变压器在线监测数据异常值检测与清洗

Transformer Online Monitoring Data Abnormal Value Detection and Cleaning

在线监测数据能够实时监测变压器状态,然而经调查发现存在数据不可靠的问题,直接关系到后续状态评估的准确性。针对在线监测数据中异常值特点,以及一般的异常状态检测方法是基于阈值,噪声数据难以及时甄别的问题,提出了一种基于灰色关联度和K-means聚类的方法。利用灰色关联法对在线监测的多元时间序列数据进行关联度挖掘,提取出关联性强的序列为后续多元序列异常数据检测提供依据;其次建立基于k-means聚类的方法建立数据的异常检测模型;最后研究了时间序列预测方法,完成趋势预测并填充缺失值和噪声值,保持数据完整性。通过某变电站的在线监测数据对此算法进行验证,结果表明该方法可及时完成异常检测及清洗,清洗后准确率93.9%,完备率可达98.6%,有较高使用价值。

The online monitoring data can monitor the status of the transformer in real time. However, it is found that there is a problem of unreliable data, which is directly related to the accuracy of the subsequent status assessment. In view of the characteristics of abnormal values in online monitoring data, and the general abnormal state detection method is based on the threshold, it is difficult to distinguish the noise data in time, so a method based on gray correlation and K-means clustering is proposed. The gray correlation method is used to mine the degree of association of online multivariate time series data, and the strong correlation sequence is extracted to provide a basis for the subsequent multivariate sequence anomaly data detection. Secondly, an anomaly detection model based on k-means clustering method is established. Finally, the time series forecasting method is studied, the trend forecast is completed and the missing values and noise values are filled to maintain data integrity. The algorithm is verified by the online monitoring data of a substation. The results show that the method can complete abnormal detection and cleaning in time. The accuracy rate after cleaning is 93.9%, and the completion rate can reach 98.6%, which has high use value.