CN101498458A - Model updating method for on-line boiler combustion optimization - Google Patents

Abstract

The invention relates to a method for updating a boiler on-line combustion optimization model. The method utilizes the data exceeding the predictive error limit of an exit model to build a new model, then utilizes combustion data of a boiler and an optimization algorithm to find an optimum new model and the weight coefficient of the existing model, and utilizes an optimum weight coefficient to combine the new model with the existing model so as to predict the working conditions of a new boiler for realizing model update. The invention overcomes the defects that the traditional update method gives up the existing model and can not utilize the study result of the existing model to fully utilize the study result of the existing model, thereby greatly reducing the amount and the time of the calculation work of model update.

Description

A kind of model update method of on-line boiler combustion optimization

Technical field

The invention belongs to information control technology field, relate to the incremental learning technology, particularly relate to a kind of model update method of on-line boiler combustion optimization.

Background technology

The burning optimization of station boiler is the important technical of energy-saving and emission-reduction, and its target is by adjusting the running status that operational factors such as boiler wind speed adjustment, coal supply obtain high efficiency or low pollution emission.The collocation of a lot of operational factors of boiler such as air distribution situation, coal supply situation and oxygen amount etc. all has direct influence to the boiler combustion state, and the configuration of operating parameters such as different air distributions, coal supply can directly cause the situation of the discharging of different boiler efficiencies and dusty gas.For given station boiler, under certain loading condiction, there is a kind of operating parameter allocation plan of optimum for the index of different fired states, can make the fired state optimization.Yet, very complicated coupled relation is arranged between the operating parameter of boiler, the configuration that find optimum operating parameter is very difficult.Along with continuous progress in science and technology, the boiler operatiopn automaticity improves constantly, but the boiler combustion optimization problem never well is resolved.

The research focus of boiler combustion optimization is by data mining at present, in a large amount of different actual operation parameters combinations, the method of applied for machines study, excavate the relational model between operational factor and boiler combustion index, utilize again and optimize algorithm carries out boiler in conjunction with the model of excavating online burning optimization.This method uses manpower and material resources sparingly, and can find the parameter configuration more excellent than artificial experiment, but because the characteristic of boiler plant changes to some extent along with the growth meeting of service time, and use coal and also can change sometimes, how guaranteeing therefore that model can upgrade fast and efficiently to adapt to news becomes the bottleneck problem of this method.This problem and modeling method, sample data are chosen and update strategy etc. all has much relations.

Summary of the invention

The objective of the invention is at the bottleneck problem in the boiler combustion optimization, propose a kind of model update method of taking into account historical learning outcome of model and new situation of change.

The present invention utilizes to exceed the data that have the model prediction limit of error and set up new model, and then utilize the boiler combustion data and optimize algorithm, seek the optimum new model and the weight coefficient of existing model, utilize optimum weight coefficient that new model and existing model are combined, jointly new boiler working condition is predicted that implementation model upgrades.This method has overcome and will have model in traditional update method and abandon fully, can not utilize the shortcoming of the learning outcome of existing model, has made full use of the learning outcome of existing model, has shortened the amount of calculation and the time of model modification greatly.

Technical scheme of the present invention is to make sample by the data of utilizing existing model prediction to exceed the limit of error, set up new model, and with new model and means such as existing model combines, establish a kind of model update method of on-line boiler combustion optimization, utilized the renewal of implementation model fast and efficiently of this method.

The step of the inventive method comprises:

Step (1) is set up the prediction error database that has had model.According to concrete boiler combustion situation with to the requirement of model prediction precision, the permission predicated error limit δ of setting model, when image data, error between judgment models predicted value and the actual motion value and the size that allows predicated error limit δ, if predicated error is greater than δ, promptly | V _c-V _s| δ, wherein V _cBe model predication value, V _sBe actual operating data, the data that then will transfinite deposit in the prediction error database, in order to the usefulness of model modification.

