CN104776446B - Combustion optimization control method for boiler - Google Patents

Abstract

The invention discloses a combustion optimization control method for a boiler. The combustion optimization control method is characterized by comprising the following steps: sampling a combustion nonlinear system of the boiler to obtain input/output data at the current moment; training the real-time sampled input/output data by an online incremental learning fuzzy neural network, building an online incremental learning predicting model of the combustion nonlinear system of the boiler; performing a nonlinear prediction control algorithm on the online incremental learning predicting model for realizing the optimization and the control of the combustion process of the boiler. According to the combustion optimization control method for the power station boiler of the online incremental learning fuzzy neural network, the nonlinear optimization problem in the predication control algorithm is solved by utilizing a particle swarm optimization algorithm through the online identification of the boiler combustion optimization model; the real-time optimization and control of the boiler combustion process are realized.

Description

A kind of boiler combustion optimization control method

Technical field

The present invention relates to a kind of optimization running technology of power boiler burning system, more particularly to a kind of online incremental learning The power boiler burning optimization control method of fuzzy neural network, belongs to thermal technics technical field.

Background technology

Burning optimization is to lift Utility Boiler Efficiency, reduce the important means of pollutant emission.Current burning optimization skill Boiler combustion efficiency and NO are set up offline using learning algorithms such as neutral net or SVMs more than art_xDischarge model, to combustion Burn optimization target values to be optimized using intelligent search algorithms such as genetic algorithms, obtain the manipulation variable of boiler combustion optimization.

Research shows, boiler has very big time-varying characteristics, As time goes on and boiler operatiopn operating mode change, The learning model of boiler combustion process will occur larger error, and the offline model set up does not adapt to this change, so as to lead Cause model mismatch.The on-line correction of model needs the longer calculating time so that the process for obtaining manipulation variable is complicated, power station pot The property regulation real-time of stove is not high, affects the performance of burning optimization.

The features such as the non-linear of boiler combustion process, large time delay, bring to boiler combustion optimization control work certain It is difficult.Model Predictive Control (MPC) is substantially control method of the class based on optimization, and it is based on forecast model, rolling optimization With the big feature of feedback compensation three, the control problem of object with big lag is can effectively solve the problem that.The precision and rolling optimization of forecast model Strategy is the key for affecting MPC performances.The solution of wherein non-linear rolling optimization is difficult to ask for, and typically can only be asked using numerical value optimizing Solution.

Therefore, it is necessary to find a kind of new method to complete the Nonlinear Model Predictive Control of power boiler burning optimization Problem, solves two big difficult points mentioned above, i.e., set up accurate Nonlinear Prediction Models in real time and obtain rolling online Optimum control amount under time domain object function.

The content of the invention

Goal of the invention：For the problem and shortage that above-mentioned prior art is present, it is an object of the invention to provide a kind of online The power boiler burning optimization control method of incremental learning fuzzy neural network, can with on-line identification boiler combustion process model, Using the nonlinear optimal problem in PSO Algorithm PREDICTIVE CONTROL, so as to improve the real-time of boiler combustion optimization control Property.

Technical scheme：For achieving the above object, the technical solution used in the present invention is：

A kind of boiler combustion optimization control method, it is characterised in that comprise the steps：

(1) boiler combustion nonlinear system is sampled, obtains the input/output data at current time；

(2) input/output data that real-time sampling is obtained is trained using online incremental learning fuzzy neural network, Set up the online incremental learning forecast model of boiler combustion nonlinear system；

(3) nonlinear Model Predictive is used the online incremental learning forecast model, is realized to boiler combustion The optimal control of journey.

In step (1), the input data is boiler operatiopn operating parameter, and the output data is boiler efficiency and cigarette Gas discharges NO_x。

The boiler operatiopn operating parameter stated includes that load, coal-supplying amount, total air, fuel throttle opening, secondary air register are opened Degree and after-flame throttle opening.

In step (2), the online incremental learning forecast model is：

Wherein, u (k)=(u₁(k),u₂(k),…,u_m(k)) represent Boiler Combustion Optimization System controlled quentity controlled variable, y (k)= (y₁(k),y₂(k),…,y_n(k)) target output of boiler combustion status is represented, d represents that the output of boiler combustion system is prolonged Late, p and q represent the input/output order of boiler combustion process nonlinear system.

