神经网络式距离保护的混合训练算法

A HYBRID LEARNING ALGORITHM FOR NEURAL NETWORK BASED DISTANCE PROTECTION

针对BP算法存在的缺陷,如训练速度慢,易收敛于局部极小点及全局搜索能力弱等,利用遗传算法能够进行全局最优化搜索这一特点,在改进的自适应遗传算法的基础上,提出了一种新的用于BP网络训练的混合算法,即改进的自适应遗传算法与BP算法相结合的混合训练方法。将所提出的混合训练方法应用于神经网络式距离保护中,利用ATP仿真计算的结果进行训练及检验,结果表明:所提出的算法与单一的BP算法相比,不仅可避免陷入局部极小点,而且提高了网络的训练速度。

For avoiding the shortcom ings existing in BPalgorithm , such as being slow in training speed, convergence to the localm inim um , and weakness in globalsearch, a new hybrid algorithm based on m odified adaptive genetic algorithm for BP netw ork is presented. Then the hybrid algorithm is applied to neuralnetw ork based distance protection. The training and testing results by ATPsim ulation show thatthe hybrid algorithm can notonly avoid convergence to the localm inim um , but also im prove the training speed for the neuralnetw ork, as com pared with the sim ple BPalgorithm . At the sam e tim e, it satisfies the dem ands ofspeed and accuracy fordistance protection. This projectissupported by Fund ofNationalEducation Com m ittee ofChina (DO96001).