CN104715142B - A kind of station boiler NOxDischarge dynamic soft-measuring method - Google Patents
A kind of station boiler NOxDischarge dynamic soft-measuring method Download PDFInfo
- Publication number
- CN104715142B CN104715142B CN201510064480.2A CN201510064480A CN104715142B CN 104715142 B CN104715142 B CN 104715142B CN 201510064480 A CN201510064480 A CN 201510064480A CN 104715142 B CN104715142 B CN 104715142B
- Authority
- CN
- China
- Prior art keywords
- particle
- model
- value
- fitness
- particles
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
- Feedback Control In General (AREA)
Abstract
The present invention discloses a kind of station boiler NOxDynamic soft-measuring method is discharged, is comprised the following steps：Data acquisition and pretreatment, initialization APSO algorithm etc., the present invention is using boiler combustion system about running the input with state parameter as model, the output of model is used as using discharged nitrous oxides concentration, history data is chosen as training sample, soft sensor modeling instrument is used as using support vector regression, with reference to nonlinear auto regressive moving average thought, it is contemplated that the input of model and the order of output variable, the ability for making soft-sensing model that there is description dynamic changing process.The present invention effectively can track and predict the change of NOx emission in boiler combustion dynamic running process, and the safety and optimization operation to station boiler have important meaning.
Description
Technical Field
The invention relates to the field of thermal technology and artificial intelligence cross technology, in particular to NO of a power station boiler x An emission dynamic soft measurement method.
Background
With the rapid development of economy and power industry in China, nitrogen oxides generated by coal-fired boilers of thermal power plants become main sources of atmospheric nitrogen oxide pollution, and NO of coal-fired units is required to meet increasingly strict environmental requirements x The control of emissions puts higher demands. Realizing NO of power station boiler x The premise of emission optimization control is to establish effective NO x And (4) discharging the soft measurement model.
Due to NO x The generation of complex combustion products of coal, and the boiler combustion system is also an extremely complex part of a power station system, and the related input variablesThe quantity is large, strong nonlinearity and strong coupling exist, and an accurate mechanism model is difficult to establish. In recent years, with the popularization of power station DCS and SIS systems, a large amount of historical operating data is reserved, and a good application environment is provided for data-driven soft measurement technologies such as neural networks and support vector machines. The support vector regression (SVMR) based on the principle of minimizing structural risk shows better generalization capability than a neural network, and well solves the difficulties of small sample, nonlinearity, overfitting and the like in soft measurement modeling.
In the prior art, NO of power station boiler is mostly modeled by adopting a steady-state modeling method x And (3) establishing a steady-state soft measurement model for discharge, wherein the steady-state model has higher precision because the sampling data under the steady-state working condition is more accurate. However, the following problems exist in the application of the steady-state soft measurement model in the actual operation of the utility boiler:
firstly, the method comprises the following steps: usually, only data matched with time sequences of various measuring points are taken as input and output of a steady-state soft measurement model, however, most of the actual operation process of the power station boiler is in dynamic change, a certain state of the system output quantity at the current moment is determined by the input state of the previous period, but not the input state of a certain time point, namely the time delay characteristic of the combustion process is ignored;
secondly, the method comprises the following steps: the steady-state soft measurement model needs the system to be kept under the steady-state working condition for a certain time to ensure the accuracy of obtaining steady-state data, but also ensures that the applicability of the steady-state soft measurement model is only limited under the steady-state working condition, and lacks the description of the dynamic change characteristic of the process, so that once the working condition point changes or disturbance occurs in the operation of the boiler, the tracking capability of the steady-state soft measurement model is greatly reduced;
thirdly, the method comprises the following steps: the steady-state operation conditions of the utility boiler are often difficult to meet, and the application of a steady-state model is also limited.
Disclosure of Invention
The invention aims to: the invention aims to solve the defects in the prior art and provide a basePower station boiler flue gas NO realized by support vector regression (SVMR) and nonlinear autoregressive moving average model framework (NARMAX) x An emissions dynamic soft measurement method.