In step (2), totally four layers of the online incremental learning structure of fuzzy neural network：

Input layer, each neuron in this layer represents an input variable of online incremental learning forecast model, wherein Use X₁, X₂..., X_rRepresent that boiler respectively runs manipulation amount u (k) and associated front p orders output y (k) successively；

Membership function layer, each input variable X_iThere is u membership function A_ij(j=1,2 ..., u), it is Gauss and is subordinate to letter Number：

Wherein μ_ijIt is x_iJ-th membership function, c_ijAnd σ_ijRespectively x_iJ-th Gaussian function center and width Degree, u is the quantity of membership function；

Fuzzy rule layer, j-th rule R_jOutput be：

Output layer, the output variable of each one input signal weighted sum of node on behalf：

Wherein y be characterize boiler combustion status optimization aim output valve, w_jFor result parameter.

In step (3), the method for the optimal control is：

(31) correction of boiler combustion nonlinear system output valve is obtained by the online incremental learning forecast model：

In current sample time k, by the past input/output of boiler combustion nonlinear system and current input u (k) by Built in line incremental learning forecast model obtains the output estimation value of boiler combustion nonlinear system

By boiler combustion nonlinear system input u (k+1) to be optimized and past input/output, boiler combustion is obtained Burn the output estimation value of nonlinear system

If the prediction deviation at k moment isUse drift correctionRepaiied Positive quantity

(32) input to boiler combustion expense linear system is optimized：

Determine that the object function of boiler combustion optimization controlled quentity controlled variable u is according to boiler combustion optimization controlled quentity controlled variable u：

Wherein y_irFor the reference locus of i-th boiler combustion significant condition output quantity, by solving boiler combustion optimization Economic goal function and obtain, y_ipIt is defeated for prediction of the corresponding i-th boiler combustion significant condition output quantity Jing after feedback compensation Go out, the dimension that m and n is respectively input into and exports, q_iAnd λ_jFor weight coefficient；

The minimum of a value of above-mentioned object function is obtained by the online rolling optimization in real time of particle cluster algorithm, optimum control amount is obtained U (k+1), acts on optimum control amount u (k+1) boiler combustion nonlinear system and is optimized control.

Beneficial effect：Compared with prior art, the present invention has advantages below, online incremental learning fuzzy neural network mould Type can be according to the time-varying characteristics of power boiler burning process nonlinear system, the structure and parameter of on-line tuning model, identification Process is simple, and adjustable parameter is few, and generalization ability is strong；Nonlinear Model Predictive Control (MPC) is optimized to boiler combustion process And control, can effectively solve the problem that the large delay characteristic of boiler combustion process, using PSO Algorithm PREDICTIVE CONTROL in it is non- Linear optimization problem, online rolling optimization in real time determines controlled quentity controlled variable, has preferable control effect to boiler combustion process.

Description of the drawings

Fig. 1 is that the boiler combustion optimization Control system architecture of the online incremental learning fuzzy neural network of the present invention is illustrated Figure.

Fig. 2 is the online increment fuzzy neural network model schematic diagram of the present invention.

Fig. 3 is the algorithm flow chart of the online increment fuzzy neural network model of the present invention.

Specific embodiment

Below in conjunction with the accompanying drawings, further elucidate the present invention, it should be understood that these embodiments be merely to illustrate the present invention and without In the scope of the present invention is limited, after the present invention has been read, those skilled in the art are to the various equivalent form of values of the invention Modification falls within the application claims limited range.

The power boiler burning optimization control method of a kind of online incremental learning fuzzy neural network that the present invention is provided, leads to Cross carries out real-time sampling to boiler combustion nonlinear system, using online incremental learning fuzzy neural network to real-time sampling data It is trained, sets up boiler combustion optimization data-driven model, it is pre- using nonlinear model to the boiler combustion optimization model Observing and controlling system (MPC) algorithm is optimized and controls.The power boiler burning optimization control method is comprised the following steps that：

(1) boiler combustion process nonlinear system is sampled, obtains the input/output data at current time；

(2) real-time sampling is trained using online incremental learning fuzzy neural network, sets up boiler combustion optimization pre- Survey model；

(3) to the model, boiler combustion process is optimized using Nonlinear Model Predictive Control (MPC) algorithm and Control.

