CN103729569A

CN103729569A - Soft measurement system for flue gas of power-station boiler on basis of LSSVM (Least Squares Support Vector Machine) and online updating

Info

Publication number: CN103729569A
Application number: CN201410025040.1A
Authority: CN
Inventors: 杨婷婷; 吕游; 刘吉臻
Original assignee: North China Electric Power University
Current assignee: North China Electric Power University
Priority date: 2014-01-20
Filing date: 2014-01-20
Publication date: 2014-04-16
Anticipated expiration: 2034-01-20
Also published as: CN103729569B

Abstract

The invention provides a power plant boiler flue gas soft measurement system based on a least squares support vector machine and online update, which belongs to the cross-technical field of thermal technology and artificial intelligence. The system selects the relevant operation and state parameters of the power plant boiler as the input of the model, the flue gas component content to be predicted as the output of the model, selects the historical operation data as the initial training sample, and uses the least squares support vector machine method to establish the initial flue gas emission. Model. In addition, based on the analysis of the time-varying characteristics of flue gas emissions, an update strategy based on sample replacement and sample addition is proposed, and two modes of deleting samples and adding samples are used to solve the parameters and update the model incrementally. . The least squares support vector machine and the online update soft sensor system proposed by the present invention can adaptively improve the model performance with the change of process characteristics, and can realize accurate prediction of flue gas emissions, which is important for the safety and optimal operation of power plant boilers significance.

Description

A Flue Gas Soft Measurement System for Power Plant Boilers Based on LSSVM and Online Update

技术领域technical field

本发明涉及一种基于最小二乘支持向量机（least squares support vectormachine，LSSVM）及在线更新的电站锅炉烟气软测量系统，属于热工技术和人工智能交叉技术领域。The invention relates to a power plant boiler flue gas soft measurement system based on a least squares support vector machine (LSSVM) and online update, belonging to the cross-technical field of thermal technology and artificial intelligence.

背景技术Background technique

为了保证电站锅炉的安全和优化运行，常常需要获取锅炉尾部烟气中飞灰含碳量和NOx排放等参数的相关信息。目前，这些参数常利用飞灰测碳仪和烟气连续监测系统（continuous emission monitoring system，CEMS）等硬件传感器来测量，但是这些仪器的安装和维护成本较高，而且由于工作在恶劣的电磁环境中，经常需要离线维修。因此，采用其他易测的锅炉运行和状态参数通过一定的数学关系模型来对烟气成分含量进行预测，具有重要的工程意义。由于燃烧过程的复杂性和不确定性，建立准确的机理模型往往是非常困难的。近年来，电站的信息化使过程运行数据的获取越来越容易，而且神经网络、支持向量机等人工智能的发展为基于数据的软测量技术提供了有效的工具。其中，最小二乘支持向量机（least squares support vector machine，LSSVM）以结构风险最小化为原则，与神经网络相比具有更好的泛化能力。而且，LSSVM利用等式约束代替不等式约束，将学习问题转化为求解线性方程组，减少了算法的复杂度。In order to ensure the safety and optimal operation of power plant boilers, it is often necessary to obtain relevant information on parameters such as carbon content in fly ash and NOx emissions in the flue gas at the tail of the boiler. At present, these parameters are often measured by hardware sensors such as fly ash carbon detectors and flue gas continuous monitoring system (continuous emission monitoring system, CEMS), but the installation and maintenance costs of these instruments are relatively high, and because they work in harsh electromagnetic environments In , offline maintenance is often required. Therefore, it is of great engineering significance to use other easily measurable boiler operation and state parameters to predict the content of flue gas components through a certain mathematical relationship model. Due to the complexity and uncertainty of the combustion process, it is often very difficult to establish an accurate mechanism model. In recent years, the informatization of power stations has made it easier to obtain process operation data, and the development of artificial intelligence such as neural networks and support vector machines has provided effective tools for data-based soft sensor technology. Among them, the least squares support vector machine (LSSVM) is based on the principle of structural risk minimization, and has better generalization ability than neural networks. Moreover, LSSVM uses equality constraints instead of inequality constraints, transforms the learning problem into solving linear equations, and reduces the complexity of the algorithm.

在利用LSSVM等方法构建烟气软测量模型时，初始样本的筛选非常重要，在从历史运行数据库中选取初始训练样本时应尽可能地使其覆盖全工况。然而事实上，数据库中存储的大都是正常的运行工况，并没有人为主动地调节和设定各个热工参数，因此很难保证所选的样本能覆盖所有工况范围。在模型建立后，运行过程中操作指令和调节参数的改变可能会带来新的工况，而模型将无法对烟气成分含量进行精确预测。另一方面，在运行过程中，煤质的变化以及设备的磨损和维修也会引起烟气排放特性的变迁，建立的初始模型在运行一段时间后预测精度会逐渐下降，若重新构建模型会带来繁重的计算负担，而且也会摒弃原模型中存在的有用信息。因此，利用模型更新来改善其性能，对实现烟气成分含量的精确测量有着重要的意义。When using methods such as LSSVM to construct a flue gas soft sensor model, the screening of initial samples is very important. When selecting initial training samples from the historical operation database, it should cover all working conditions as much as possible. However, in fact, most of the data stored in the database are normal operating conditions, and there is no active adjustment and setting of various thermal parameters, so it is difficult to ensure that the selected samples can cover all operating conditions. After the model is established, changes in operating instructions and adjustment parameters during operation may bring about new working conditions, and the model will not be able to accurately predict the content of flue gas components. On the other hand, during the operation process, changes in coal quality and equipment wear and maintenance will also cause changes in flue gas emission characteristics. The prediction accuracy of the established initial model will gradually decrease after a period of operation. To heavy computational burden, but also discard the useful information existing in the original model. Therefore, using model update to improve its performance is of great significance to achieve accurate measurement of smoke component content.

