Transactions of China Electrotechnical Society  2024, Vol. 39 Issue (17): 5521-5533    DOI: 10.19595/j.cnki.1000-6753.tces.231033
Current Issue| Next Issue| Archive| Adv Search |
Outlier Detection and Data Reconstruction of Dissolved Gas in Oil for Power Transformers
Jiang Jun1, Zhang Wenqian1, Li Bo2, Li Xiaohan3, Fan Lidong4
1. Jiangsu Key Laboratory of New Energy Generation and Power Conversion Nanjing University of Aeronautics and Astronautics Nanjing 211106 China;
2. State Grid Anyang Power Supply Company Anyang 455000 China;
3. State Grid Jiangsu Electric Power Co. Ltd Research Institute Nanjing 211103 China;
4. Hangzhou Qianjiang Electric Group Co. Ltd Hangzhou 311243 China

Download: PDF (2963 KB)   HTML (1 KB) 
Export: BibTeX | EndNote (RIS)      
Abstract  Transformer plays a vital role of power transmission and voltage conversion in power system, and its health is of great significance to the safe operation and reliability of the whole power grid. Dissolved gas analysis (DGA) in oil is an important indicator for transformer condition evaluation and health evaluation, especially in the process of power equipment condition maintenance and the promotion of digital power grid. Its data quality and data accumulation are of great value for transformer health index analysis, equipment management and even smart grid development. However, due to the degradation of performance, transmission interruption or other interference factor during the operation of the on-line sensors, the monitoring data occur errors and outliers, which greatly affects the use of data, and results in a waste of resources.
Considering that Copula based outlier detection (COPOD) algorithm could handle multidimensional, high and large-scale data sets, and isolation forest (IForest) algorithm had a better detection effect on dense data sets, the two algorithms were combined to mark more outliers by setting a higher detection threshold as preliminary test. To reduce the influence of both ends of data on the detection algorithm caused by the trend of data, the method of sliding window was used to detect outliers by the two algorithms respectively, and the intersection of the results was taken as the outlier data set. Due to the high threshold, the suspect outlier data set were mixed with some normal DGA data. Further, Grubbs test with high sensitivity, wide applicability and suitable for single detection was introduced, and its performance is better in small samples. Then the suspect outlier data was tested twice through Grubbs method for improving the accuracy of outlier identification, and the misidentified normal data was excluded from the detection results. After the outliers are removed, Transformer neural network is used to fill in the outliers. Mask filling and observation reconstruction were adopted to optimize the loss function and train the network structure. Finally, the predicted values using neural network were used to fill in the data to construct DGA data.
Simulation data and actual online monitoring data of 11 500 kV oil-immersed transformers in operation were used to verify the performance of the algorithm. Results show that, the number of correct identification points, the number of correct identification outlier points and the average area of the receiver operating characteristic (ROC) curve of the proposed algorithm are improved by 3.5%, 29.4% and 5.0% on the same gas series compared with the traditional K-nearest neighbors (KNN). For data filling, compared with the bidirectional scaling (BiScaler) algorithm, the mean of root mean square error (RMSE) is 7.29 μL/L, and the mean of mean absolute error (MAE) is 2.66 μL/L. The performance is improved by 9.7% and 9.2% respectively, effectively improving the quality and utilization of data.
The following conclusions can be drawn from the simulation analysis: (1) The combined algorithm of COPOD, IForest and Grubbs, which is used to identify outliers in transformer DGA online data, is more accurate than the traditional algorithm KNN, the misidentification of outliers is also reduced. (2) By optimizing the loss function using mask filling and observation reconstruction, the Transformer neural network is trained to improve the prediction performance and effectively complete the data filling task. (3) After the transformer DGA data is filled, the data is more concentrated and the quality is significantly improved without changing the overall data distribution. This method has moderate complexity and high reliability,and it is suitable for large-scale data preprocessing in practical applications.
Key wordsPower transformer      dissolved gases in oil      joint detection method      outlier detection      data reconstruction technique     
Received: 03 July 2023     
PACS: TM411  
Service
E-mail this article
Add to my bookshelf
Add to citation manager
E-mail Alert
RSS
Articles by authors
Jiang Jun
Zhang Wenqian
Li Bo
Li Xiaohan
Fan Lidong
Cite this article:   
Jiang Jun,Zhang Wenqian,Li Bo等. Outlier Detection and Data Reconstruction of Dissolved Gas in Oil for Power Transformers[J]. Transactions of China Electrotechnical Society, 2024, 39(17): 5521-5533.
URL:  
https://dgjsxb.ces-transaction.com/EN/10.19595/j.cnki.1000-6753.tces.231033     OR     https://dgjsxb.ces-transaction.com/EN/Y2024/V39/I17/5521
Copyright © Transactions of China Electrotechnical Society
Supported by: Beijing Magtech