For using boiler structure, burner arrangement form the features such as choose suitable boiler operatiopn operating parameter and make For the input quantity of combustion model, boiler efficiency and environment protection emission NO_xFor optimization aim output quantity, so as to obtain boiler combustion system The real-time input/output data of system model；

Described real-time boiler operatiopn operating parameter includes load, coal-supplying amount, total air, fuel throttle opening, secondary Throttle opening, after-flame throttle opening；

Boiler efficiency and flue gas NO are set up using a kind of online incremental learning fuzzy neural network_xModel.For building The |input paramete of mould and the output parameter of sign boiler combustion status are expressed asWherein x_kRepresent kth group as defeated Enter the boiler operating parameter vector of data, y_k+dRepresent that kth group characterizes the vector of boiler combustion status as output parameter, d is represented The output of boiler combustion system postpones.Such as the online incremental learning fuzzy neural network model schematic diagram that Fig. 2 is the present invention, model Structure has four layers：

1. input layer, each neuron in this layer represents input variable X of boiler combustion system model_i(i=1, 2 ..., r), r is input variable number；

2. membership function layer, each input variable X_iThere is u membership function A_ij(j=1,2 ..., u), it is Gauss and is subordinate to Function：

Wherein μ_ijIt is x_iJ-th membership function, c_ijAnd σ_ijThe center of respectively j-th Gaussian function and width, u It is the quantity of membership function；

3. fuzzy rule layer, the node layer number has reacted number of fuzzy rules, for calculating T- models of each rule triggering power Son is figured for multiplication, j-th rule R_jOutput be：

4. output layer, the output variable of each one input signal weighted sum of node on behalf：

Wherein y be characterize boiler combustion status optimization aim output valve, w_jFor result parameter.

As shown in figure 3, online incremental learning fuzzy neural network algorithm is broadly divided into 4 parts：The generation of neuron, premise Parameter Estimation, weight adjustment and tailoring technique.Model during iteration, according to the system of setting miss by computation model error Difference parameter k_eProduce neuron；In addition, model can also calculate the distance for being input to the neuronal center for having existed, according to setting Boundary sizes k_dProduce neuron.The k of setting_eAnd k_dAlso dynamic can adjust during iterative calculation.Neuron is produced Afterwards, the center c of its member function_ijAnd width cs_jSize give an initial value, in iterative calculation below, can according to k_d And k_eSize is being accordingly adjusted.Weight adjustment refers to weight w of neuron output layer_kSize it is true by linear least square It is fixed.Tailoring technique is mainly model can calculate importance index of the neuron to system, according to importance come dynamic cutting footpath Can be adjusted according to input data dynamic to base unit, therefore model structure, will not be excessively complicated.

The fuzzy rule of network model is gradually changed from scratch, is decided whether to increase according to the sample of training Or reduce by a rule, and set response parameter.The following detailed description of algorithmic procedure, detailed algorithm flow chart such as Fig. 3：

1) initialize and predefine initial parameter ε_min, ε_max, e_min, e_max, k_mf, k_s, k_eer, wherein ε_minAnd ε_maxIt is fuzzy The setting value of the minimum and maximum of regular perfect set, e_minIt is preferable Accuracy Error, e_maxIt is maximum error, k_eAnd k_dFor pre- The threshold value being associated with e and ε respectively is first set, it can be with dynamic adjustment, k in learning process_mfIt is the adjacent membership function of control The constant of similarity, k_eerIt is default constant for regular importance；

2) first group of sample data (x₁, y₁) obtain after, produce first fuzzy rule, parameter is as follows：

c₁=x₁,σ₁=σ₀,ω₁=ω₀, wherein ω₀,σ₀For predefined parameter；

3) from the beginning of second group of sample, to each group of new samples (x_k, y_k) calculate md_kJ (), it is observation data and j-th strip The mahalanobis distance of the central point of fuzzy rule, finds out md_k,min=md_k(J),And computing system is missed Difference

4) md is worked as_k,min>k_d, e_k>k_eMono- fuzzy rule of Shi Zengjia, it is assumed that u fuzzy rule has been produced is new to produce rule Initial parameter then is allocated according to the following rules：Multidimensional input variable x_kProject to corresponding one-dimensional membership function empty Between, calculate dataWith boundary set Φ_iEuclidean distance between (j)Wherein Φ_i∈{x_imin,c_i1, c_i2,...,c_iu,x_imax, while findingIf ed_i(j_n)≤k_mf, thenIt is not used in the dimension to produce The new membership function of life.Otherwise distribute a new Gaussian function, the width of Gaussian function and center are respectively：

5) after producing new rule, importance η of calculation error reduction rate and j-th strip rule_j(j=1,2 ..., u), error Reduce rate matrix Δ=(ρ₁,ρ₂,...,ρ_u), wherein jth arranges (r+1) individual error slip of j-th rule of correspondence.If η_j<k_err, then j-th rule is deleted；

6) if new rule need not be increased, condition md is met_k,min<k_d, e_k>k_e, then input variable x_iMembership function Width cs_ijIt is modified to σ_new,ij=ζ * σ_ij, wherein ζ is decay factor；

7) result parameter is adjusted, and n observation data sample of r input variable produces u fuzzy rule, its matrix form For W φ=Y.Optimized parameter W^*Determination be formulated as minimize | | W φ-T | |₂Linear problem, T be desired output (i.e. Sample is exported).W is determined using Generalized Inverse Method^*For：W^*=T (φ^Tφ)^-1φ^T。

8) see whether to complete learning process, if the return to step 3 without if), otherwise terminate whole learning process.

