[关键词]
[摘要]
为建立一个有效的电站锅炉效率与NOx排放浓度预测模型,在最小二乘支持向量回归算法(Least Squares Support Vector Regression,LSSVR)基础上进行改进,提出了约束支持向量回归算法(Constraint Support Vector Regression,CSVR),通过优化支持向量的选择策略,来增强算法泛化能力和对不良数据的抵御能力。初始数据经主成分分析(Principal Component Analysis,PCA)后,输入基于CSVR算法的锅炉燃烧模型进行训练,并将建模结果与LSSVR算法和BP神经网络算法进行了比较。结果表明:使用PCA对数据预处理后,输入变量维数由五维降到三维,简化了模型结构,同时又保留了输入数据的主要特征。在相当的平均预测误差水平上,CSVR算法选用支持向量数目分别只有83个和117个,远少于LSSVR算法选用的900个;CSVR的最大预测相对误差只有3%,远低于LSSVR的25.8%,BP算法介于两者之间。
[Key word]
[Abstract]
In order to build an effective prediction model for power plant boiler efficiency and NOx emission concentration,based on the improved least squares support vector regression (LSSVR) algorithm,the constraint support vector regression (CSVR) algorithm was proposed to strengthen the algorithm generalization ability and the resistance ability to bad data by optimizing the selection strategy of support vector.After the initial data was subjected to principal component analysis (PCA),it was input into the boiler combustion model based on CSVR algorithm for training,and the modeling results were compared with the LSSVR algorithm and the BP neural network algorithm.The results show that the dimension of input variables is reduced from 5 to 3 after data preprocessing by PCA,which simplifies the model structure and retains the main characteristics of input data.At a comparable level of average prediction error,the numbers of support vectors used by the CSVR algorithm are only 83 and 117 respectively,which are both far less than the number of 900 used by the LSSVR algorithm.The maximum prediction relative error of CSVR is only 3%,which is much lower than 25.8% of LSSVR,and the BP algorithm is somewhere in between.
[中图分类号]
TM621.2
[基金项目]