变电站在线监测多维信息聚合技术

Multidimensional Information of Aggregation Technology for On-line Monitoring of Substation

为有效突破现有变电站在线监测数据孤立,信息不能互通互联,全景多维信息难以融合等瓶颈,以变电站设备在线监测为对象,从在线监测全景多维信息利用角度出发,提出了一种基于局部相关系数与支持向量机的变电站设备在线监测物联网信息聚合模型,并将该模型运用于在线监测数据较为完善的变压器设备中.首先对变压器油色谱及油温数据进行归一化处理,其次对预处理后的数据以当前采样点的前M个点构成一组时窗长度进行局部相关系数的计算,再次利用历史数据确定支持向量机样本划分训练的界限和阀值.将相关系数作为支持向量机的输入矩阵,进行多分支支持向量机数据训练,最终根据支持向量机3次训练结果即利用相关度在0.9～1、0.8～0.9、0.6～0.8、低于0.6刻画变压器正常、异常、未预警、预警、未告警、告警6种运行状态.当其余子系统数据完善时,再利用多个分支支持向量机的训练结果结合经验权重分析,最终进行决策.该模型具有可拓展、多分支、融合程度大的优点,可实现监测信息非直接因果关系下,有效挖掘监测信息间隐性关联关系.

To effectively break through the obstacle that information of the existing isolation transformer on-line moni- toring data can not be interconnected and multi-dimensional holographic information can not be made full use of, taking on-line monitoring of substation equipment as the object, from the viewpoint of making full use of information of online monitoring panoramic multidimensional data, we proposed an aggregation model based on partial correlation coefficients and support vector machine（SVM）, and used the model to online monitor more comprehensive data of transformer equipment.First,transformer oil chromatography and oil temperature data were normalized.Then, the data after pretreat- ment with M points before the current sampling points were adopted to calculate correlation coefficients. Thirdly, historical data were used to determine the sample so as to train the support vector machine limits and thre- sholds.Finally ,the correlation coefficients were used as the input matrix of the SVM, and the multi-branch SVM was used to implement the data training. The results were based on the three training results： the use of correlation in 0.8-0.9, 0.6~0.8, 0.9-1, less than 0.6 of the normal, abnormal, no warning, warning, no alarm, alarm six running states. When the data were improved, the training results of the multi branch SVM were used for the empirical analysis and the decision making.The results show that this model has the advantages of being scalable, multi-branch;a large degree of integration can be achieved in non-monitored information, and effective mining implicit relationship between monitoring informa- tion can be realized.