Step (2) is set up new model.When master mould need upgrade, utilize data in the prediction error database as training sample, sample can be expressed as

X wherein _iRepresent the boiler operating parameter vector of i group, y as the input data _iRepresent the parameter of i group as the sign boiler combustion state of output parameter, adopt algorithm of support vector machine to set up new model, kernel function is elected RBF as:

$K(x_i,x_j)=\mathrm{\φ}\left(x_i\right)\·\mathrm{\φ}\left(x_j\right)=\exp \left|\left(\frac{{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\σ}}^2}\right)\right|$

φ (x) is a mapping function, establishes the object function of being asked to be: f (x _i)=w φ (x _i)+b, f (x _i) being the boiler combustion index prediction value of model output, w is the weight coefficient vector, b is an intercept.Introduce relaxation factor ξ ^* _i〉=0 and ξ _i〉=0 and allow error of fitting ε, model can be by in constraint:

$\left\{\begin{array}{c}{y}_i-w\·\mathrm{\φ}\left(x_i\right)-b\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ w\·\mathrm{\φ}\left(x_i\right)+b-y_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{*}\\ {\mathrm{\ξ}}_i\≥0\\ {\mathrm{\ξ}}_i^{*}\≥0\end{array}\right.i=1,\·\·\·,N,$ Under the condition, minimize:

$\min R(w,\mathrm{\ξ},{\mathrm{\ξ}}^{*})=\frac{1}{2}w\·w+c\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\mathrm{\ξ}+{\mathrm{\ξ}}^{*}$

Obtain, wherein constant C 0 be penalty coefficient.This minimization problem is a convex quadratic programming problem, introduces Lagrangian:

$L(w,b,\mathrm{\ξ},{\mathrm{\ξ}}^{*},\mathrm{\α},{\mathrm{\α}}^{*},\mathrm{\γ},{\mathrm{\γ}}^{*})=\frac{1}{2}w\·w+C\underset{i=1}{\overset{N}{\mathrm{\Σ}}}(\mathrm{\ξ}+{\mathrm{\ξ}}^{*})-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i[y_i-({\mathrm{\ξ}}_i+\mathrm{\ϵ}+f\left(x_i\right))]$

$-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{*}[{\mathrm{\ξ}}_i^{*}+\mathrm{\ϵ}+f\left(x_i\right)-y_i]-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\γ}}_i{\mathrm{\ξ}}_i+{\mathrm{\γ}}_i^{*}{\mathrm{\ξ}}_i^{*})$

Wherein: ${\mathrm{\α}}_i,{\mathrm{\α}}_i^{*}\≥0,$ ${\mathrm{\γ}}_i,{\mathrm{\γ}}_i^{*}\≥0,$ Be Lagrange's multiplier.

At the saddle point place, function L is about w, b, ξ _i, ξ _i ^*Minimal point, also be α _i,

, γ _i,

Maximal point, minimization problem are converted into the maximization problems of asking its dual problem.

Lagrangian L is about w at the saddle point place, b, ξ _i, ξ _i ^*Minimal point:

$\left\{\begin{array}{c}\frac{\∂}{\∂w}L=0\→w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})\mathrm{\φ}\left(x_i\right)\\ \frac{\∂}{\∂b}L=0\→\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0\\ \frac{\∂}{\∂{\mathrm{\ξ}}_i}L=0\→C-{\mathrm{\α}}_i-{\mathrm{\γ}}_i=0\\ \frac{\∂}{\∂{\mathrm{\ξ}}_i^{*}}L=0\→C-{\mathrm{\α}}_i^{*}-{\mathrm{\γ}}_i^{*}=0\end{array}\right.$

Can get the dual function of Lagrangian:

$-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})\mathrm{\ϵ}+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})y_i$

At this moment,

$w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})\mathrm{\φ}\left(x_i\right)$

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})K(x,x_i)+b$

According to Ku En-Plutarch (KKT) conditional theorem, have following formula to set up at saddle point:

$\left\{\begin{array}{c}{\mathrm{\α}}_i[\mathrm{\ϵ}+{\mathrm{\ξ}}_i-y_i+f\left(x_i\right)]=0\\ {\mathrm{\α}}_i^{*}[\mathrm{\ϵ}+{\mathrm{\ξ}}_i+y_i-f\left(x_i\right)]=0\end{array}\right.i=1,\·\·\·,N$

By following formula as seen, ${\mathrm{\α}}_i\·{\mathrm{\α}}_i^{*}=0,$ α _iWith

Can not be non-zero simultaneously, can get:

$\left\{\begin{array}{cc}\begin{array}{c}{\mathrm{\ξ}}_i{\mathrm{\γ}}_i=0\\ {\mathrm{\ξ}}_i^{*}{\mathrm{\γ}}_i^{*}=0\end{array}& i=1,\·\·\·,N\end{array}\right.$

Can obtain b from following formula, obtain model.Empty the prediction error database, in order to upgrade the usefulness of image data next time.

