基于大容量样本挖掘及贝叶斯堆栈泛化集成算法的电站锅炉NO_(x)稳态建模

Steady-state NO_(x) modeling using big data mining and Bayesian stacking generalization ensemble algorithm for utility boilers

氮氧化物(NO_(x))是电站锅炉燃烧的副产物,准确预测其排放对于环境保护和锅炉的燃烧优化具有重要意义。为解决NO_(x)稳态特性建模中面临的参数高维、运行数据时变性强、试验样本有限等问题,提出了一种基于大容量样本挖掘及贝叶斯集成算法的建模框架。首先,采用孤立森林、R-Value算法识别非稳态点、离群点、停滞点,获得高质量稳态样本库;然后,基于随机森林递归特征消除法进行变量选择;最后,建立贝叶斯优化堆栈泛化集成(BO-SGEM)模型对NO_(x)进行预测。以某660 MW机组锅炉72000条历史运行数据为例进行NO_(x)排放预测,结果表明:在9种离群点诊断算法中,孤立森林算法性能最优,稳态诊断和特征选择的结果与锅炉燃烧机理相符;BO-SGEM模型的精度与泛化能力均优于堆栈泛化集成学习模型及支持向量回归机、极端随机树、梯度上升树等算法模型。

Nitrogen oxides(NO_(x))are byproducts from utility boilers’combustion,and accurate prediction of their emission concentrations is of great significance to environmental protection and boiler combustion optimization.To solve the problems in NO_(x) steady-state modeling such as high dimensional parameters,strong time-varying characteristics,and limited test samples,a modeling framework based on big data mining and Bayes ensemble learning is proposed.Firstly,a high-quality steady-state databaseis obtained through recognition of non-steady points,outliers and stagnation values using isolated forest and R-Value algorithms.Then,the variable selection is carried out by using random forest-recursive feature elimination method.Finally,the Bayes operation-stacking generalization ensemble learning model(BO-SGEM)is established to predict the NO_(x) emission.Taking 72000 history operational data of a 660 MW utility boiler as an example,the NO_(x) emission is predicted and the results show that,the isolated forest algorithm has the best performance among 9 outlier detection algorithms,its results of steady-state diagnosis and feature selection are consistent with the boiler combustion mechanism.The accuracy and generalization ability of BO-SGEM are better than other base models such as stacking generalization ensemble(SGEM)model,support vector regressor(SVR)model,extremely randomized trees(ET)modeland gradient boosting decision tree(GBDT)model.