The technical scheme is as follows: the invention relates to a NO of power station boiler x An emissions dynamic soft measurement method comprising the steps of:
(1) Data acquisition and preprocessing: determining model input variable and output variable, and acquiring original sample from DCS systemFor original sampleAdopting a normalization pretreatment method and measuring NO under various working conditions x Converting the concentration to a value below 6% oxygen; wherein x is i ∈R m Representing the input samples of the ith group of models, m being the number of input variables, y i E, R represents the output sample of the ith group of models, and n is the number of samples;
because the monitoring data of the environmental protection bureau are all unified to set a standard combustion condition: o in the residual flue gas 2 The concentration of (2) is 6%, therefore, in the patent, the residual flue gas O in actual combustion is referred to under each working condition 2 The concentration of (a) is not necessarily 6%, so that conversion is required according to a conversion formula;
(2) Initializing a self-adaptive particle swarm algorithm: each particle p (i) contains m +3 variables, including m input variables x i Corresponding order d i Order d corresponding to 1 output variable y y The penalty factor C and the nuclear parameter sigma of the SVMR model are two parameters which need to be manually set during SVMR modeling, and the values of the two parameters have great influence on the soft measurement result of the model, so that the two parameters need to be optimized together;
(3) Arranging input data and output data into a nonlinear autoregressive moving average model framework NARMAX model structure: input variable x contained by particle p (i) i Order d of i And order d of the output variable y y The input data and the output data are arranged into a NARMAX model structure, and then the input X (t) and the output Y (t) of the soft measurement model are obtained through calculation as follows:
in the above formula, T is the sampling period of the collected measured point data, and X (T) is the model input quantity corresponding to the sampling time T, and comprises m input variables X i Consecutive d before sampling instant t i The state quantity of one sampling period, and the succession d of measured quantities y before the sampling instant t y A state quantity for each sampling period; y (t) is a model output quantity corresponding to the sampling time t and is a state quantity of the measured Y at the sampling time t;
(4) Initializing the SVMR model: using a Gaussian kernel function K (x) i ,x j )＝exp(-||x i -x j || 2 /(2σ 2 ) RBF as the kernel function of SVMR modeling, the SVMR model is set with the penalty factor C and the kernel parameter σ contained by the particle p (i):
in the above formula, the first and second carbon atoms are,α i and b is a support vector regression machine parameter,is a support vector, l is the number of the support vectors, and sigma is an RBF nuclear parameter;
(5) Carrying out SVMR training and fitting, and calculating individual fitness f of particles i And (3) keeping the optimal individuals: taking X (t) as SVMR model input, Y (t) as SVMR model output, taking the first 3/4 of the sorted sample set as training samples, taking the second 1/4 as prediction samples, and carrying out model training and fitting; calculating the number of particles by taking the deviation index of the prediction sample as a fitness functionFitness f i And obtaining the optimal fitness f of the population m And retaining its particle position P m Then the fitness function for particle p (i) is:
in the above formula, n is the length of the predicted time series, y j Andrespectively an actual measurement value and a model prediction value at the jth moment;
(6) Optimizing convergence judgment: if f m <f e Wherein f is e To expect a convergence threshold, when K = K max If yes, then the optimization is terminated, and the step (9) is executed; otherwise, letting k = k +1, executing step (7); wherein K is the current iteration number, K max For maximum number of iterations, the optimization range of the order is [ d ] min ,d max ]；
(7) Calculating the average fitness of the particle population, and dividing the particle population into three subgroups: the fitness value of the particle p (i) in the kth iteration is f i k The fitness value of the population-optimal particle is f m (ii) a The average fitness of the particle swarm isMake the fitness value better thanIs obtained by averaging the fitness of the particlesAccording to the individual fitness f of the particles i k Dividing the data into three subgroups of local optimization, balance optimization and global optimization, and respectively calculating corresponding inertia weight w and learning factor c 1 、c 2 Classifying corresponding to the step (7.3), and calculating the three coefficients;
(8) Individual velocity of particlesAnd location updating: according to the inertia weight w and the learning factor c obtained in the step (7) 1 、c 2 For individual velocity V of the particles i And position P i Performing calculation update according to the following formula
v in ＝wv in +c 1 rand()(b in -p in )+c 2 rand()(b gn -p in )
p in ＝p in +v in
Wherein v is in Is the particle velocity, p in As particle position, b in For the optimal position of the individual, b gn The optimal position of the group is obtained;
(9) Outputting an optimal result, and finishing the establishment of a dynamic soft measurement model: output f m D contained in the corresponding optimum individual i 、d y The penalty factor C and the nuclear parameter sigma are used as setting parameters to establish a final dynamic soft measurement SVMR model;
(10) And (3) realizing dynamic soft measurement: for any newly acquired sample, after being preprocessed, the dynamic soft measurement SVMR model obtained in the step 9 is input, and then corresponding soft measurement output can be obtained.
Further, in the step (1), NO before the SCR reactor at the tail part of the boiler furnace of the power station is selected x Inlet concentration as output y i Selecting NO in the dynamic operation process of the boiler unit x Combustion system operating variables with influence of emissions as input variables x i (ii) a While the original sampleThe dynamic continuous operation working condition data which is large in coverage range and operates in a variable load mode in the DCS.