It is to the model that the nonlinear system of boiler combustion process is set up using online incremental learning fuzzy neural network：

Wherein, u (k)=(u₁(k),u₂(k),…,u_m(k)) represent Boiler Combustion Optimization System controlled quentity controlled variable, y (k)= (y₁(k),y₂(k),…,y_n(k)) output quantity of boiler combustion status is represented, p and q represents boiler combustion process nonlinear system Input/output order.

Determine that the output of boiler combustion nonlinear system postpones d by the fitting precision of off-line model, in present sample K is carved, the output of system is obtained by built forecast model by the past input/output of boiler combustion system and current input u (k) EstimateBy system input u (k+1) to be optimized and past input/output, the output estimation of system is obtained ValueDue to the reasons such as noise jamming or model mismatch, forecast model outputWith reality output y (k+d) Deviation is commonly present, if the prediction deviation at k moment isUse drift correctionObtain Obtain correction

Boiler combustion process model is nonlinearity, then for solving the optimization problem of PREDICTIVE CONTROL signal sequence It is Solution of Nonlinear Optimal Problem.Shorter control signal sequence can reduce the complexity of optimization problem, increase control Robustness.Using one-step prediction control signal, the object function for determining boiler combustion optimization controlled quentity controlled variable u is：

Wherein y_irFor the reference locus of i-th boiler combustion significant condition output quantity, boiler combustion is solved by off-line model Burn the economic goal function of optimization and obtain；y_ipIt is corresponding i-th boiler combustion significant condition output quantity Jing after feedback compensation Prediction output；The dimension that m and n is respectively input into and exports；q_iAnd λ_jFor weight coefficient.

For above-mentioned Predictive Control of Nonlinear Systems algorithm, the fitness function for determining boiler combustion optimization controlled quentity controlled variable is formula (5), the minimum of a value of formula (5) object function is obtained as Rolling optimal strategy using particle swarm optimization algorithm, obtains boiler combustion Optimum control amount u (k+1) of optimization, acts on optimum control amount u (k+1) boiler combustion process nonlinear system and is controlled System.

Determine boiler combustion future controlled quentity controlled variable u (k+1)=(u (k using the online rolling optimizations in real time of PSO₁+1),u(k₂+ 1),…,u(k_m+1)).Particle population size is L, and particle i is expressed as pop_i=(u_i,v_i,l_i,F_ibest,F_i), wherein：u_i= (u_i1,u_i2,…,u_im) represent particle i position vector, v_i=(v_i1,v_i2,…,v_im) represent that the history that particle i is passed through is best Position, F_ibestRepresent the adaptive optimal control value of particle i, F_iRepresent the current adaptive value of particle i.In the t time iteration, the speed of particle i Degree, displacement more new formula, the adjustment formula of inertia weight is as follows：

Wherein：c₁,c₂For accelerated factor, r₁,r₂For the random number between [0,1].All particles find most in whole colony Good position is g=(g₁,g₂,…,g_m), it is that PSO optimizes the optimum control amount for obtaining, the i.e. optimum control of boiler combustion system Amount.Whole boiler combustion optimization control algolithm step is as follows：

1) initialization system state, forecast model and PSO parameters, and in order to improve the arithmetic speed of model, first to pot Stove burning historical data carries out off-line training, obtains Nonlinear Prediction Models under boiler original state；

2) sampling instant k, on the basis of Nonlinear Prediction Models under the offline boiler original state set up, using online Incremental learning fuzzy neural network is trained to present sample data, online amendment boiler combustion optimization forecast model in real time；

3) to fixed boiler combustion optimization controlled quentity controlled variable u (k), the output y (k+d) of system, by online incremental learning Fuzzy Neural Network Prediction Model is obtainedIf boiler combustion controlled quentity controlled variable u (k+1) of optimization undetermined is particle in PSO Position vector, brings forecast model into, the estimation output of etching system when obtaining k+1By drift correction, the estimation is defeated Go out and obtain the adaptation value function F of particle；

4) current adaptive value F of each particle is compared_iWith itself adaptive optimal control value F_ibestIf, F_i<F_ibest, then update F_ibestAnd l_i.Compare the adaptive optimal control value and global optimum's adaptive value of particle, if F_ibest<F_global, then F is updated_globalAnd g；

5) weight of each particle, speed and displacement are updated by formula (6).Check whether and reach maximum iteration time, be then Exit, g is controlled quentity controlled variable u (k+1) for optimizing, otherwise continue PSO iteration；

6) optimum control amount u (k+1) is acted on into boiler combustion optimization nonlinear system, reality is carried out to boiler combustion process When control；

7) k, i.e. k+1 → k, return to step 2 are increased), the whole calculating process of repetition.

