改进PSO-BPNN的电力变压器故障诊断与模式识别

为了优化反向传播网络相关学习参数,提出一种粒子群优化辅助BP神经网络(BPNN)的新方法.以变压器油中气体体积分数百分比构造故障特征,将BP网络的初始权值和阈值进行实数编码,以对应PSO中的粒子,实现BP网络的离线训练与在线分析,对变压器故障模式做出判断.结果表明,该算法更合理地更新了粒子的位置和速度,最优地设置了全局极值,有效克服了粒子的早熟收敛,获得的故障诊断准确率高达91%,并大大提升了BP网络的收敛速度.该算法为此类设计提供了有效的模型参考.

Fault diagnosis and pattern recognition of power transformer based on improved PSO-BPNN

In order to optimize the relative learning parameters for back propagation neural network(BPNN), a novel particle swarm optimization(PSO)aided BPNN method was proposed. The fault features were established with the gas volume fraction in the oil of power transformer. In addition, the initial weight and threshold values of BPNN were coded in real number form, and were related to the particles in the PSO. The off-line training and on-line analysis of BPNN were realized, and the fault patterns of transformer were judged. The results show that the algorithm updates the position and velocity of particles more reasonably, and the global extremum is optimally set. The premature convergence of particles is effectively overcome, and the accuracy rate of fault diagnosis approaches as high as 91%. Besides, the convergence rate of BPNN gets dramatically enhanced, and the mentioned algorithm provides an effective model reference for this sort of designs.