基于逻辑回归的增量式异常用电行为检测方法

用户异常用电行为的检测是电力公司需要重点解决的问题。目前异常用电检测通常采用数据分析的方法,主要包括聚类和分类两种,在处理固定数据集时校测准确率和效率均较高。但是此类方法在处理增量数据时,每次数据增量更新时均需要将增量数据与原始数据合并后重新建模才能获得新的检测模型,而用户的用电数据是频繁更新的且最新的数据更能体现出用户的用电习惯,因此在异常用电行为检测时必须考虑增量数据,而现有检测方法在进行增量式异常用电行为检测时效率很低。为解决数据增量式更新的情况下异常用电行为检测方法性能低下的问题,提出了一种基于逻辑回归的增量式异常用电行为检测方法,仅需对增量数据进行建模即可得到面向全局数据集的检测模型,无需对全局数据进行重新建模,提高检测算法的执行效率。当用户电量数据产生增量时,仅需对增量数据构建检测模型,再与原始数据的检测模型相结合,即可得到基于全部数据的检测模型。实验结果表明,该方法在保证检测结果准确性的同时,极大地提高了算法执行效率,且对计算和存储资源的需求较低。

An Incremental Method of Abnormal Electricity Consumption Based on Logistic Regression

Detection of abnormal electricity consumption is an urgent problem for power supply companies. Data analysis methods, including clustering and classification, are typically adopted into abnormal electricity detection. These methods have high accuracy and performance when processing fixed data set. However, when the data set is incrementally updated, the model for the whole data set must be re-computed once the data set is updated. Meanwhile, the electricity consumption data for each user is updated every day, and the updated data is most relevant to the current behavior of the user. Thus, incremental updates must be considered when detecting abnormal electricity consumption, and existing detection methods cannot adopted because of their low performance caused by re-modeling. In order to improve the performance of the existing methods, an incremental detection method of abnormal electricity consumption based on logical regression is proposed. In this method, we only need to construct the detection model for the updated data, and the model for the whole data set can be obtained without re-modeling the whole data set, which highly improves the performance. When the electricity consumption data is updated, only the incremental data should be analyzed to get a new logistic regression model, and the regression model for the whole data can be simply computed with the new model and the original model. The experimental results show that this method not only guarantees the accuracy of detection results, but also greatly improves the efficiency of algorithm execution, and has a low demand for computing and storage resources.