Claims (4)

1. a kind of boiler combustion optimization control method, it is characterised in that comprise the steps：

(1) boiler combustion nonlinear system is sampled, obtains the input/output data at current time；

(2) input/output data that real-time sampling is obtained is trained using online incremental learning fuzzy neural network, is set up The online incremental learning forecast model of boiler combustion nonlinear system；

(3) nonlinear Model Predictive is used the online incremental learning forecast model, is realized to boiler combustion process Optimal control；

In step (2), the online incremental learning forecast model is：

$\hat{y}(k+d)=f\left(y\right(k+d-1),y(k+d-2),...,y(k+d-p),u(k),u(k-1),...,u(k-q+1\left)\right)$

Wherein, u (k)=(u₁(k),u₂(k),…,u_m(k)) represent Boiler Combustion Optimization System controlled quentity controlled variable, y (k)=(y₁(k), y₂(k),…,y_n(k)) represent boiler combustion status target output, d represent boiler combustion system output postpone, p and q tables Show the input/output order of boiler combustion process nonlinear system；

In step (3), the method for the optimal control is：

(31) correction of boiler combustion nonlinear system output valve is obtained by the online incremental learning forecast model：

In current sample time k, by the past input/output of boiler combustion nonlinear system and current input u (k) by being built Online incremental learning forecast model obtains the output estimation value of boiler combustion nonlinear system

By boiler combustion nonlinear system input u (k+1) to be optimized and past input/output, boiler combustion is obtained non- The output estimation value of linear system

If the prediction deviation at k moment isUse drift correctionObtain correction

(32) input to boiler combustion expense linear system is optimized：

Determine that the object function of boiler combustion optimization controlled quentity controlled variable u is according to boiler combustion optimization controlled quentity controlled variable u：

$\min J=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{q}_i{\&lsqb;y_{ir}(k+d+1)-y_{ip}(k+d+1)\&rsqb;}^2+\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j{\&lsqb;u_j(k+1)-u_j\left(k\right)\&rsqb;}^2$

Wherein y_irFor the reference locus of i-th boiler combustion significant condition output quantity, by the economy for solving boiler combustion optimization Object function and obtain, y_ipThe prediction for being corresponding i-th boiler combustion significant condition output quantity Jing after feedback compensation output, m The dimension for being respectively input into n and exporting, q_iAnd λ_jFor weight coefficient；

The minimum of a value of above-mentioned object function is obtained by the online rolling optimization in real time of particle cluster algorithm, optimum control amount u (k+ is obtained 1), optimum control amount u (k+1) is acted on into boiler combustion nonlinear system and is optimized control.

2. boiler combustion optimization control method according to claim 1, it is characterised in that in step (1), the input number According to for boiler operatiopn operating parameter, the output data is boiler efficiency and fume emission NO_x。

3. boiler combustion optimization control method according to claim 2, it is characterised in that described boiler operatiopn manipulation ginseng Number includes load, coal-supplying amount, total air, fuel throttle opening, secondary air register aperture and after-flame throttle opening.

4. boiler combustion optimization control method according to claim 1, it is characterised in that in step (2), the online increasing Amount totally four layers of structure of fuzzy neural network of study：

Input layer, each neuron in this layer represents an input variable of online incremental learning forecast model, wherein using X₁,

X₂..., X_rRepresent that boiler respectively runs manipulation amount u (k) and associated front p orders output y (k) successively；

Membership function layer, each input variable X_iThere is u membership function A_ij, it is Gauss member function：

${\mathrm{\&mu;}}_{ij}\left(x_i\right)=\exp \&lsqb;-{(x_i-c_{ij})}^2/{\mathrm{\&sigma;}}_{ij}^2\&rsqb;$

Wherein μ_ijIt is x_iJ-th membership function, j=1,2 ..., u, c_ijAnd σ_ijRespectively x_iJ-th Gaussian function in The heart and width, u is the quantity of membership function；

Fuzzy rule layer, j-th rule R_jOutput be：

Output layer, the output variable of each one input signal weighted sum of node on behalf：

Wherein y be characterize boiler combustion status optimization aim output valve, w_jFor result parameter.