摘要
电站锅炉NO_x排放是大气污染物的重要来源,建立有效的预测模型是降低NO_x排放的基础。NO_x的排放特性受多个热工变量的影响,针对变量间的相关性和强耦合性,提出一种基于互信息变量选择和长短期记忆神经网络的预测模型,实现对NO_x排放的动态预测。以互信息"最小冗余最大相关"为准则对特征变量进行重要性排序和变量选择。在变量筛选过程中采用序列前向选择方法,以模型预测精度为目标确定最优输入特征集和最佳模型参数。将筛选出来的特征变量集作为LSTM预测模型的输入,并采用多层网格搜索算法优化网络超参数,建立了NO_x排放动态预测模型。基于某660 MW超超临界燃煤机组的运行数据对模型进行验证,实验结果表明该方法能够有效地减少模型输入变量的数目,降低变量间的信息冗余,同时提高了预测模型的精度和鲁棒性。
The NO x emission during coal combustion is one of the major environmental pollutants,which makes it necessary to establish an effective NO x prediction model.The emission characteristics of NO x are influenced by many thermal variables.Considering the correlation and strong coupling of variables,this paper proposed a prediction model based on mutual information variable selection and long-short term memory(LSTM)neural network to realize the dynamic prediction of NO x emission.We carried out importance ranking and selection on characteristic variable based on the principle of“minimum redundancy and maximum correlation”in mutual information.In the process of variable selection,we adopted the sequence forward selection method to determine the optimal input set and model parameters to improve model prediction accuracy.We established the dynamic prediction model of NO x emission by taking the selected variables as the input of LSTM neural network and applying multi-layer grid search in optimizing model parameters.We validated the model by the operating data of a 660 MW coal-fired boiler.The experimental results have showed that the proposed method effectively decreases the number of input variables,reduces the information redundancy among input variables and improve the prediction accuracy and robustness of the model.
作者
杨国田
王英男
李新利
刘凯
YANG Guotian;WANG Yingnan;LI Xinli;LIU Kai(School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China)
出处
《华北电力大学学报（自然科学版）》
CAS
北大核心
2020年第3期66-74,共9页
Journal of North China Electric Power University：Natural Science Edition
基金
中央高校基本科研业务费专项资金资助项目(2018QN052).
关键词
NOX排放
动态预测
互信息
长短期记忆
深度学习
NOx emission
dynamic prediction
mutual information
long-short term memory(LSTM)
deep learning