发明内容Contents of the invention

本发明的目的在于克服现有烟气排放的时变特性，提出了一种基于LSSVM及在线更新的电站锅炉烟气软测量系统。The purpose of the present invention is to overcome the time-varying characteristics of the existing flue gas discharge, and propose a power plant boiler flue gas soft measurement system based on LSSVM and online update.

一般而言，热工过程中烟气排放特性的变化主要由两方面的因素引起：(1)在运行过程中入炉煤质的变化以及设备的磨损和维修等因素，导致过程特性发生变化；这种特性的变化是不可逆的，也即特性变化后不会再回到先前的运行状态。(2)由于生产操作指令以及调节参数的改变从而出现一些新的工况状态；这种特性变化是可逆的，因为随着调节参数的继续变化，过程有可能从现状态切换到先前历史工况中已有的状态。针对这两种特性变化，对应的模型更新方法也有所不同。对于第一种特性变化，需要删除旧样本信息。这是因为旧样本是对先前运行过程的描述，而运行过程已发生了不可逆变化，这些样本便没有任何价值，需要用新的样本来替代，对这种变化的更新应该以样本替换的形式来实现。而第二种变化则是过程正常运行状态的改变和切换，因此需要将新样本信息融入到旧样本中，以此来拓展模型的工作范围，对这种变化的更新应该以样本追加的形式来实现。Generally speaking, the change of flue gas emission characteristics in the thermal process is mainly caused by two factors: (1) The change of coal quality in the furnace during operation, the wear and maintenance of equipment and other factors lead to the change of process characteristics; This characteristic change is irreversible, that is, it will not return to the previous operating state after the characteristic change. (2) Due to changes in production operation instructions and adjustment parameters, some new working conditions appear; this characteristic change is reversible, because as the adjustment parameters continue to change, the process may switch from the current state to the previous historical working conditions existing state. For these two feature changes, the corresponding model update methods are also different. For the first feature change, the old sample information needs to be deleted. This is because the old sample is a description of the previous running process, and the running process has undergone irreversible changes, these samples have no value and need to be replaced by new samples, the update of this change should be done in the form of sample replacement accomplish. The second change is the change and switching of the normal operating state of the process. Therefore, it is necessary to integrate the new sample information into the old sample to expand the working range of the model. The update of this change should be in the form of sample addition. accomplish.

因此，本发明提出通过LSSVM构建初始烟气排放模型，然后利用样本追加和样本替换来实现模型的增量更新。该方法预测精度高、成本低、计算速度快，有利于应用于工程实践之中。Therefore, the present invention proposes to construct an initial smoke emission model through LSSVM, and then use sample addition and sample replacement to realize incremental update of the model. This method has high prediction accuracy, low cost and fast calculation speed, which is beneficial to be applied in engineering practice.

基于LSSVM及在线更新的电站锅炉烟气软测量系统，所述系统包括：A power plant boiler flue gas soft measurement system based on LSSVM and online update, the system includes:

1）LSSVM模型建立单元：收集初始训练样本来构建LSSVM模型，其中：通过传感器测量发电机功率、各磨煤机给煤量、各磨煤机入口一次风量、各层二次风和燃尽风风门开度信号，并将测量值存入DCS历史数据库中；选择上述测量值作为软测量模型的输入变量，要预测的烟气成分含量作为模型的输出变量，从历史运行数据库中选取覆盖范围大且具有代表性的若干段工况作为初始训练样本，记为

其中x_i∈R^p表示第i组输入样本，对应于测量的发电机功率、各磨煤机给煤量、各磨煤机入口一次风量、各层二次风和燃尽风风门开度，y_i∈R为第i组输出样本，对应于烟气成分的含量，p为输入变量个数，n为样本数量，并构建LSSVM模型；1) LSSVM model building unit: collect initial training samples to construct LSSVM model, in which: through sensors to measure generator power, coal feed volume of each coal mill, primary air volume of each coal mill inlet, secondary air and burnout air of each layer The damper opening signal, and the measured value is stored in the DCS historical database; the above-mentioned measured value is selected as the input variable of the soft sensor model, the smoke component content to be predicted is used as the output variable of the model, and the large coverage area is selected from the historical operation database. And several representative operating conditions are used as initial training samples, denoted as

where x _i ∈ R ^p represents the i-th group of input samples, corresponding to the measured generator power, the coal feed volume of each coal mill, the primary air volume of each coal mill inlet, the secondary air of each layer and the opening of the overburning air door, y _i ∈ R is the i-th group of output samples, corresponding to the content of smoke components, p is the number of input variables, n is the number of samples, and constructs the LSSVM model;

LSSVM模型可描述为以下优化问题：The LSSVM model can be described as the following optimization problem:

$\underset{w w,, b b,, ξ ξ}{min min} J J ((w w,, ξ ξ)) = = \frac{11}{22} {w w}^{T T} w w + + \frac{11}{22} γ γ {Σ Σ}_{i i = = 11}^{n no} {ξ ξ}_{i i}^{22} - - - - - - ((11))$

