Neural network approach to sizing an electrical machine

Purpose - The conventional starting point for the design of an electrical machine (or any low-frequency electromagnetic device) is known as "sizing". In this process, a simple magnetic circuit is used to estimate the main geometric parameters. This does not work for many devices, particularly where eddy currents and non-linearity dominate. The purpose of this paper is to investigate an approach using a neural network trained on a large database of existing designs as a general sizing system. Design/methodology/approach - The approach is based on a combination of a radial basis function neural network and a database of stored performances of electrical machines. The network is trained based on a set of typical performance requirements for a machines design problem. The resulting design is analyzed using finite elements to determine if the design performance is acceptable. Findings - The number of neurons in the network was varied to determine the approximation and generalization capabilities. The finite element analysis showed that the network produced initial design parameters which resulted in an appropriate performance. Research limitations/implications - The research has looked at only one class of machine. Further work is needed on a range of machines to determine how effective the approach can be. Practical implications - The approach can provide a good initial design and thus can reduce overall design time significantly. Originality/value - The paper proposes a novel, fast and effective generalized approach to sizing low frequency electromagnetic devices.