Further, the original sample is processed in the step (1)The specific process adopting the normalization pretreatment comprises the following steps:
scaling the original value of each input data to be within a specified interval according to the following formula:
wherein, x is a value before normalization, x' is a value after normalization, and max (x) and min (x) are respectively a maximum value and a minimum value of an interval after normalization;
because the oxygen content of the flue gas is inconsistent under various working conditions, NO is convenient to compare x The value of NO measured under each condition is determined according to the following formula x The concentration is converted to a value at 6% oxygen:
where ρ '(NOx) and ρ' (O) 2 ) Respectively representing converted NO x Concentration and O 2 Concentrations, rho (NOx) and rho (O) 2 ) Respectively representing actually measured NO x Concentration and O 2 And (4) concentration.
Further, in the step (7), the individual fitness f of the particles is determined according to i k The process of dividing the particle population into three subgroups was:
(7.1) calculating the mean fitness of the population of particles, and then setting the fitness value of the particle p (i) in the kth iteration to f i k The fitness value of the population-optimal particle is f m Mean fitness value of population of particles of
(7.2) the fitness value is better than that of theBy averaging the fitness of the particlesDefinition ofThe early-maturing convergence degree of the particle swarm is evaluated, and the smaller delta is, the more the particle swarm tends to converge early;
(7.3) finally, according to the individual fitness f of the particles i k It is classified into the following three categories:
(1) if it isThen the particle is a better particle in the group and is divided into a local optimizing subgroup, and a smaller inertia weight w is taken according to the following formula to make the particle perform local search, and a larger c is taken at the same time 1 And smaller c 2 To enhance the learning ability of the particle individuals and strengthen the local optimizing ability of the particles:
(2) if it isThe particle is a common particle in the group and is divided into a balance optimizing group, and the balance optimizing group has good global optimizing capability and local optimizing capability, maintains the inertia weight w unchanged, and takes c 1 ＝c 2 =1, learning ability to balance the particle's own cognition and social cognition;
(3) if it isThen the particle is the worse particle in the group, which is divided into the global optimizing group, and the larger w is selected to increase the variation amplitude and strengthenGlobal search capability of population of particles, while taking c to be smaller 1 And a larger c 2 And strengthening the social learning ability of the particles, and adjusting the strategy according to the following formula:
in the step, w is the inertia weight of the particle swarm algorithm and is used for balancing the global search capability and the local search capability of the algorithm, larger w can enhance the global search capability of the algorithm, and smaller w can enhance the local convergence capability of the algorithm and accelerate the convergence speed of the algorithm. The value of w is usually in the range of [0.4,1.5].
In the formula, c 1 And c 2 A value range of a learning factor of the algorithm is set [ c 1min ,c 1max ]The larger corresponding interval [ (c) 1min +c 1max )/2,c 1max ]Smaller, i.e. corresponding to the interval [ c 1min ,(c 1min +c 1max )/2]And the specific values are calculated according to corresponding calculation formulas under the three different particle classification conditions in the step (7.3).
In the formula, c 1max And c 2max Are respectively cognitive learning factors c 1 And social learning factor c 2 Maximum value of the value range set, c 1min And c 2min Is the minimum value of the value range; w is an initial set value of the inertia weight, w min For setting the minimum value of the inertial weight, k 1 、k 2 Is an adjustment factor, k, of the inertial weight w 1 Determining an upper search limit, k, for the inertial weight w 2 For controlling the convergence speed of the inertial weight w.
Has the advantages that: the invention combines the NARMAX model idea with a support vector regression modeling method, and simultaneously realizes the integral optimization of model parameters by using a self-adaptive particle swarm algorithm, compared with the prior art, the method has the following advantages:
(1) The invention considers the order of the input and output variables of the soft measurement model, establishes the dynamic soft measurement model of the nitrogen oxide, can realize the soft measurement of the nitrogen oxide in the dynamic change process of the unit variable load and the like, and effectively improves the soft measurement precision;
(2) The invention adopts the self-adaptive particle swarm algorithm to integrally optimize the order of the input and output variables and the model parameters;
(3) The invention can directly utilize the historical operation data of the measuring points read from the power plant DCS system, does not need field test, is easy for engineering application, has low cost and reliable prediction result.
Drawings
FIG. 1 is a schematic diagram of the dynamic soft measurement principle of the present invention;
FIG. 2 is a flow chart of an embodiment;
FIG. 3 is a diagram of the predicted effect of the embodiment.
Detailed Description
The technical solution of the present invention is described in detail below, but the scope of the present invention is not limited to the embodiments.