其中，J为目标函数，

是核空间映射函数，w为权重向量，γ为惩罚系数，ξ_i为误差变量，b为偏差，上标T表示矩阵的转置；为解此优化问题，定义Lagrange函数如下：Among them, J is the objective function,

is the kernel space mapping function, w is the weight vector, γ is the penalty coefficient, ξ _i is the error variable, b is the deviation, and the superscript T represents the transposition of the matrix; in order to solve this optimization problem, the Lagrange function is defined as follows:

其中，α=[α₁,…,α_n]^T为Lagrange乘子；利用Lagrange函数对各变量求偏导，并令导数值为零可得到：Among them, α=[α ₁ ,…,α _n ] ^T is the Lagrange multiplier; use the Lagrange function to calculate the partial derivative of each variable, and set the derivative value to zero to get:

消去中间变量w和ξ_i，将其转化为求解线性方程组：Eliminate the intermediate variables w and ξ _i , and transform it into a system of linear equations:

$[\begin{matrix} 00 & {\overset{&RightArrow; &Right Arrow;}{I I}}^{T T} \\ \overset{&RightArrow; &Right Arrow;}{I I} & Ω Ω + + 11 / / γI γI \end{matrix}] [\begin{matrix} b b \\ a a \end{matrix}] = = [\begin{matrix} 00 \\ y the y \end{matrix}] - - - - - - ((44))$

其中y=[y₁,…,y_n]^T，I为n×n阶单位矩阵，Ω={Ω_k|k,l=1,…,n}，且定义为核函数；通过求解方程组得到α和b的值为：where y=[y ₁ ,…,y _n ] ^T , I is an n×n order identity matrix, Ω={Ω _k |k,l=1,…,n}, and Defined as a kernel function; the values of α and b are obtained by solving the equation system:

$\{\begin{matrix} b b = = \frac{{\overset{&RightArrow; &Right Arrow;}{I I}}^{T T} {H h}^{- - 11} y the y}{\overset{&RightArrow; &Right Arrow;}{I I} {H h}^{- - 11} \overset{&RightArrow; &Right Arrow;}{I I}} \\ a a = = {H h}^{- - 11} ((y the y - - \overset{&RightArrow; &Right Arrow;}{I I} b b)) \end{matrix} - - - - - - ((55))$

其中为特征矩阵；in is the characteristic matrix;

从而得到初始的烟气含量的软测量LSSVM模型为：Thus, the soft-sensing LSSVM model of the initial smoke content is obtained as:

$h h ((x x)) = = {Σ Σ}_{i i = = 11}^{n no} {α α}_{i i} K K ((x x,, {x x}_{i i})) + + b b - - - - - - ((66))$

其中核函数选取为高斯径向基函数K(x,x_i)=exp(-||x-x_i||²/σ²)，σ为核函数参数，h(x)为烟气成分含量的预测值；The kernel function is selected as the Gaussian radial basis function K(x, _xi )=exp(-||xx _i || ² /σ ² ), σ is the kernel function parameter, and h(x) is the prediction of the smoke component content value;

2）烟气成分含量预测单元：利用LSSVM模型建立单元所建立的模型对烟气成分含量进行预测，也即将传感器新测得的发电机功率、各磨煤机给煤量、各磨煤机入口一次风量、各层二次风和燃尽风风门开度数据作为输入变量x_q，利用上式得到烟气成分含量的软测量值 2) Flue gas component content prediction unit: use the model established by the LSSVM model building unit to predict the flue gas component content, that is, the generator power newly measured by the sensor, the coal feed of each coal mill, and the inlet of each coal mill The data of the primary air volume, the secondary air of each layer and the opening of the overburning air damper are used as the input variable x _q , and the soft measurement value of the smoke component content is obtained by using the above formula

3）样本预测误差计算单元：当实际的烟气成分含量的传感器测量值y_q采集到后，计算样本(x_q,y_q)的预测误差Er：3) Sample prediction error calculation unit: when the sensor measurement value y _q of the actual smoke component content is collected, calculate the prediction error Er of the sample (x _q , y _q ):

$Er Er = = | | \frac{{\overset{^^}{y the y}}_{q q} - - {y the y}_{q q}}{{y the y}_{q q}} | | \times \times 100100 % % - - - - - - ((77))$

4）预测误差判断单元：判断预测误差：若Er>Δ，Δ为误差阈值，则进入最近样本点选取单元；否则需要判断判断测试样本是否结束，若结束则系统运行结束，否则进入烟气成分含量预测单元；4) Prediction error judging unit: Judging the prediction error: if Er>Δ, Δ is the error threshold, then enter the nearest sample point selection unit; otherwise, it needs to judge whether the test sample is over, if it is over, the system operation ends, otherwise enter the smoke component Content prediction unit;

5）最近样本点选取单元：从历史运行数据中选取距新采样样本(x_q,y_q)最近的样本点(x_k,y_k)：5) The nearest sample point selection unit: select the sample point (x _k , y _k ) closest to the new sampling sample (x _q , y _q ) from the historical operating data:

$k k = = \underset{i i}{arg arg} \underset{i i = = 11,, . . . . . .,, n no}{min min} {| | | | {x x}_{q q} - - {x x}_{i i} | | | |}_{22} - - - - - - ((88))$

6）更新类型判定单元：对新采样的数据样本进行判断，根据以下准则确定更新类型：6) Update type judgment unit: judge the newly sampled data samples, and determine the update type according to the following criteria:

(i)若||x_k-x_q||₂>δ₁，则对模型实施样本追加更新，即直接将新采样样本(x_q,y_q)加入到先前的历史数据库中；(i) If ||x _k -x _q || ₂ >δ ₁ , implement additional sample update on the model, that is, directly add the new sampling sample (x _q , y _q ) to the previous historical database;

