基于随机森林的变压器故障识别与实例分析

Transformer fault recognition based on random forest and case analysis

为提高变压器运行故障识别的准确性,解决现有诊断方法存在问题,提出了一种基于随机森林(RF)的变压器故障识别方法。首先,获取油中溶解的氢烃气体数据,分为训练组和测试组。以基尼不纯度(Gini impurity)为准则,利用预处理后的训练组数据构建大量决策树(decision tree),生成随机森林。通过随机森林中每棵决策树对测试数据的识别和投票,得到测试结果,并评估各特征参量重要性。最后,通过测试数据,比较改进三比值法、支持向量机(SVM)与随机森林的分类效果;并利用3种方案对某变电站运行主变实际故障运行数据进行识别。数据识别结果和实例分析表明,所采用的基于随机森林的变压器故障识别方法较改进三比值法、SVM法准确率更高,解决了现有诊断方法存在的局部最小值、过拟合等问题,且在实例中的诊断结果与变压器返厂检修结论一致,具有良好的效果。

In order to improve the accuracy of transformer operation fault identification and solve the problems of existing methods, proposes a transformer faults recognition method based on random forest(RF). Firstly, dividing the hydrogen and hydrocarbon gas data into the training group and test group that in the oil. And using Gini impurity as a criterion, a large number of decision trees are constructed with the processed training data to build a random forest. Through the identification and voting of the test data by each decision tress in the forest, the test results and the importance of feature parameters are obtained and evaluated. Finally, through the test data, compares classification effect of improved three-ratio method、the support vector machine(SVM) and the random forest, and three schemes are used to identify the actual fault data of a certain substation operation transformer. The data recognition results and the case analysis show that the transformer fault recognition method based on random forest used by has higher accuracy than improved three-ratio method and SVM, it solves the problems of existing methods such as local minimum and over fitting, and the diagnosis results in the case are consistent with the conclusion of the transformer maintenance, which is effective.