Abstract: Boiler combustion process modeling is critical for the accuracy and reliability of combustion optimization.Firstly,a steady-state detection (SSD) algorithm is used to extract steady-state samples for building steady-state combustion process model.Considering the unbalance of samples over power load,a new method of data division,which divides the available data into training subset and test subset according to load,is proposed to improve model generalization.Then,a single factor graph analysis is conducted to determine searching range of three SVM model structural parameters.After choosing the model parameters by combining grid search with cross validation,four multiple inputs single output SVM models,including boiler efficiency,NO_x emission,flue gas temperature and the unburned carbon in fly ash,are established based on the divided data.The results are demonstrated that four models all have a good generalization capability.

Key words: combustion process modeling, support vectors regression, steady-state detection, parameter selection

摘要： 为建立燃烧过程稳态模型，首先利用稳态检测算法提取稳态样本；针对稳态数据中的不均衡性，提出了一种基于负荷划分数据的方法，即根据负荷工况将样本划分成训练子集与测试子集，以提高模型的泛化性能。利用单因素图形分析方法确定3个模型参数的搜索范围，将网格搜索与交叉验证相结合选择最优的模型参数，在此基础上建立了一个300 MW燃煤火电厂机组锅炉燃烧过程的支持向量机模型，包括锅炉效率、NO_x排放量、排烟温度和飞灰含碳量4个过程输出。结果表明，经过参数优化的4个输出模型均具有很好的泛化性能。

关键词: 燃烧过程建模, 支持向量回归, 稳态检测, 参数选择