可解释的变负荷下燃煤机组SCR反应器入口NO_(x)质量浓度预测模型  

Interpretable prediction model for NO_(x) mass concentration at SCR reactor inlet in coal-fired power plants under flexible operating conditions

作　　者：李影 卓建坤[2] 吴逸凡 樊永刚 姚强 李水清[2] LI Ying;ZHUO Jiankun;WU Yifan;FAN Yonggang;YAO Qiang;LI Shuiqing(School of Electrical Engineering,Xinjiang University,Urumqi 830046,China;Key Laboratory for Thermal Science and Power Engineering of Ministry of Education,Tsinghua University,Beijing 100084,China)

机构地区：[1]新疆大学电气工程学院,新疆乌鲁木齐830046 [2]清华大学热科学与动力工程教育部重点实验室,北京100084

出　　处：《热力发电》2024年第7期119-128,共10页Thermal Power Generation

基　　金：自治区重大科技专项项目(2023A01005-1)。

摘　　要：燃煤电厂灵活调峰过程NO_(x)测量往往存在滞后现象,导致选择性催化还原(selective satalytic reduction,SCR)脱硝喷氨控制系统响应不及时,易造成喷氨量过高或过低,从而造成SCR反应器出口NO_(x)质量浓度波动剧烈和氨逃逸率增大。为实现喷氨阀门的提前快速调节并考虑影响燃煤锅炉NO_(x)排放量的因素存在耦合性,提出了一种基于卷积神经网络(convolutional neural network,CNN)和长短时记忆(long short-term memory,LSTM)神经网络混合模型的SCR反应器入口NO_(x)预测模型。利用一台330 MW燃煤电站锅炉的运行参数,通过Pearson系数法计算特征变量之间的相关性,筛选出相关性较大的特征,并定义模型的输入矩阵和输出矩阵,采用随机搜索算法进行优化,以提高预测性能。进一步利用SHAP算法对黑箱模型进行解释,并通过Simulink仿真验证了带有NO_(x)预测的控制效果。结果表明:CNN-LSTM预测模型在调峰负荷变化时,能够以较高的精度预测SCR反应器入口NO_(x)质量浓度的变化,并能提前25s为喷氨控制系统提供反馈;优化后的喷氨控制策略降低了出口NO_(x)质量浓度与设定值间的标准差(降低28%),并提升了NH_(3)/NO_(x)的响应速度,减小最大氨逃逸量22%。该研究结果可为灵活调峰机组的智慧SCR脱硝技术及燃烧优化提出有效的指导。There is a delay in NO_(x) measurement for flexible operations in coal-fired power plants,which leads to a delayed response in ammonia injection control system of selective catalytic reduction(SCR)reactor,resulting in potential over or under-injection of ammonia and significant fluctuations in NO_(x) mass concentration at outlet of the SCR reactor.To enable proactive adjustment of ammonia injection and considering the interconnected factors influencing the NO_(x) emissions from coal combustion,a prediction model for NO_(x) mass concentration at the SCR reactor inlet is proposed based on convolutional neural networks(CNNs)and long short-term memory neural(LSTM)networks.By using operational parameters from a 330 MW coal-fired power plant,a Pearson coefficient method is employed to calculate the correlation between feature variables.Significant features are extracted to define the model input matrix and output matrix.The random search algorithm is used for hyper-parameters optimization to enhance predictive performance.The SHAP algorithm is then applied to interpret the model structure and explain the black-box model.Finally,the control effects of model with NO_(x) concentration prediction is verified through Simulink simulation.The results indicate that,the CNN-LSTM prediction model demonstrates higher predictive accuracy for the variable NO_(x) mass concentration at the SCR reactor inlet during the frequent load fluctuations.It can provide feedback to the ammonia injection control system of 25 seconds in advance.The optimized ammonia injection control strategy not only reduces the standard deviation between the NO_(x) mass concentration at the SCR reactor outlet and the set value by 28%,but also improves the response speed of NH_(3)/NO_(x) regulation,reducing the maximum ammonia slip by 22%.The research findings can provide guidance for intelligent SCR denitration system and combustion optimizing operating during flexible operation of coal-fired power plants.

关 键 词：NO_(x)预测 燃煤机组 CNN-LSTM模型 SHAP 灵活调峰 

分 类 号：X773[环境科学与工程—环境工程]