The Research of Anomaly Detection Method for Transformer Oil Temperature Based on Hybrid Model of Non-Supervised Learning and Decision Forests

2018 2nd International Conference on Power and Energy Engineering

Xiao, Fei^1 ; Leng, Xiwu^2 ; Ye, Kang^1 ; Hu, Youlin^1 ; Li, Xiongli^3 ; Zhu, Licheng^3

State Grid Shanghai Electric Power Company Power Dispatch Control Centre, Shanghai, China^1

National Power Dispatch Communication Centre, Beijing, China^2

Tellhow Software Co. Ltd., Nanchang, China^3

关键词: Anomaly detection;    Anomaly detection methods;    Competitive advantage;    Decision forest;    Hypothesis tests;    Main transformer;    Non-supervised clustering;    Temperature threshold value;   

Others  :  https://iopscience.iop.org/article/10.1088/1755-1315/192/1/012020/pdf
DOI  :  10.1088/1755-1315/192/1/012020

来源: IOP

【 摘 要 】

The anomaly detection of transformer's oil temperature is critical and valuable issue for the safe operation of transformers and power system. In terms of the defects of traditional anomaly detection approaches of transformer's oil temperature, such as high investment, poor generality, and non-real time, this paper proposed a hybrid model with non-supervised learning and decision forests method to detect anomaly of transformer's oil temperature. Based on non-supervised clustering algorithm, firstly, the clusters of transformers' working conditions are explored from big data sets of transformers. After that, the abnormal temperature threshold value of each cluster is deduced by hypothesis tests method and utilizes to tag anomaly in data sets of working conditions. Finally, the data sets with anomaly tags are inputted into random decision forests model to construct the classifier of abnormal oil temperature and generate the rules for anomaly detection. This method was validated by empirical data of main transformer in Shanghai, and the results represented its conspicuous competitive advantages to traditional oil temperature anomaly detection methods in the factors of real-time and accuracy.