Abstract: Transverse Flux Permanent Magnet Generator (TFPMG) is a new but appears to be interesting candidate in direct-driven wind turbines for high electromotive force (EMF) density in air-gap and high efficiency at low speed. However, the output voltage of TFPMG with high quality is the crucial issue for back to back converter. To lower harmonic contents of EMF, a novel modeling and optimization method which incorporates Least Squares Support Vector Machines (LSSVM), Orthogonal Design Method (ODM), and Particle Swarm Optimization (PSO) in generator design procedure was presented. Based on the three-dimensional finite element method with scalar potential, the relationships between the machine parameters with leakages were investigated. The ODM was used to generate geometric information of trial samples. Function approximation of sample information performed by the LSSVM regression analysis provides an efficient way for parameters optimization with large-scale iterative computation. With the global search ability as well as the independence on original solution characteristics, PSO algorithm was employed to find the optimal design, which can guarantee the harmonic coefficients of non-load output voltage up to 14.36%, lower than original design. The measured result obtained from a 1.5 kW prototype test rig evaluates the effectiveness of the nonparametric modeling and the optimization method.