Example (b): the dynamic soft measurement method based on the nonlinear autoregressive moving average and the support vector regression is adopted to realize NO measurement on the four-corner tangential supercritical direct current boiler of a 600MW coal-fired power generating unit of a certain power plant as a research object x Dynamic soft measurement of emissions.
As shown in fig. 2, the specific steps of this embodiment are as follows:
(1) Data acquisition and preprocessingTreating: NO for power station boiler combustion system x The emission is dynamically measured in a soft manner, and the model output y (t) is selected as NO before the SCR reactor at the tail of the furnace x The NO of the boiler is measured by the inlet concentration, model input variable selection condition parameters (load and coal feeding quantity), main combustion zone parameters (primary air door opening degree and secondary air door opening degree), burnout zone parameters (burnout air door opening degree), state parameters (oxygen quantity) and the like x The operation variables of the boiler combustion system with the influence of emission are four input variables in total; selecting data in a variable load regulation process as a training sample, wherein the load change range is 364.7 MW-554.4 MW, the duration is 62min, the sampling period is set to be 20s, and 186 groups of data are collected; the original value of each input data is scaled to the interval [0,1 ] by the following formula using normalization processing]And measuring the NO measured under each working condition x Converting the concentration to a value below 6% oxygen;
(2) Initializing a self-adaptive particle swarm algorithm: setting the population size of a particle swarm S =50 and the maximum iteration number K max =1000; each particle p (i) contains 7 variables in total, 4 of which are model input orders d i 1 output order d y And 2 SVMR model parameters, setting the optimization range of the time delay order as [1,20 ]](i.e. [20s, 400s)]) The penalty factor C is taken as [1,2000 ]]The width of the nucleus σ is [0.01,1000]Randomly generating an initial population of particles within the interval;
(3) The input data and the output data are arranged into a NARMAX model structure: order value d contained in input data and output data as p (i) i And d y Arranging the soft measurement model into an NARMAX model structure, and calculating to obtain the input X (t) and the output Y (t) of the soft measurement model as follows:
taking X (t) as SVMR model input and Y (t) as SVMR model output, the structure of the finally obtained soft measurement model is shown in figure 1, wherein X is X l_c (t-T),x l_c (t-2T),...,x l_c (t-d l_c T) represents past d l_c The load in each time interval andcoal feed as input for the current model, d l_c Representing the order corresponding to the load and the coal feeding quantity as input variables of the soft measurement model; x is a radical of a fluorine atom a_f (t-T),x a_f (t-2T),...,x a_f (t-d a_f T)、 x sofa (t-T),x sofa (t-2T),...,x sofa (t-d sofa T) andrespectively representing the opening degree of the primary and secondary air doors, the opening degree of the burnout air door and the oxygen amount in the past periods as input quantities of a model, d a_f 、d sofa Andrespectively representing the corresponding order, y, as input to the soft-metric model NOx (t-T),y NOx (t-2T),...,y NOx (t-d y T) is past d y The output y at each time is used as the input quantity of the model, d y Is the order of the output variable;
(4) Initializing the SVMR model: setting an SVMR model according to a penalty factor C and a kernel width sigma contained in p (i), and establishing a particle p (i) corresponding to X (t) and Y (t) as SVMR model input and output
SVMR-NARMAX model; wherein the kernel function is selected as a Gaussian kernel function (RBF), and the model form is as follows:
(5) Carrying out SVMR training and fitting, calculating individual fitness of the particles, and reserving optimal individuals: according to the established model, fitting of training samples and prediction of test samples are carried out, a Mean Square Error (MSE) evaluation index is used as a particle fitness function, and the following calculation is carried out:
wherein n is the length of the prediction time seriesDegree, y j Andrespectively an actual measurement value and a model predicted value at the jth moment; selecting min (f) i k ) I.e. the individual with the smallest prediction deviation as the optimal individual f of the population m Reserving;
(6) Optimizing convergence judgment: if f m <f e Or K = K max In which f is e If the convergence threshold is desired, the optimization is terminated, and step 9 is executed; otherwise, let k = k +1, execute step 7;
(7) Calculating the average fitness of the particle group, and dividing the particle group into three subgroups: the fitness value of the particle p (i) in the kth iteration is f i k The fitness value of the population-optimal particle is f m (ii) a Particle population mean adaptation value ofMake the fitness value better thanIs obtained by averaging the particle fitness valueDefinition ofTo evaluate the degree of premature convergence of the population, a smaller Δ indicates a tendency of the population toward premature convergence. The particles are classified into the following three categories according to their individual fitness values:
