【作者】 王丽平；

【导师】 吕游；

【作者基本信息】 华北电力大学(北京) ， 控制工程， 2022， 硕士

【摘要】 我国煤炭资源丰富,是一个以煤炭为主的能源消费型大国,火力发电在全国总发电量中占有较大的比例,火电厂燃煤锅炉燃烧煤炭所产生的氮氧化物(NOx)是主要的污染源之一,已引起社会的广泛关注。国家对火电厂气体排放的控制也做出了相应的要求,降低NOx的排放浓度迫在眉睫,目前在实际的火电厂中,大多数通过选择性催化还原(SCR)技术来控制NOx的排放,而对于燃烧过程中的NOx浓度,通常电厂会使用烟气自动监控系统(CEMS)来进行NOx的实时跟踪测量,但在实际电厂中,测量结果与真实值之间存在一定的迟延时间,此外,由于新能源所具有的特性,当其机组大规模并网时,火电机组需要通过快速变负荷来增加新能源利用率,会导致NOx浓度发生显著变化,不能实时、准确的反映SCR入口 NOx浓度的变化,进而不能及时有效的指导SCR反应器做出相应动作调整。为解决上述问题,本研究以某660MW燃煤机组为研究对象,设计了一种考虑时延的SCR脱硝系统入口 NOx浓度自适应预测模型,主要研究内容如下:(1)燃烧系统中每个采样时刻测得的锅炉侧数据与SCR系统入口 NOx浓度数据均存在迟延时间,直接通过采集到的数据建立SCR脱硝系统入口 NOx浓度预测模型不能精准地将锅炉侧参数和NOx之间的非线性关系表达出来,为了解决这一问题,本文基于互信息法估计了 SCR脱硝系统入口 NOx生成和锅炉侧参数之间的迟延时间,考虑到各个变量之间的耦合性,使得模型的输入数据更符合实际工程,提高了预测模型的准确性。(2)本文建立了基于LSTM的SCR脱硝系统入口 NOx浓度预测模型,LSTM神经网络凭借其独特的门控设计解决了序列的长期依赖问题,单一的模型容易受到各种因素的影响从而影响模型的训练精度,引入卷积神经网络,利用其特征提取的优势,建立CNN-LSTM预测模型来提高模型的稳预测精度,利用相同的电厂实际数据集分别对两种模型进行对比分析,实验表明,后者的模型预测精度更高。(3)部分系统的某些特性具有时变性,原有的离线模型的预测精度不能满足实际需求,为解决这一问题,采用贝叶斯优化算法自动调整优化模型的超参数,利用欧氏距离对输入数据进行重构,结合CNN-LSTM神经网络预测模型实现模型的在线更新,通过实际数据对模型性能进行验证,该方法可以在系统特性发生改变时,仍然有较好的预测精度和训练速度,一定程度上保障模型的有效性,在火电厂中有一定的实际应用价值。更多还原

【Abstract】 China is rich in coal resources,is a coal-based energy consumption of large countries,thermal power generation in the country’s total power generation accounted for a large proportion,coal-fired boilers in coal-fired boilers to produce nitrogen oxides(NOx)is one of the main sources of pollution,has caused widespread concern in society.The state has also made corresponding requirements for the control of gas emissions from thermal power plants,and it is urgent to reduce the emission concentration of NOx,at present,in the actual thermal power plants,most of them control the emissions of NOx through selective catalyst reduction(SCR)technology,and the NOx concentration during combustion,usually the power plant will use a continuous emission monitoring system(CEMS)to carry out real-time tracking measurement of NOx,but due to the location of the measurement point,the blockage of the sampling pipeline,the measurement result and the actual value of the existence of a certain delay time and other factors,in addition,due to the characteristics of new energy,when its unit is connected to the grid on a large scale,the thermal power unit needs to increase the utilization rate of new energy through rapid load change,which will lead to significant changes in NOx concentration,and can not react in real time and accurately to the change of the NOx concentration in the SCR inlet.Therefore,it is impossible to guide the SCR reactor to make corresponding action adjustments in a timely and effective manner,so it is important to establish an accurate prediction model of the inlet NOx concentration of the SCR denitrification system.In order to solve the above problems,this study designs an adaptive prediction model of the inlet NOx concentration of SCR denitrification system considering the delay,taking a 660MW coal-fired unit as the research object,and the main research contents are as follows:(1)There is a delay time between the boiler side data measured at each sampling time in the combustion system and the Inlet NOx concentration data of the SCR system,and the prediction model of the inlet NOx concentration of the SCR denitrification system is established directly from the collected data,which cannot accurately express the nonlinear relationship between the boiler side parameters and the NOx,in order to solve this problem,this paper estimates the delay time between the inlet NOx generation and the boiler side parameters of the SCR denitrification system based on the mutual information method,taking into account the coupling between the various variables.This makes the input data of the model more in line with the actual engineering and improves the accuracy of the prediction model.(2)This paper establishes the NOx concentration prediction model at the input of the SCR denitrification system based on LSTM,the LSTM neural network solves the long-term dependence problem of the sequence by virtue of its unique gating design,the single model is susceptible to various factors to affect the training accuracy of the model,the convolutional neural network is introduced,and the CNN-LSTM prediction model is established to improve the stable prediction accuracy of the model,and the two models are compared and analyzed using the same actual data set of the power plant.Experiments have shown that the model of the latter has higher prediction accuracy.(3)Some characteristics of some systems have time-varying nature,the prediction accuracy of the original offline model can not meet the actual needs,in order to solve this problem,the Bayesian optimization algorithm is used to automatically adjust the hyperparameters of the optimization model,the use of Euclidean distance to reconstruct the input data,combined with the CNN-LSTM neural network prediction model to achieve online update of the model,through the actual data to verify the model performance,the method can be changed when the system characteristics change,still have a better prediction accuracy and training speed,To a certain extent,the effectiveness of the guarantee model has a certain practical application value in thermal power plants.更多还原