(ii)若||x_k-x_q||₂≤δ₁，则对模型实施样本替换更新，即用新采样样本(x_q,y_q)来替换先前历史数据库中满足条件||x_k-x_q||₂≤δ₂的相似样本；(ii) If ||x _k -x _q || ₂ ≤ δ ₁ , implement sample replacement update for the model, that is, use new sample (x _q , y _q ) to replace the previous history database satisfying the condition ||x _k - similar samples for x _q || ₂ ≤ δ ₂ ;

其中δ₁由历史训练数据样本之间的平均距离决定，δ₂设为0.5δ₁；Among them, _δ1 is determined by the average distance between historical training data samples, and _δ2 is set to _0.5δ1 ;

7）特征矩阵更新单元：根据确定的更新类型，对LSSVM模型建立单元获得的初始LSSVM模型进行增量更新，即对特征矩阵H^-1的计算更新，其中，将更新策略分解为样本增加和样本删减两种模式：根据更新类型确定子单元的确定方案，若实施样本追加更新，则直接进行样本增加；若实施样本替换更新，则先进行有关样本删减，然后再进行样本增加；7) Feature matrix update unit: according to the determined update type, incrementally update the initial LSSVM model obtained by the LSSVM model building unit, that is, calculate and update the feature matrix H ^-1 , where the update strategy is decomposed into sample addition and sample Two modes of deletion: determine the determination plan of the subunit according to the update type, if the sample addition update is implemented, the sample increase will be performed directly; if the sample replacement update is implemented, the relevant sample deletion will be performed first, and then the sample increase will be performed;

(i)样本删减模式(i) Sample subtraction mode

记要删减的样本为(x_s,y_s)，交换特征矩阵H中的第n行和第s行、第n列和第s列后，得到新的特征矩阵为：Remember that the sample to be deleted is (x _s , y _s ), after exchanging the nth row and sth row, nth column and sth column in the feature matrix H, the new feature matrix is obtained as:

其中 $k = K (x_{s}, x_{s}) + \frac{1}{γ},$ k_s=[K(x₁,x_s),…,K(x_s-1,x_s),K(x_n,x_s),K(x_s+1,x_s),…,K(x_n-1,x_s)]^T；in $k = K (x_{the s}, x_{the s}) + \frac{1}{γ},$ k _s =[K(x ₁ ,x _s ),…,K(x _s-1 ,x _s ),K(x _n ,x _s ),K(x _s+1 ,x _s ),…,K( x _n-1 ,x _s )] ^T ;

若记交换第n行和第s行：

第n列和第s列：

对应的初等矩阵分别为

和

则有

而且：If you remember to exchange the nth row and the sth row:

Column n and column s:

The corresponding elementary matrices are

and

then there is

and:

${H h}_{00}^{- - 11} = = {T T}_{n no,, s the s}^{- - 11} {H h}^{- - 11} {T T}_{n no,, s the s}^{- - 11} = = {T T}_{n no,, s the s} {H h}^{- - 11} {T T}_{n no,, s the s} - - - - - - ((1010))$

若H₀的逆矩阵记为：If the inverse matrix of H ₀ is written as:

${H h}_{00}^{- - 11} \overset{Δ Δ}{= =} [\begin{matrix} {h h}_{1111} & {h h}_{1212} \\ {h h}_{1212}^{T T} & {h h}_{22 twenty two} \end{matrix}] - - - - - - ((1111))$

则根据分块求逆矩阵，可以得到：Then according to the block inversion matrix, we can get:

${H h}_{00} = = {[\begin{matrix} {h h}_{1111} & {h h}_{1212} \\ {h h}_{1212}^{T T} & {h h}_{22 twenty two} \end{matrix}]}^{- - 11} = = [\begin{matrix} {(({h h}_{1111} - - {h h}_{1212} {h h}_{22 twenty two}^{- - 11} {h h}_{22 twenty two}^{T T}))}^{- - 11} & * * \\ * * & * * \end{matrix}] - - - - - - ((1212))$

对比式和式可得到新的特征矩阵H₁的逆矩阵为：Comparing the formula and the formula, the inverse matrix of the new characteristic matrix H ₁ can be obtained as:

${H h}_{11}^{- - 11} = = {h h}_{1111} - - {h h}_{1212} {h h}_{22 twenty two}^{- - 11} {h h}_{1212}^{T T} - - - - - - ((1313))$

这里只给出删除单个样本(x_s,y_s)的情形，若要删除多个样本则依次进行；Here only the case of deleting a single sample (x _s , y _s ) is given, and if multiple samples are to be deleted, proceed sequentially;

(ii)样本增加模式(ii) Sample augmentation mode

记要增加的样本为(x_t,y_t)，仍记当前模型的特征矩阵为H，则新样本下的特征矩阵H₂可以记为：Note that the sample to be added is (x _t , y _t ), and the feature matrix of the current model is still recorded as H, then the feature matrix H ₂ under the new sample can be written as:

其中 $k_{t} = {[K (x_{1}, x_{t}), K (x_{2}, x_{t}), . . ., K (x_{n}, x_{t})]}^{T}, k = K (x_{t}, x_{t}) + \frac{1}{γ};$ in $k_{t} = {[K (x_{1}, x_{t}), K (x_{2}, x_{t}), . . ., K (x_{no}, x_{t})]}^{T}, k = K (x_{t}, x_{t}) + \frac{1}{γ};$