(1) if it isThe particles are the better particles in the population, divided into local optimization sub-groups, and the smaller w is taken to change in a small range and the larger c is taken 1 And smaller c 2 To enhance the learning ability of the particle individuals and strengthen the local optimizing ability of the particles
(2) If it isThe particles are common particles in a group and are divided into balanced optimizing groups, and meanwhile, the particles have good global optimizing capacity and local optimizing capacity, w is maintained to be unchanged, and c is taken 1 ＝c 2 =1, learning ability to balance particle self-cognition and social cognition;
(3) if it isThe particles are poor particles in the population, are divided into global optimizing subgroups, and the larger w is selected to increase the variation amplitude, strengthen the global searching capability of the particle population, and simultaneously, the smaller c is selected 1 And a larger c 2 And strengthening the social learning ability of the particles, and adjusting the strategy according to the following formula:
wherein is obtainedc 1max ＝c 2max ＝1.8，c 1min ＝c 2min ＝0.2，k 1 ＝1.5，k 2 ＝0.3；
(8) Particle individual velocity and location update: obtaining the inertia weight w and the learning factor c corresponding to each particle according to the calculation result of (7) 1 、c 2 The velocity and position of the individual particles are updated according to the following formula:
v in ＝wv in +c 1 rand()(b in -p in )+c 2 rand()(b gn -p in ) (12)
p in ＝p in +v in (13)
(9) Outputting an optimal result, and finishing the establishment of a dynamic soft measurement model: output f m D contained in the corresponding optimum individual i 、d y The penalty factor C and the nuclear parameter sigma are shown in the following table 1, and the parameters are used as setting parameters to establish a final dynamic soft measurement SVMR model;
TABLE 1
(10) And (3) realizing dynamic soft measurement: and (4) preprocessing any newly acquired sample, and inputting the dynamic soft measurement SVMR model obtained in the step (9) to obtain corresponding soft measurement output.
In this embodiment, the front 140 groups are training samples, and the rear 40 groups are testing samples, as shown in FIG. 3, the present invention can be used to test NO in the dynamic operation process of the boiler x Accurate tracking and prediction of emission change to realize NO x Dynamic soft measurement of emissions.
Claims (3)
1. NO of power station boiler x An emission dynamic soft measurement method, characterized by: the method comprises the following steps:
(1) Data acquisition and pretreatment: determining model input variables and output variables fromCollecting original samples in DCSFor original sampleAdopting a normalization pretreatment method and measuring NO under various working conditions x Converting the concentration to a value below 6% oxygen; wherein x is i ∈R m Representing the input samples of the ith group of models, m being the number of input variables, y i E.g. R represents the output samples of the ith group of models, and n is the number of samples;
(2) Initializing a self-adaptive particle swarm algorithm: each particle p (i) contains m +3 variables, including m input variables x i Corresponding order d i Order d corresponding to 1 output variable y y And a penalty factor C and a nuclear parameter sigma of the SVMR model of the support vector regression;
(3) Arranging input data and output data into a nonlinear autoregressive moving average model framework NARMAX model structure: input variable x contained by particle p (i) i Order d of i And order d of the output variable y y The input data and the output data are arranged into a NARMAX model structure, and then the input X (t) and the output Y (t) of the soft measurement model are obtained through calculation as follows:
in the above formula, T refers to a sampling period for collecting the measurement point data; x (t) is a model input quantity corresponding to the sampling time t and comprises m input variables X i Consecutive d before the sampling instant t i The state quantity of one sampling period, and the succession d of measured quantities y before the sampling instant t y A state quantity of each sampling period; y (t) is a model output quantity corresponding to the sampling time t, and is a state quantity of the measured Y at the sampling time t;
(4) Initializing SVMR model: the SVMR model is set up with a penalty factor C and a kernel parameter σ contained by the particle p (i) using a gaussian kernel function, RBF:
in the above-mentioned formula, the compound has the following structure,α i and b is a support vector regression machine parameter,is a support vector, l is the number of the support vectors, and sigma is an RBF nuclear parameter;
(5) Carrying out SVMR training and fitting, and calculating individual fitness f of particles i And keeping the optimal individuals: taking X (t) as SVMR model input, Y (t) as SVMR model output, taking the first 3/4 of the sorted sample set as training samples, taking the second 1/4 as prediction samples, and carrying out model training and fitting; calculating the individual fitness f of the particles by taking the deviation index of the prediction sample as a fitness function i And obtaining the optimal fitness f of the population m And its particle position Pm is preserved, the fitness function of the particle p (i) is:
