基于聚类分析和用户画像的用气负荷异常检测

Anomaly Detection of Gas Load Based on Clustering Analysis and User Profile

将燃气日瞬时流量和日用气量作为研究对象,提出基于K-means聚类、特征标签、用户画像、k折交叉验证和岭回归的用气负荷异常检测方法。结合实例,对该异常检测方法进行探讨。将案例用户某段时间的瞬时流量组成数据集,使用K-means算法进行聚类分析,将用气分为工艺生产和停工小火两类用气行为,得到工艺生产数据集。针对工艺生产数据集中的每个样本,得到6个特征标签(日最大负荷、日均负荷、日用气时段百分比、日用气量、用气负荷相似度、用气负荷冲击度)。将特征标签归一化后绘制修正箱线图,即用户画像,剔除了异常样本。使用k折交叉验证和岭回归算法构建异常评价标准。利用岭回归算法构建异常评价模型。将案例用户另一段时间的瞬时流量输入异常评价模型,进行负荷异常检测,与实际结果对照,得到该异常检测方法的准确率达到90%以上。

The daily instantaneous gas flow and daily gas consumption are used as the research objects,and a gas load anomaly detection method based on K-means clustering,feature label,user profile,k-fold cross-validation and ridge regression is proposed.This anomaly detection method is discussed combined with an example.The instantaneous flow rate of the case user for a certain period of time is composed into a data set,and the K-means algorithm is used for clustering analysis to classify the gas consumption into two categories of gas consumption behaviours:process production and pilot at stoppage,and the process production data set is obtained.For each sample in the process production data set,the six feature labels are obtained(maximum daily load,average daily load,percentage of daily gas consumption period,daily gas consumption,gas load similarity and gas load impact).After normalizing the feature labels,a modified box line diagram,i.e.the user profile,is drawn,and abnormal samples are eliminated.Anomaly evaluation criteria are constructed using k-fold cross-validation and a ridge regression algorithm.Anomaly evaluation model is constructed using a ridge regression algorithm.The instantaneous flow rate of the case user for another period of time is inputted into the anomaly evaluation model for load anomaly detection.Compared with the actual results,the accuracy of the anomaly detection method is more than 90%.