由分块矩的求逆公式可得到H₂的逆为：From the inverse formula of the block moment, the inverse of _H2 can be obtained as:

${H h}_{22}^{- - 11} = = [\begin{matrix} {H h}^{- - 11} + + {H h}^{- - 11} {k k}_{t t} {ρ ρ}^{- - 11} {k k}_{t t}^{T T} {H h}^{- - 11} & - - {H h}^{- - 11} {k k}_{t t} {ρ ρ}^{- - 11} \\ - - {ρ ρ}^{- - 11} {k k}_{t t}^{T T} {H h}^{- - 11} & {ρ ρ}^{- - 11} \end{matrix}] - - - - - - ((1515))$

其中 $ρ = k - k_{t}^{T} H^{- 1} k_{t} .$ in $ρ = k - k_{t}^{T} h^{- 1} k_{t} .$

8）软测量模型更新单元：将求得的新的H^-1带入式，得到相应的模型参数α和b，实现对烟气软测量模型h(x)的更新；8) Soft sensor model update unit: bring the obtained new H ^-1 into the formula to obtain the corresponding model parameters α and b, and realize the update of the flue gas soft sensor model h(x);

本发明利用LSSVM构建初始的烟气软测量模型，并对模型在线更新，减少了模型的计算复杂度，有利于工程实现，能够对锅炉烟气各成分进行精确地预测。本发明将LSSVM建模方法与增量更新相结合，具有以下显著优势：The invention utilizes LSSVM to construct an initial flue gas soft measurement model, and updates the model online, thereby reducing the computational complexity of the model, facilitating engineering realization, and accurately predicting components of boiler flue gas. The present invention combines the LSSVM modeling method with incremental updating, and has the following significant advantages:

1)选择电站发电机功率、各磨煤机给煤量、各磨煤机入口一次风量、各层二次风和燃尽风风门开度作为模型的输入，能够对锅炉烟气排放特性进行全面的描述；1) Select the generator power of the power station, the coal feed rate of each coal mill, the primary air volume of each coal mill inlet, the secondary air of each layer and the opening of the overburning air damper as the input of the model, which can comprehensively analyze the characteristics of boiler flue gas emission. description of;

2)从历史运行数据库中选取覆盖范围大且具有代表性的若干段工况作为训练样本来构建LSSVM初始烟气软测量模型，具有较高的预测精度；2) From the historical operation database, several representative working conditions with large coverage are selected as training samples to construct the LSSVM initial smoke soft sensor model, which has high prediction accuracy;

3)将更新策略分为样本追加和样本替换，针对排放特性变化的本质来更新模型；3) The update strategy is divided into sample addition and sample replacement, and the model is updated according to the essence of emission characteristics changes;

4)利用增量的方法实施更新，减少了计算的复杂度；4) The incremental method is used to implement the update, which reduces the computational complexity;

5)应用本发明，不增加任何硬件，而且易于工程现场应用，成本低，预测结果精确可靠。5) The application of the present invention does not add any hardware, and is easy to apply on the engineering site, with low cost and accurate and reliable prediction results.

附图说明Description of drawings

图1是烟气排放特性的变化和对应的样本更新过程；Figure 1 shows the change of smoke emission characteristics and the corresponding sample update process;

图2是某电站锅炉的结构示意图；Fig. 2 is a structural schematic diagram of a power plant boiler;

图3是本发明系统的结构框图；Fig. 3 is a structural block diagram of the system of the present invention;

图4是利用本发明对某燃煤锅炉烟气排放中NOx的含量进行软测量得到的预测结果对比示意图。其中，前1100组为初始的训练样本，后270组为测试样本。Fig. 4 is a schematic diagram of comparison of prediction results obtained by using the present invention to perform soft measurement of NOx content in flue gas emissions from a certain coal-fired boiler. Among them, the first 1100 groups are the initial training samples, and the last 270 groups are the test samples.

具体实施方式Detailed ways

下面结合附图和实施例对本发明作详细说明。本实施例在以本发明技术方案为前提下进行实施，但本发明的保护范围不限于下述的实施例。The present invention will be described in detail below in conjunction with the accompanying drawings and embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, but the protection scope of the present invention is not limited to the following embodiments.

