基于随机森林特征优选与MAEPSO-ELM算法的变压器DGA故障诊断

Transformer DGA fault diagnosis based on the random forest feature optimization and MAEPSO-ELM algorithm

针对变压器故障智能诊断方法中输入特征不同影响诊断结果以及粒子群算法优化极限学习机准确率低的问题,提出基于随机森林特征优选与多尺度协同变异粒子群极限学习机的变压器DGA故障诊断方法。首先,基于故障样本DGA数据建立候选特征集,采用随机森林算法计算各特征重要性评分并降序排列,通过序列前向选择法筛选最优输入特征;其次,针对极限学习机参数选择困难的问题,引入多尺度协同变异粒子群算法进行优化;最后,将之与IEC三比值法以及不同组合极限学习机诊断性能进行比较。实例表明所提方法诊断精度更高。

A transformer DGA fault diagnosis method is proposed based on the random forest feature optimization and multi-scale cooperative mutation particle swarm limit learning machine for the problems that different input characteristics effects the diagnosis results and the low accuracy of particle swarm algorithm optimization limit learning machine.Firstly,the candidate feature set is established on the basis of the DGA data in the fault sample.The random forest algorithm is utilized to calculate the feature importance scores and rank them in a descending order.The optimal input features are then selected by the sequence forward selection method.Next,aiming at the problem of difficult parameter selection of extreme learning machine,a multi-scale cooperative mutation particle swarm optimization algorithm is introduced for optimization.Finally,the method is compared for the diagnostic performance with the IEC three-ratio method and different combinations of extreme learning machines.An example shows that the proposed method has higher diagnostic accuracy.