Step (3) is determined the weight of new model and existing model.Gather the different operating conditions of up-to-date boiler data down as test samples, use the master mould prediction and predict weighted-average method, check data is predicted, i.e. y with new model _i=α D _n+ β D _o, y wherein _iBe i group Optimization of operating target prediction value, D _nBe new model predicted value, D _oBe original model predication value, α is a new model predicted value weight coefficient, β is the prediction weight coefficient of original model, and alpha+beta=1, definite employing particle cluster algorithm iteration optimizing of α and β is determined, definition particle cluster algorithm initial position Z vector respectively tie up component, be respectively new model weight and original model weight beta, object function is: min J=∑ | y _i-Y _i|, Y wherein _iIt is the actual motion value of i group operating mode.When J had reached minimum, reaches setting value or finished iterations, training was finished, and obtains the weight coefficient of new model and existing model.

Step (4) combines original model with the weight proportion of new model by optimum, constitute new model, i.e. D=α D _n+ β D _o, wherein D realizes the renewal of boiler combustion optimization model for the model after upgrading.

The model update method that the present invention proposes has made full use of the useful information that existing model comprised, significantly reduced the workload of model modification, improve the efficient of model modification, satisfied the actual requirement of boiler combustion on-line optimization, guaranteed the real-time and the accuracy of boiler combustion optimization.

The specific embodiment

A kind of model update method of on-line boiler combustion optimization, concrete steps are:

(1) sets up the prediction error database that has had model.According to concrete boiler combustion situation with to the requirement of model prediction precision, the permission predicated error limit δ of setting model, when image data, error between judgment models predicted value and the actual motion value and the size that allows predicated error limit δ, if predicated error is greater than δ, promptly | V _c-V _s| δ, wherein V _cBe model predication value, V _sBe actual operating data, the data that then will transfinite deposit in the prediction error database, in order to the usefulness of model modification.

(2) set up new model.When master mould need upgrade, utilize data in the prediction error database as training sample, sample can be expressed as

X wherein _iRepresent the boiler operating parameter vector of i group, y as the input data _iRepresent the parameter of i group as the sign boiler combustion state of output parameter, adopt algorithm of support vector machine to set up new model, kernel function is elected RBF as:

$K(x_i,x_j)=\mathrm{\φ}\left(x_i\right)\·\mathrm{\φ}\left(x_j\right)=\exp \left|\left(\frac{{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\σ}}^2}\right)\right|$

φ (x) is a mapping function, establishes the object function of being asked to be: f (x _i)=w φ (x _i)+b, f (x _i) being the boiler combustion index prediction value of model output, w is the weight coefficient vector, b is an intercept.Introduce relaxation factor ξ ^* _i〉=0 and ξ _i〉=0 and allow error of fitting ε, model can be by in constraint:

$\left\{\begin{array}{c}{y}_i-w\·\mathrm{\φ}\left(x_i\right)-b\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ w\·\mathrm{\φ}\left(x_i\right)+b-y_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{*}\\ {\mathrm{\ξ}}_i\≥0\\ {\mathrm{\ξ}}_i^{*}\≥0\end{array}\right.i=1,\·\·\·,N,$ Under the condition, minimize:

$\min R(w,\mathrm{\ξ},{\mathrm{\ξ}}^{*})=\frac{1}{2}w\·w+c\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\mathrm{\ξ}+{\mathrm{\ξ}}^{*}$

Obtain, wherein constant C 0 be penalty coefficient.This minimization problem is a convex quadratic programming problem, introduces Lagrangian:

$L(w,b,\mathrm{\ξ},{\mathrm{\ξ}}^{*},\mathrm{\α},{\mathrm{\α}}^{*},\mathrm{\γ},{\mathrm{\γ}}^{*})=\frac{1}{2}w\·w+C\underset{i=1}{\overset{N}{\mathrm{\Σ}}}(\mathrm{\ξ}+{\mathrm{\ξ}}^{*})-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i[y_i-({\mathrm{\ξ}}_i+\mathrm{\ϵ}+f\left(x_i\right))]$

$-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{*}[{\mathrm{\ξ}}_i^{*}+\mathrm{\ϵ}+f\left(x_i\right)-y_i]-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\γ}}_i{\mathrm{\ξ}}_i+{\mathrm{\γ}}_i^{*}{\mathrm{\ξ}}_i^{*})$

Wherein: ${\mathrm{\α}}_i,{\mathrm{\α}}_i^{*}\≥0,$ ${\mathrm{\γ}}_i,{\mathrm{\γ}}_i^{*}\≥0,$ Be Lagrange's multiplier.