本实施例对某660MW电站锅炉烟气排放中NOx的含量进行软测量。参考图1，考虑一单入单出烟气排放特性的变化和对应的样本更新过程，图1(a)中工况状态I中的样本是从历史数据库中选出的具有代表性的初始样本，样本空间为x∈[x₁,x₂]，并基于I中的样本建立初始的LSSVM烟气排放模型y=f(x)。运行过程中调节参数的改变将带来新的工况，运行状态将转换到状态II以及状态III（如图中虚线箭头所示），此时样本在x∈[x₃,x₄]和x∈[x₅,x₆]范围内。而这种状态转换是可逆的，也即是说，随着过程的继续，运行工况有可能重新回到状态I。因此，需要把工况状态II和III中的样本追加入初始状态I中，对由初始状态I建立的排放模型进行拓展，使其覆盖更大的工况范围，从而完成模型的更新。拓展后的模型如图1(b)所示，可以看出，经过更新后模型由初始运行范围x∈[x₁,x₂]拓展到x∈[x₁,x₆]。图1(c)中所示，排放特性在初始工况x∈[x₁,x₂]上发生了不可逆变化，从工况状态I变化到状态IV（如图中实线箭头所示）。引起这种变化的原因很可能是设备的磨损或煤质等外界因素发生了改变，而且，这种工况改变是不可逆的。这时需要用新采集的样本来替换旧样本以实现模型更新，使其能够描述新的过程特性。综合考虑图1(a)和(c)中的过程变化，更新后的烟气排放模型及样本分布如图1(d)所示。In this embodiment, soft measurement is performed on the content of NOx in the flue gas emission of a 660MW power plant boiler. Referring to Fig. 1, consider the change of a single-input-single-out flue gas emission characteristic and the corresponding sample update process. The sample in working condition I in Fig. 1(a) is a representative initial sample selected from the historical database , the sample space is x∈[x ₁ ,x ₂ ], and the initial LSSVM smoke emission model y=f(x) is established based on the samples in I. The change of the adjustment parameters during the operation will bring a new working condition, and the operating state will be converted to state II and state III (as shown by the dotted arrow in the figure). At this time, the sample is between x∈[x ₃ ,x ₄ ] and x ∈[x ₅ ,x ₆ ] range. And this state transition is reversible, that is to say, as the process continues, the operating condition may return to state I again. Therefore, it is necessary to add the samples in working condition states II and III to the initial state I, and expand the emission model established by the initial state I to cover a larger range of working conditions, so as to complete the update of the model. The expanded model is shown in Figure 1(b). It can be seen that the updated model extends from the initial operating range x∈[x ₁ ,x ₂ ] to x∈[x ₁ ,x ₆ ]. As shown in Figure 1(c), the emission characteristics have undergone an irreversible change in the initial operating condition x∈[x ₁ ,x ₂ ], from state I to state IV (as shown by the solid arrow in the figure). The reason for this change is likely to be the change of external factors such as equipment wear or coal quality, and this change in working conditions is irreversible. At this time, it is necessary to replace the old samples with new samples to update the model so that it can describe the new process characteristics. Considering the process changes in Figure 1(a) and (c), the updated smoke emission model and sample distribution are shown in Figure 1(d).

图2是某燃煤锅炉的结构示意图。如图2所示，锅炉形式为单炉膛Π型锅炉，并采用切圆燃烧方式，以获得沿炉膛水平断面较为均匀的空气动力场。主燃烧器分上下两组布置，并拉开一定的距离，降低燃烧器区域热负荷，有效减少炉膛的结焦。在上层煤粉喷嘴上方布置有四层分离燃尽风（SOFA）喷嘴，以补充燃料后期燃烧所需要的空气。尾部烟道中装有烟气连续监测系统（continuousemission monitoring system，CEMS），用来测量烟气排放中NOx的含量。但是，CEMS在工作过程中，经常需要离线维修，因此为了锅炉的安全和优化运行，需要构建NOx排放的软测量模型来进行冗余测量。Figure 2 is a schematic diagram of the structure of a coal-fired boiler. As shown in Figure 2, the boiler is a single furnace Π-type boiler, and adopts a tangential combustion method to obtain a relatively uniform aerodynamic field along the horizontal section of the furnace. The main burners are arranged in upper and lower groups, and they are separated by a certain distance to reduce the heat load of the burner area and effectively reduce the coking of the furnace. Four layers of separated burn-off air (SOFA) nozzles are arranged above the upper layer of pulverized coal nozzles to supplement the air required for the post-combustion of fuel. The flue gas continuous monitoring system (continuousmission monitoring system, CEMS) is installed in the tail flue to measure the NOx content in the flue gas emission. However, during the working process of CEMS, offline maintenance is often required. Therefore, for the safety and optimal operation of the boiler, it is necessary to build a soft sensor model of NOx emission for redundant measurement.

根据相关机理分析，选择影响锅炉NOx排放的以下参数作为模型的输入变量：选择发电机功率描述负荷的影响，将6台磨煤机给煤量、6台磨煤机入口一次风量和6个燃料风开度信号进行主成分分析，并利用提取得到的特征主成分变量来描述一次风粉量沿炉高分配对NOx排放的影响，选择8个二次风门开度(AA、AB、BC、CC、DD、DE、EF、FF)来描述二次配风方式对NOx排放的影响，选择4层燃尽风门开度(UA、UB、UC、UD)来描述燃尽风的影响。以上所有参数的运行值通过传感器测量并存入DCS历史数据库中，也即电站历史数据库中均有对应的测点。According to the relevant mechanism analysis, the following parameters that affect boiler NOx emissions are selected as the input variables of the model: generator power is selected to describe the impact of load, the coal feed volume of 6 coal mills, the primary air volume at the inlet of 6 coal mills and the 6 fuel Principal component analysis was performed on the wind opening signal, and the extracted characteristic principal component variables were used to describe the influence of primary air powder distribution along the furnace height on NOx emissions. Eight secondary air door openings (AA, AB, BC, CC , DD, DE, EF, FF) to describe the impact of the secondary air distribution mode on NOx emissions, and select four layers of burn-up damper openings (UA, UB, UC, UD) to describe the impact of burn-up air. The operating values of all the above parameters are measured by sensors and stored in the DCS historical database, that is, there are corresponding measuring points in the power plant historical database.

从历史数据库中选取机组负荷跨度较大（从300MW到660MW）的连续一周的以上各参数的运行数据，采样周期为60s。对数据进行清洗后，将其分为两组：其中1100组作为初始训练样本，另外270组未参加训练的工况段作为测试样本。Select the operation data of the above parameters for a continuous week with a large unit load span (from 300MW to 660MW) from the historical database, and the sampling period is 60s. After the data is cleaned, it is divided into two groups: 1100 groups are used as initial training samples, and the other 270 groups of working conditions that have not participated in training are used as test samples.

