Neural Network Models for Predicting Magnetization Surface Switched Reluctance Motor: Classical, Radial Basis Function, and Physics-Informed Techniques

Neural networks have increasingly been utilized in electric drive systems to enhance modeling, control, and optimization. These data-driven techniques enable accurate predictions of complex nonlinear behaviors, including the magnetization characteristics of electric machines. This paper investigates the use of neural networks for predicting magnetization surfaces in switched reluctance motors, a key aspect of their design and operational efficiency. Three neural networks-based methods are studied: classical neural networks, radial basis function neural networks, and physics-informed neural networks. Experimental data from a 7.5 kW switched reluctance motor are used to assess the capabilities of each approach. Moreover, the study evaluates predictive accuracy, computational requirements, and the ability to reflect physical dynamics. Results demonstrate that classical neural networks and radial basis function networks can model magnetization surface, but with inaccuracy due to failure to comply with flux behavior, with radial basis function networks excelling in computational efficiency. Physics-informed neural networks achieve the highest accuracy by integrating physical laws into the learning process. This research highlights the potential of neural networks techniques in advancing switched reluctance motors modeling, paving the way for improved electric drive systems.