaches are often used, which substantially reduce the accuracy of the ANN but it increases the speed of dataset generation. Therefore, it is an important task to know before generating the dataset which assumptions can be applied and how much error they will introduce in the ANN. This paper compare several traditionally used approaches for ANN generation and how they contributed to the error when forced convection is used, determining which approaches generate a higher error and the time saved from using those assumptions under certain boundary conditions. A new approach is also proposed for global fan modelling using only 2 variables.