请参考图3，本发明提出的基于LSSVM及在线更新的电站锅炉烟气中NOx的软测量系统的框图，该系统包括：Please refer to Fig. 3, the block diagram of the soft measurement system of NOx in the utility boiler flue gas based on LSSVM and online update that the present invention proposes, and this system comprises:

LSSVM模型建立单元：将初始的1100组样本作为训练样本，记为其中x_i∈R^p表示第i组输入样本，对应于发电机功率、各磨煤机给煤量、各磨煤机入口一次风量、各层二次风和燃尽风风门开度，y_i∈R为第i组输出样本，对应于烟气成分NOx的含量，p为输入变量个数，n为初始训练样本数量，并利用初始样本构建LSSVM模型，实现对NOx含量的预测；LSSVM model building unit: take the initial 1100 sets of samples as training samples, denoted as where x _i ∈ R ^p represents the i-th group of input samples, corresponding to the generator power, the coal feed volume of each coal mill, the primary air volume of each coal mill inlet, the secondary air of each layer and the opening of the overburning air door, y _i ∈R is the i-th group of output samples, corresponding to the NOx content of the flue gas component, p is the number of input variables, n is the number of initial training samples, and the initial samples are used to construct the LSSVM model to realize the prediction of the NOx content;

烟气成分含量预测单元：对新采样的样本x_q预测，得到预测值

Smoke component content prediction unit: predict the newly sampled sample x _q to obtain the predicted value

样本预测误差计算单元：根据式计算样本(x_q,y_q)的预测误差Er；Sample prediction error calculation unit: calculate the prediction error Er of the sample (x _q , y _q ) according to the formula;

预测误差判断单元：判断预测误差：若Er>Δ，则则进入更新单元，这里设定Δ=0.07；Δ为误差阈值，，否则进入测试样本判断单元；Prediction error judgment unit: judge prediction error: if Er>Δ, then enter the update unit, here set Δ=0.07; Δ is the error threshold, otherwise enter the test sample judgment unit;

最近样本点选取单元：从历史数据中选取距新采样样本最近的样本点(x_k,y_k)；The nearest sample point selection unit: select the sample point (x _k , y _k ) closest to the new sampling sample from the historical data;

更新类型确定单元：根据新采样样本(x_q,y_q)与其最近的样本(x_k,y_k)之间的距离||x_q-x_k||₂来确定更新类型；Update type determination unit: determine the update type according to the distance ||x _q -x _k || ₂ between the new sampling sample (x _q , y _q ) and its nearest sample (x _k , y _k );

特征矩阵更新单元：：根据确定的更新类型，由式～式增量计算特征矩阵H的逆；Feature matrix update unit: according to the determined update type, calculate the inverse of the feature matrix H incrementally from formula to formula;

软测量模型更新单元：将求得的新的H^-1代入式，得到相应的模型参数α和b，对模型进行更新。Soft sensor model update unit: Substitute the obtained new H ^-1 into the formula to obtain the corresponding model parameters α and b, and update the model.

为了验证LSSVM及在线更新的软测量模型的预测效果，同时还建立了未实施更新的LSSVM模型进行对比，采用方根误差(R_MSE)和平均相对误差(M_RE)作为模型预测效果的评价指标： $R_{MSE} = \sqrt{Σ_{i = 1}^{N} {(y_{i} - {\hat{y}}_{i})}^{2} / N}, M_{RE} = \frac{1}{N} Σ_{i = 1}^{N} | \frac{y_{i} - {\hat{y}}_{i}}{y_{i}} | \times 100 % .$ LSSVM模型更新前后对测试样本的预测结果对比见表1。In order to verify the prediction effect of LSSVM and the online updated soft sensor model, the LSSVM model without update is also established for comparison, and the root square error (R _MSE ) and mean relative error (M _RE ) are used as the evaluation indicators of the model prediction effect : $R_{MSE} = \sqrt{Σ_{i = 1}^{N} {({the y}_{i} - {\hat{the y}}_{i})}^{2} / N}, m_{RE} = \frac{1}{N} Σ_{i = 1}^{N} | \frac{{the y}_{i} - {\hat{the y}}_{i}}{{the y}_{i}} | \times 100 % .$ The comparison of the prediction results of the test samples before and after the update of the LSSVM model is shown in Table 1.

表1Table 1

如表1所示，采用LSSVM更新后的软测量模型的性能与未更新的模型相比有很大的改善，预测精度得到了提高。图4是利用本发明和未加更新的LSSVM模型对锅炉烟气中NOx排放进行建模得到的预测结果对比图。其中，前1100组为初始的训练样本，后270组为测试样本。由图可以看出，当排放特性改变时，未加更新的LSSVM模型对NOx的预测误差逐渐增大，而由于本发明施加了在线更新，能一直保持较高的预测精度。As shown in Table 1, the performance of the soft sensor model updated with LSSVM is greatly improved compared with the non-updated model, and the prediction accuracy is improved. Fig. 4 is a comparison chart of prediction results obtained by using the present invention and the LSSVM model without updating to model NOx emissions in boiler flue gas. Among them, the first 1100 groups are the initial training samples, and the last 270 groups are the test samples. It can be seen from the figure that when the emission characteristics change, the prediction error of the LSSVM model without updating gradually increases for NOx, but due to the online update applied by the present invention, it can always maintain a high prediction accuracy.