in the above formula, n is the predicted time series length, y j Andrespectively an actual measurement value and a model prediction value at the jth moment;
(6) Optimizing convergence judgment: if f m <f e Wherein f is e To expect a convergence threshold, when K = K max If so, the optimization is terminated, and the step (9) is executed; otherwise, let k = k +1, perform step (7); wherein K is the current iteration number, K max For maximum number of iterations, the optimization range of the order is [ d ] min ,d max ]；
(7) Calculating the average fitness of the particle population, and dividing the particle population into three subgroups: the fitness value of the particle p (i) in the kth iteration is f i k The fitness value of the population-optimal particle is f m (ii) a The average fitness of the particle swarm isTo make the fitness value better thanIs averaged to obtainAccording to the individual fitness f of the particles i k Dividing the data into three subgroups of local optimization, balance optimization and global optimization, and respectively calculating corresponding inertia weight w and learning factor c 1 、c 2 ；
Wherein, the fitness f is adapted according to the individual particle i k The process of dividing the particle population into three subgroups was:
(7.1) calculating the mean fitness of the population of particles, and then setting the fitness value of the particle p (i) in the kth iteration to f i k The fitness value of the population-optimal particle is f m Mean fitness value of population of particles of
(7.2) to outperform the fitness valueIs obtained by averaging the fitness of the particlesDefinition ofFor evaluating precocity of a particle populationThe smaller the Delta is, the more the particle population tends to be precocious and convergent;
(7.3) finally, according to the individual fitness f of the particles i k It is classified into the following three categories:
(1) if it isThen the particle is the better particle in the group, and is divided into local optimizing groups, and the smaller inertia weight w is taken according to the following formula, so that the inertia weight w varies in a small range, and the larger c is taken at the same time 1 And smaller c 2 To enhance the learning ability of the particle individuals and strengthen the local optimizing ability of the particles:
(2) if it isThe particle is shown as a common particle in the group, is divided into a balanced optimizing subgroup, has good global optimizing capability and local optimizing capability at the same time, maintains the inertia weight w unchanged, and takes c 1 ＝c 2 =1, learning ability to balance particle self-cognition and social cognition;
(3) if it isThe particles are the poor particles in the group, the particles are divided into global optimizing groups, and the larger w is selected to increase the variation amplitude, strengthen the global searching capability of the particle group, and simultaneously, the particles are selected to be betterSmall c 1 And a larger c 2 And strengthening the social learning ability of the particles, and adjusting the strategy according to the following formula:
wherein, c 1max And c 2max Are respectively cognitive learning factor c 1 And social learning factor c 2 Maximum value of the set value range, c 1min And c 2min Is the minimum value of the value range; w is a 0 Is an initial setting value of the inertia weight, w min For setting the minimum value of the inertial weight, k 1 、k 2 Is an adjustment factor, k, of the inertial weight w 1 Determining the search ceiling, k, of the inertial weight w 2 For controlling the convergence speed of the inertial weight w;
(8) Particle individual velocity and location update: according to the inertia weight w and the learning factor c obtained in the step (7) 1 、c 2 For individual velocity V of the particles i And position P i Performing calculation updating;
(9) Outputting an optimal result, and completing the establishment of a dynamic soft measurement model: output f m D contained in the corresponding optimum individual i 、d y The penalty factor C and the nuclear parameter sigma are used as setting parameters to establish a final dynamic soft measurement SVMR model;
(10) And (3) realizing dynamic soft measurement: and (4) preprocessing any newly acquired sample, and inputting the dynamic soft measurement SVMR model obtained in the step (9) to obtain corresponding soft measurement output.
2. The utility boiler NO of claim 1 x An emission dynamic soft measurement method, characterized by: in the step (1), NO in front of the SCR reactor at the tail of the boiler furnace of the power station is selected x Inlet concentration as output y i Selecting NO in the dynamic operation process of the boiler unit x Combustion system operating variable with influence of emission as input variable x i (ii) a While the original sampleThe dynamic continuous operation working condition data which is large in coverage range and operates in a variable load mode in the DCS.
