基于偏互信息的变量选择方法及其在火电厂SCR系统建模中的应用  

Variable Selection Method Based on Partial Mutual Information and Its Application in Power Plant SCR System Modeling

作　　者：刘吉臻 秦天牧 杨婷婷 吕游 

机构地区：[1]新能源电力系统国家重点实验室(华北电力大学),北京市昌平区102206

出　　处：《中国电机工程学报》2016年第9期2438-2443,共6页Proceedings of the CSEE

基　　金：国家重点基础研究发展计划(973计划)资助(2012CB215203);北京高等学校青年英才计划项目(YEPT0705);中央高校基本科研业务费专项资金资助(2015XS69)~~

摘　　要：数据驱动模型被广泛应用于工业过程中,最优变量集的选取对模型性能非常重要。针对工业过程中建模对象普遍具有的强非线性以及变量间的相关和耦合特性,采用偏互信息方法(partial mutual information,PMI)进行变量选择。利用benchmark验证了PMI方法的有效性并将其应用于火电厂SCR烟气脱硝系统。将选取的最优变量集作为支持向量机(support vector machine,SVM)的输入,建立了PMI-SVM模型。此外,将PMI方法与人工神经网络方法(artificial neural network,ANN)结合,构成PMI-ANN模型。将两种PMI模型与原始的SVM和ANN模型进行对比,结果表明PMI方法降低了模型的复杂度,提高了模型的学习和泛化能力。Data-driven model is widely used in industrial process. However, selecting an appropriate set of input variables is important for obtaining high-quality models. In view of the industrial process modeling objects generally have strong nonlinear and the variables are relevant and coupled, this paper adopted partial mutual information （PMI） method for variable selection. Benchmark was used to validate the effectiveness of the PMI method and PMI was employed to the variable selection of power plant SCR denitration system. PMI-SVM model was established based on the selected optimal variable set as model input. In addition, combining PMI method with artificial neural network （ANN）, the PMI-ANN model was constituted. Comparing two PMI models with the SVM and ANN models without the variable selection, the results show that the PMI method reduces the complexity of the model, the learning and generalization ability of the models are both improved.

关 键 词：偏互信息 变量选择 支持向量机 烟气脱硝 数据建模 

分 类 号：TK39[动力工程及工程热物理—热能工程]