本发明还提供一种模型，该模型根据以上所述的基于LSSVM及在线更新的电站锅炉烟气软测量方法建模得到。本发明提出了基于LSSVM及在线更新的电站锅炉烟气软测量系统，减少了模型的计算复杂度，有利于工程实现，对电站锅炉的安全和优化运行有着重要的意义。The present invention also provides a model, which is obtained according to the above-mentioned method based on LSSVM and online updating of the flue gas soft sensor method of the utility boiler. The invention proposes a power plant boiler flue gas soft measurement system based on LSSVM and online update, which reduces the computational complexity of the model, is beneficial to engineering realization, and has important significance for the safety and optimal operation of the power plant boiler.

上述实例用来说明本发明，而不是对其进行限制。在本发明的权利要求保护范围内，任何对对本发明的修改都落入本发明的保护范围内。The above examples are presented to illustrate the invention, not to limit it. Within the protection scope of the claims of the present invention, any modification to the present invention falls within the protection scope of the present invention.

Claims

1. A utility boiler flue gas soft sensor system based on LSSVM and online update, characterized in that the soft sensor system includes a model prediction subsystem, a performance detection subsystem and a model update subsystem, and the model prediction subsystem is connected to To the performance prediction subsystem, the performance prediction subsystem is connected to the model update subsystem, and the model update subsystem is fed back to the model prediction subsystem, wherein the model prediction subsystem includes an LSSVM model building unit and a smoke component content prediction unit; The performance detection subsystem includes a sample prediction error calculation unit and a prediction error judgment unit; the model update subsystem includes a nearest sample point selection unit, an update type judgment unit, a feature matrix update unit, and a soft sensor model update unit.

2. soft sensing system according to claim 1, is characterized in that,

The LSSVM model building unit is used to collect initial training samples to build the LSSVM model;

The smoke component content prediction unit uses this model to predict the smoke component content;

The sample prediction error calculation unit is used to calculate the prediction error Er of the sample (x _q , y _q ) after the sensor measurement value y _q of the actual smoke component content is collected;

The prediction error judging unit is used to judge the prediction error: if Er>Δ, and Δ is the error threshold, then enter the nearest sample point selection unit; otherwise, it needs to judge whether the test sample is over, if it is over, the system operation ends, otherwise enter the smoke Gas component content prediction unit;

The nearest sample point selection unit is used to select the sample point (x _k , y _k ) closest to the new sampling sample (x _q , y _q ) from the historical operation data, where

The update type determining unit judges and determines the update type according to the distance between the new sampling sample (x _q , y _q ) and its nearest sample (x _k , y _k );

The feature matrix update unit, according to the determined update type, incrementally updates the initial LSSVM model obtained by the LSSVM model building unit, that is, updates the calculation of the feature matrix H ^-1 ;

The soft sensor model updating unit is used to calculate the obtained new model characteristic matrix H ^-1 to obtain corresponding model parameters α and b, realize the update of the flue gas soft sensor model h(x), and then use the new The model predicts the content of smoke components.

3. The soft sensor system according to claim 2, characterized in that, constructing the LSSVM model is specifically: measuring the power of the generator, the coal feed rate of each coal mill, the primary air volume of each coal mill inlet, and the secondary air volume of each layer through sensors. Wind and overburning air door opening signals, and store the measured values in the DCS historical database; select the above measured values as the input variables of the soft sensor model, and the smoke component content to be predicted as the output variable of the model, from the historical operation database Select several representative operating conditions with a large coverage area as the initial training samples, denoted as

where x _i ∈ R ^p represents the i-th group of input samples, corresponding to the measured generator power, the coal feed volume of each coal mill, the primary air volume of each coal mill inlet, the secondary air of each layer and the opening of the overburning air door, y _i ∈ R is the i-th group of output samples, corresponding to the content of smoke components, p is the number of input variables, n is the number of samples, and the LSSVM model is constructed.

4. The soft sensor system according to claim 2, wherein the prediction of the flue gas component content is specifically: the generator power newly measured by the sensor, the coal feed rate of each coal mill, the inlet of each coal mill The data of the primary air volume, the secondary air of each layer and the opening of the overburning air damper are used as the input variable x _q , which is substituted into the soft sensor LSSVM model formula of the initial smoke content:

Get the soft measurement value of the content of smoke components

The kernel function is chosen as the Gaussian radial basis function

σ is the kernel function parameter, and h(x) is the predicted value of smoke component content.

5. soft sensor system according to claim 4 is characterized in that, wherein the calculation of predictive error Er is as follows:

Er = | \frac{{\hat{the y}}_{q} - {the y}_{q}}{{the y}_{q}} | \times 100 % .

6. The soft sensor system according to claim 2, characterized in that, the judgment criterion adopted in the update type determination unit is:

(i) If ||x _k -x _q || ₂ >δ ₁ , implement additional sample update on the model, that is, directly add the new sampling sample (x _q , y _q ) to the previous historical database;

(ii) If ||x _k -x _q || ₂ ≤ δ ₁ , implement sample replacement update for the model, that is, use new sample (x _q , y _q ) to replace the previous history database satisfying the condition ||x _k - similar samples for x _q || ₂ ≤ δ ₂ ;

Among them, δ ₁ is determined by the average distance between historical training data samples, and δ ₂ is set to 0.5δ ₁ .

7. The soft sensor system according to claim 2, characterized in that, the update strategy of the feature matrix update unit is divided into two modes: sample increase and sample deletion: if the sample is added and updated, the sample is directly increased; If the sample replacement update is implemented, the sample is deleted first, and then the sample is added.