3. The utility boiler NO of claim 1 x An emission dynamic soft measurement method, characterized by: the step (1) is to use the original sampleThe specific process adopting the normalization pretreatment comprises the following steps:
scaling the original value of each input data to within a specified interval according to the following formula:
wherein, x is a value before normalization, x' is a value after normalization, and max (x) and min (x) are respectively a maximum value and a minimum value of an interval after normalization;
because the oxygen content of the flue gas is inconsistent under various working conditions, NO is convenient to compare x The value of NO measured in each operating condition is determined according to the following formula x The concentration is converted to a value at 6% oxygen:
wherein ρ' (NO) x ) And ρ' (O) 2 ) Respectively representing converted NO x Concentration and O 2 Concentration, ρ (NO) x ) And ρ (O) 2 ) Respectively representing actually measured NO x Concentration and O 2 And (4) concentration.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510064480.2A CN104715142B (en) | 2015-02-06 | 2015-02-06 | A kind of station boiler NOxDischarge dynamic soft-measuring method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510064480.2A CN104715142B (en) | 2015-02-06 | 2015-02-06 | A kind of station boiler NOxDischarge dynamic soft-measuring method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104715142A CN104715142A (en) | 2015-06-17 |
| CN104715142B true CN104715142B (en) | 2018-02-02 |
Family
ID=53414465
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510064480.2A Active CN104715142B (en) | 2015-02-06 | 2015-02-06 | A kind of station boiler NOxDischarge dynamic soft-measuring method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104715142B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105224730A (en) * | 2015-09-17 | 2016-01-06 | 西南林业大学 | The original NO of a kind of high pressure common rail electric-controlled diesel engine 2forecasting of discharged quantity method |
| CN105605610A (en) * | 2016-01-13 | 2016-05-25 | 上海交通大学 | Boiler combustion optimization method based on particle swarm algorithm |
| CN106527176A (en) * | 2016-10-21 | 2017-03-22 | 哈尔滨理工大学 | MFOA (modified fruit fly optimization algorithm)-SVM (support vector machine)-based boiler thermal efficiency and NOX modeling method |
| CN106991507A (en) * | 2017-05-19 | 2017-07-28 | 杭州意能电力技术有限公司 | A kind of SCR inlet NOx concentration on-line prediction method and device |
| CN107657104A (en) * | 2017-09-20 | 2018-02-02 | 浙江浙能台州第二发电有限责任公司 | Boiler combustion system dynamic modelling method based on Online SVM |
| CN108319146B (en) * | 2018-03-09 | 2020-08-11 | 西安西热控制技术有限公司 | Discrete particle swarm training-based method for radial basis function neural network |
| CN110986085B (en) * | 2019-11-21 | 2021-07-27 | 东南大学 | Four-corner tangential boiler combustion optimization control method based on air distribution mode optimization |
| CN111413098B (en) * | 2020-01-13 | 2021-06-01 | 哈尔滨工程大学 | Natural gas engine combustion system nonlinear feature identification method based on IMEP time sequence |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101882239A (en) * | 2010-05-28 | 2010-11-10 | 华东理工大学 | Ethylene cracking severity modeling method based on expert knowledge and neutral network |
| CN101699229B (en) * | 2009-10-28 | 2011-09-07 | 东南大学 | On-line soft sensing method for flow of inlet water of ozonation contact reactor in water plant |
| CN102778538B (en) * | 2012-07-06 | 2014-12-10 | 广东电网公司电力科学研究院 | Soft measuring method based on improved SVM (Support Vector Machine) for measuring boiler unburned carbon content in fly ash |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7006900B2 (en) * | 2002-11-14 | 2006-02-28 | Asm International N.V. | Hybrid cascade model-based predictive control system |
-
2015
- 2015-02-06 CN CN201510064480.2A patent/CN104715142B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101699229B (en) * | 2009-10-28 | 2011-09-07 | 东南大学 | On-line soft sensing method for flow of inlet water of ozonation contact reactor in water plant |
| CN101882239A (en) * | 2010-05-28 | 2010-11-10 | 华东理工大学 | Ethylene cracking severity modeling method based on expert knowledge and neutral network |
| CN102778538B (en) * | 2012-07-06 | 2014-12-10 | 广东电网公司电力科学研究院 | Soft measuring method based on improved SVM (Support Vector Machine) for measuring boiler unburned carbon content in fly ash |
