IEEE TRANSACTIONS ON MAGNETICS, VOL. 43, NO. 12, DECEMBER 2007 4089 Switched Reluctance Motor Design Using Neural-Network Method With Static Finite-Element Simulation H. Sahraoui , H. Zeroug , and H. A. Toliyat Department of Electrical Engineering,National Polytechnic School, Algiers 16200, Algeria Department of Electrical Engineering, University of Sciences and Technology Houari-Boumediene, Algiers 16111, Algeria Advanced Laboratory of Electric Machines and Power Electronic,Texas A&M University, College Station, TX 77843–3128 USA The paper describes a neural network method for optimal design of a switched reluctance motor (SRM). The approach maximizes average torque while minimizing torque ripple, considering mainly the stator and rotor geometry parameters. Before optimization takes place, an experimental validation of the SRM model, based on the finite-element method, is performed. The validation predicts average torque and torque ripple characteristics for several motor configurations while stator and rotor pole arcs are varied. The numerical results are highly nonlinear, and a function approximation of the data is therefore difficult to implement. We therefore interpolate the data by using a neural network based on a generalized radial basis function. The computed results allow us to search for optimum motor parameters. The optimum design was confirmed by numerical field solutions. Index Terms—Design, finite-element method, modeling, neural-network modeling, optimization, simulation, SRM drives. I. INTRODUCTION T HE switched reluctance motor (SRM) is becoming an in- creasingly high potential drive in many applications. The capabilities of this drive include high power density, high ef- ficiency, and an inherent fault tolerance. Its simple and robust mechanical construction over a wide speed range has made this drive very attractive, and therefore it can be adopted as an al- ternative to many other ac drives. [1], [2]. There are, however, several disadvantages which are still preventing this motor from finding significant industrial applications. Acoustic noise is one major problem of the motor, particularly at low speed [3]–[5]. Therefore, it is necessary to minimize torque ripple during low speed operation. Although it offers numerous benefits as mentioned above, the design of SRM, however, is generally not straightforward. The difficulty of the SRM modeling is mainly due to the highly non- linear characteristics of the motor [3], [4]. Several design pa- rameters, such as the number of phases, pole arc, bore diam- eter, air gap, etc., should be tailored according to the require- ments of a specific application. However, it was shown in sev- eral works [5], [6] that torque output as well as the torque ripple are sensitive mainly to stator and rotor pole arcs. In [7], SRM pole arcs, and , are found to have significant importance on the performance of the motor. In fact, it was shown that the motor needs to be designed with sufficient pole overlap be- tween the stator and the rotor to ensure enough torque during phase commutation. Furthermore, widening of the pole arc of the stator makes it possible to suppress torque ripple and in- creases the value of the torque [6], [8]. This paper pays at- tention to shapes of the stator and rotor pole arcs, because of their direct effect on the inductance or torque determination. It means that these parameters should be investigated at the very first stage of the design. Optimization schemes have been car- ried out using an analytical method which takes into account Digital Object Identifier 10.1109/TMAG.2007.907990 most motor design parameters, as well as control parameters [9], [11]. Sahin [12] uses force and permeance data numeri- cally computed for a unit length geometry with identical slot- ting for the rotor and stator, to carry out a comprehensive op- timization procedure, using a back-propagation algorithm for interpolation. In his optimization (torque ripple minimization), he uses an objective function neural-network based which takes into account the excitation level as well as the motor geom- etry parameter constraints. The method shows the ripples pre- dicted from this approach agree well with those determined through the finite-element method (FEM). However, its draw- back—the search of a global optimum is determined heuristi- cally and therefore presents some complexity and lengthy proce- dures in optimum determination—can to some extent outweigh the advantages obtained through the neural-network optimiza- tion being carried out. Further, usually the motor operates in highly saturated conditions, which is a very desirable feature. Because of this and due to further simplifying associated as- sumptions, optimization analytical schemes may become inac- curate. Other optimization schemes using FEM are sought in order to account for these operating conditions, and any com- plex geometry involved [10]. Furthermore, in order to carry out geometry optimization with sufficient accuracy, numerous field solutions are required. The FEM method was combined with the heuristic method to determine the SRM shape that produces high torque consid- ering static and dynamic simulations [13]. But this approach can be time consuming and requires adjustment of the meshing at every iteration—a task which is difficult to achieve. Interpo- lation techniques are therefore sought. The interpolation using the nonlinear least square method was implemented to produce a suitable function approximation. However, this has shown limitations as the error can become significant and hence the optimization may not be accurate. In this work, a much simpler and more accurate method is pre- sented, using some design guidelines knowledge as described in [5]. The motor design parameters involved are mainly the rotor and stator pole arcs . In order to account for the geometry as well as for the nonlinearity of material utilized, 0018-9464/$25.00 © 2007 IEEE