Deep Neural Network for Magnetic Core Loss Estimation using the MagNet Experimental Database

Magnetic components play a critical role in power electronics systems and their evolution towards higher power density and efficiency. Nevertheless, accurately modelling magnetic core losses is not a trivial task, requiring extensive measurements. In the context of the general advances of Machine Learning technologies in power electronics applications, this paper presents a Deep Neural Network (DNN) approach to core loss estimations. Various internal parameters of the DNN are tested and compared, to identify the optimal DNN structure for the core loss estimation, including the number of hidden layers, number of neurons, data transformation, and different activation functions. The training data-set comprises the MagNet database for N87 toroid magnetic cores, based on an experimental data acquisition system capable of automatically measuring various magnetic cores under arbitrary excitation signals. The results of the DNN models indicate that a DNN with suitable parameters can robustly and accurately model the core losses. The attainable accuracy is well within the required range for magnetic core losses. The optimal structure proposed in this paper consists of 10 hidden layers with sigmoid activation functions, 10 neurons in each layer, integrating a log-transformation and data normalization. The model is validated with extensive experimental tests similar to the MagNet measurement system. Furthermore, tests at higher switching frequencies up to 1MHz indicate that the model can predict losses for parameters outside the range of its training data. With the achieved performance, the DNN can benefit various power electronics engineering challenges such as loss estimation for inductor design.