Non-Patent Citations (16)
| Title |
|---|
| Dynamic soft sensors for NOx emissions in a circulating fluidized bed boiler;Liukkonen, Mika,et al.,;《 ICAE 2011 - International Conference on Applied Energy》;20110518;483-490 * |
| 一种云自适应粒子群优化的模糊PID控制器设计;应明峰，等;《计算机测量与控制》;20131231;第21卷(第12期);3278-3281 * |
| 一种动态改变惯性权重的自适应粒子群算法;任子晖，等;《计算机科学》;20090228;第36卷(第2期);227-230 * |
| 化工过程软测量建模方法研究进展;曹鹏飞，等;《化工学报》;20130331;第63卷(第3期);788-800 * |
| 基于云模型改进的粒子群K均值聚类算法;杨书佺，等，;《现代计算机》;20140430(第04期);15-19 * |
| 基于多关联向量机的动态软测量建模;李川，等;《系统仿真学报》;20090630;第21卷(第12期);3513－3517 * |
| 基于支持向量机的燃烧飞灰含碳量与NOx排放特性建模;成海南，等;《能源研究与利用》;20090131(第1期);1－4 * |
| 基于改进粒子群算法的电站锅炉NOx排放预测控制及优化;冯磊华，等;《中南大学学报（自然科学版）》;20110731;第42卷(第7期);2018－2022 * |
| 基于最小二乘支持向量机的电站锅炉燃烧优化;顾燕萍，等，;《中国电机工程学报》;20100615;第30卷(第17期);91-97 * |
| 基于粒子蜂群算法优化的多支持向量机软测量建模方法研究;李广宇;《中国优秀硕士学位论文全文数据库信息科技辑（月刊）》;20120515(第05期);I140-43 * |
| 基于遗传算法的燃煤电站锅炉整体燃烧优化方法研究;孙巧玲，等;《热能动力工程》;20040131;第19卷(第1期);85－89 * |
| 火电厂热工参数软测量关键技术和方法研究;乔弘，;《中国博士学位论文全文数据库工程科技II辑（月刊）》;20091015(第10期);C042-17 * |
| 电站锅炉烟气含氧量软测量方法研究;张捷夫;《中国优秀硕士学位论文全文数据库工程科技II辑（月刊）》;20150115(第01期);C042-720 * |
| 电站锅炉燃烧系统软测量及稳态优化研究;崔宇;《中国优秀硕士学位论文全文数据库工程科技II辑（月刊）》;20100715(第07期);C042-103 * |
| 自适应混沌粒子群优化算法;赵志刚，等;《计算机工程》;20110831;第37卷(第15期);128-130 * |
| 非线性NARMAX模型结构与参数一体化辨识的改进算法;王晓，等;《信息与控制》;20000430;第29卷(第2期);102－110 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104715142A (en) | 2015-06-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104715142B (en) | A kind of station boiler NOxDischarge dynamic soft-measuring method | |
| CN107526292B (en) | A method of the regulation ammonia spraying amount based on inlet NOx concentration prediction | |
| CN104534507B (en) | A kind of boiler combustion optimization control method | |
| CN110597070B (en) | Method for identifying model parameters of thermal power generating unit system | |
| CN112613237B (en) | CFB unit NOx emission concentration prediction method based on LSTM | |
| CN112488145B (en) | NO based on intelligent methodxOnline prediction method and system | |
| CN110263395A (en) | The power plant's denitration running optimizatin method and system analyzed based on numerical simulation and data | |
| CN112465239B (en) | Desulfurization system operation optimization method based on improved PSO-FCM algorithm | |
| CN109670625A (en) | NOx emission concentration prediction method based on Unscented kalman filtering least square method supporting vector machine | |
| CN113440990A (en) | EMD-LSTM based outlet SO2Concentration prediction method | |
| CN106991507A (en) | A kind of SCR inlet NOx concentration on-line prediction method and device | |
| Hao et al. | Prediction of nitrogen oxide emission concentration in cement production process: a method of deep belief network with clustering and time series | |
| CN115145152A (en) | Boiler combustion and denitration process collaborative optimization control method | |
| CN104331736A (en) | RBF (Radial Basis Function) neural network based supercritical boiler nitric oxide discharging dynamic predication method | |
| Liu et al. | Temporal hypergraph attention network for silicon content prediction in blast furnace | |
| CN108647483A (en) | A kind of SCR inlet NO based on fuzzy tree modeling methodXThe flexible measurement method of concentration | |
| CN114815616A (en) | Intelligent regulation and control method and system for mode conversion of turbine stamping combined type engine | |
| CN111199304A (en) | Multi-target combustion optimization method based on data-driven fusion strategy | |
| CN112836902A (en) | Coal combustion calorific capacity prediction method based on improved BP neural network | |
| CN114740713B (en) | Multi-objective optimization control method for wet flue gas desulfurization process | |
| CN113488111B (en) | Ammonia injection amount optimization modeling method for SCR denitration system | |
| CN110647560B (en) | Mechanism and brain-like intelligence combined online soft measurement method for coal quality of fired coal of power plant | |
| Liu et al. | A prediction method of NOx in thermal power plants using GC-LSTM neural network | |
| CN113935230A (en) | Implementation of NO based on attention mechanism LSTM modelxEmission amount prediction method | |
| Wang et al. | Data-driven multi-objective intelligent optimal control of municipal solid waste incineration process |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |