RBF网络参数优化方法及其在开关磁阻电机建模中的应用

 基于全局搜索的进化算法——粒子群算法（QPSO）和一种局部搜索算法——结构化的非线性参数优化方法（SNPOM），提出了一种混合的优化算法估计RBF神经网络中的参数——网络中心、线性参数、非线性参数，初始化一定数目的种群作为SNPOM的初始值，得到其适应值，通过选择、交叉、替换策略更新种群，完成网络中心初始值的寻优。再用SNPOM方法进一步优化，以提高SNPOM算法的全局搜索能力。仿真结果表明，混合优化方法比单独采用SNPOM法更优，且优于其他算法。并针对开关磁阻电机（SRM）高度非线性的开发重点和难点，用RBF网络进行SRM建模，将QPSO-SNPOM算法应用于RBF模型参数优化中，仿真实验结果表明，该算法较SNPOM算法精度更高、泛化能力更强，较遗传混合算法更快，训练后的RBF模型完全满足开关磁阻电机特性。

Method of RBF Network Parameters for the Optimization of Switched Reluctance Motor Modeling

Based on the evolutionary algorithm, Quantum-behaved Particle Swarm Optimization (QPSO), and the local search strategy, Structured Nonlinear Parameter Optimization Method (SNPOM), a hybrid parameter optimization algorithm (QPSO-SNPOM) for RBF neural networks is proposed. The approach starts with a population of random initial parameter values, and updates the population by selection, crossover, and replacement according to the fitness values obtained by the SNPOM. It is shown through simulation tests that the combination provides better results than either method alone(QPSO or SNPOM) and many other existing algorithms. The switched reluctance motor is a new development in motor technologies, and its characteristics, such as the small starting current, the strong starting torque, and simple structure, have made it a very attractive design. However, it has proven difficult to develop accurate SRM models because SRM exhibits highly-nonlinear characteristics. Therefore, we construct a RBF network model of the SRM, and apply our new method, QPSO-SNPOM, to optimize the parameters of the RBF network model. The simulation results show that the RBF training error is smaller, and the system is more capable of generalization when optimized by this new method rather than by SNPOM. In addition, the training time is shorter than it is when applying evolutionary algorithms. We also confirm that the trained RBF network completely models the characteristics of the SRM.