变量选择与支持向量机相结合的SO2排放特性建模

摘要

Coal-fired power plant boiler is one of the major sources of SO2 pollution emission.In circulating fluidized bed (CFB) boiler,the unstable desulfurization in the furnace leads to that the desulfurization of desulfurization tower is not timely,ultimately resulting in excessive SO2 emission.Establishing an effective SO2 emission prediction model can effectively solve this problem.The SO2 emission characteristics are influenced by many thermal parameters which are correlation and coupling.In this regard,this paper presents a SO2 emission prediction model based on variable selection and SVM.Based on the field operation data of a 300 MW CFB boiler,the BP neural network was applied to reduce the dimension and the complexity of the input variables,and the filtered variables were used as the input of BP-SVM model.And then K-fold cross validation method was used to determine the optimal parameters of the model by grid search to establish the SO2 emissions BP-SVM model.Compared with the SVM model without variable selection,the results show that the BP-SVM model after the variable selection can effectively reduce the model complexity and improve the generalization ability.%燃煤电站锅炉SO2排放是大气污染的主要来源之一,建立有效的SO2排放预测模型有利于解决循环流化床(CFB)锅炉因炉内脱硫不稳定导致脱硫塔脱硫不及时而引起的SO2排放超标的问题.SO2的排放特性受众多热工参数影响,且各参数间存在相关性与耦合性,对此本文提出一种基于变量选择与支持向量机(SVM)的SO2排放预测模型.基于某300 MW CFB锅炉现场运行数据,采用BP神经网络降低输入变量的维度与复杂度,将筛选后的输入变量作为BP-SVM模型的输入,采用K-折交叉验证的方法通过网格搜索确定最优模型参数,建立SO2排放BP-SVM模型.将BP-SVM模型与未经变量选择的SVM模型对比分析,结果表明经过变量选择后的BP-SVM模型可以有效降低模型复杂度,提高模型泛化能力.