基于孤立森林算法的电力调度流数据异常检测方法

调度是电力系统安全运行的保障。针对具有"概念漂移"特点的调度监测流数据,基于离线数据分析或简单阈值判定的异常检测方法,存在与生产系统实时运行状态结合不紧密、依赖专家经验等问题。提出了一种基于孤立森林算法的电力调度流数据异常检测方法,利用历史数据集训练构建多个子森林异常检测器,组成基森林异常检测器;据此,在线根据滑动窗口中数据的异常情况及缓冲区数据量大小,触发检测器更新。提出一种根据异常偏差率大小筛选子森林异常检测器的更新策略,解决因模型随机更新导致异常检测器整体性能下降的问题。以服务器和某省级电网调度中心业务流数据集作为训练与测试样本,验证了所提方法在异常检测查全率及查准率等综合性能上的先进性及其在实际系统应用中的可行性。

An Approach of Data Anomaly Detection in Power Dispatching Streaming Data Based on Isolation Forest Algorithm

Dispatching is the guarantee of safe operation of power system. For scheduling monitoring flow data with "concept drift" feature, there are many issues in some aspects when using traditional offline data analysis or simple threshold judgement of anomaly detection methods, such as defective tightness between existing production systems and real-time operation state, excessive dependence on expert experience,etc. This paper puts forward an algorithm based on isolated forest data anomaly detection method of electric power dispatching flow, using historical data set to train and build many child forest anomaly detectors and compose base forest anomaly detector. Therefore, the detector update is triggered online according to abnormal data situation in sliding window and data size in cache area. An update strategy selecting sub-forest abnormal detector according to the size of abnormal deviation rate is proposed to solve dropping issue of the abnormal detector performance caused by random update.Using the flow of business data set from servers and provincial power grid dispatching centers as training and testing samples,the proposed method is verified on its advancement on comprehensive performance of anomaly detection such as recall and precision, feasible in practical system application.