油中溶解气体分析的变压器故障诊断新方法

New Method of Power Transformer Fault Diagnosis by Dissolved Gas-in-oil Analysis

为克服电气分析应用中BP网络算法遇到的困难,改进基本遗传算法并将其与BP算法相结合构成混合算法应用到基于油中溶解气体分析的变压器故障诊断的小波神经网络的训练中。混合算法先利用改进后的遗传算法优化小波神经网络的4个初始值,再利用梯度下降算法训练小波神经网络。针对实际情况,对所采用的遗传算法运用实数编码,采用通过自适应调整的交叉和变异概率,使遗传算法在广泛的空间搜索,向解的方向尽快收敛。仿真结果表明,该算法有效解决了小波网络若初值设置不合理,极易进入局部极小的区域使网络振荡增大、不收敛及GA算法独立训练神经网络速度缓慢等缺点。用训练过的该小波神经网络模型对456台次的变压器故障进行验证和诊断的仿真结果表明,该算法具有较快的收敛速度和较高的计算精度,证实了该算法应用于电力变压器故障诊断的有效性。

Power transformer is one of the key equipment of the power system, so it is valuable for power transformers to discover the incipient fault timely and truly. BP algorithm is used to iterate by gradient decline method of error function, so its convergence speed is slow, and easily settles into local small extremum, and the concrete extremum position is closely related to the initialization of the weights. Traditional genetic algorithm is similar to the exhaustive heuristic search, though it is a global search. That causes inevitably the searching time too long. This paper improves the simple genetic algorithm and makes up of the hybrid algorithm combining genetic algorithm with BP algorithm. It is applied to the wavelet neural network for the power transformer fault diagnosis by dissolved gas-irroil analysis. The hybrid algorithm makes use of the improved genetic algorithm to optimize the four original parameter of the wavelet neural network first, then makes use of the gradient descent algorithm to train the wavelet neural network. In view of actual situation, in the genetic algorithm, the chromosome code adopted real number and self-adaptive probabilities of the crossover and mutation is adopted in order that the genetic algorithm not only maintains the wavelet neural network characteristics, but also avoids the wavelet neural network defects. Simulation results show that the problem is solved and wavelet neural network settles into local small extremum so easily that the network surging will increase and the network will not be convergent if the initialization is unreasonable, and overcomes the shortcoming that the speed is too slow if using genetic algorithm to train neural network independently. Simulation result of 456 power transformer fault diagnosis by the trained wavelet network with hybrid algorithm indicates that the train speed and the diagnosis accuracy is improved to some extent. That shows the validity of that method in power transformer fault diagnosis by dissolved gas-in-oil analysis.