At the saddle point place, function L is about w, b, ξ _i, ξ _i ^*Minimal point, also be

Maximal point, minimization problem are converted into the maximization problems of asking its dual problem.

Lagrangian L is about w at the saddle point place, b, ξ _i, ξ _j ^*Minimal point:

$\left\{\begin{array}{c}\frac{\∂}{\∂w}L=0\→w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})\mathrm{\φ}\left(x_i\right)\\ \frac{\∂}{\∂b}L=0\→\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0\\ \frac{\∂}{\∂{\mathrm{\ξ}}_i}L=0\→C-{\mathrm{\α}}_i-{\mathrm{\γ}}_i=0\\ \frac{\∂}{\∂{\mathrm{\ξ}}_i^{*}}L=0\→C-{\mathrm{\α}}_i^{*}-{\mathrm{\γ}}_i^{*}=0\end{array}\right.$

Can get the dual function of Lagrangian:

$-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})\mathrm{\ϵ}+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})y_i$

At this moment,

$w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})\mathrm{\φ}\left(x_i\right)$

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})K(x,x_i)+b$

According to Ku En-Plutarch (KKT) conditional theorem, have following formula to set up at saddle point:

$\left\{\begin{array}{c}{\mathrm{\α}}_i[\mathrm{\ϵ}+{\mathrm{\ξ}}_i-y_i+f\left(x_i\right)]=0\\ {\mathrm{\α}}_i^{*}[\mathrm{\ϵ}+{\mathrm{\ξ}}_i+y_i-f\left(x_i\right)]=0\end{array}\right.i=1,\·\·\·,N$

By following formula as seen, ${\mathrm{\α}}_i\·{\mathrm{\α}}_i^{*}=0,$ α _iWith

Can not be non-zero simultaneously, can get:

$\left\{\begin{array}{cc}\begin{array}{c}{\mathrm{\ξ}}_i{\mathrm{\γ}}_i=0\\ {\mathrm{\ξ}}_i^{*}{\mathrm{\γ}}_i^{*}=0\end{array}& i=1,\·\·\·,N\end{array}\right.$

Can obtain b from following formula, obtain model.Empty the prediction error database, in order to upgrade the usefulness of image data next time.

(3) determine the weight of new model and existing model.Gather the different operating conditions of up-to-date boiler data down as test samples, use the master mould prediction and predict weighted-average method, check data is predicted, i.e. y with new model _i=α D _n+ β D _o, y wherein _iBe i group Optimization of operating target prediction value, D _nBe new model predicted value, D _oBe original model predication value, α is a new model predicted value weight coefficient, β is the prediction weight coefficient of original model, and alpha+beta=1, definite employing particle cluster algorithm iteration optimizing of α and β is determined, definition particle cluster algorithm initial position Z vector respectively tie up component, be respectively new model weight and original model weight beta, object function is: min J=∑ | y _i-Y _i|, Y wherein _iIt is the actual motion value of i group operating mode.When J had reached minimum, reaches setting value or finished iterations, training was finished, and obtains the weight coefficient of new model and existing model.

(4) original model is combined with the weight proportion of new model by optimum, constitute new model, i.e. D=α D _n+ β D _o, wherein D realizes the renewal of boiler combustion optimization model for the model after upgrading.

Claims (1)

1, a kind of model update method of on-line boiler combustion optimization is characterized in that the step of this method comprises:

Step (1). set up the prediction error database there has been model: according to concrete boiler combustion situation with to the requirement of model prediction precision, the permission predicated error limit δ of setting model, when image data, error between judgment models predicted value and the actual motion value and the size that allows predicated error limit δ, if predicated error is greater than δ, promptly | V _c-V _s| δ, the data that then will transfinite deposit in the prediction error database, wherein V _cBe model predication value, V _sBe actual operating data;

Step (2). set up new model: when master mould need upgrade, utilize data in the prediction error database as training sample, schedule of samples is shown

X wherein _iRepresent the boiler operating parameter vector of i group, y as the input data _iRepresent the parameter of i group as the sign boiler combustion state of output parameter, adopt algorithm of support vector machine to set up new model, kernel function is elected RBF as

$K(x_i,x_j)=\mathrm{\φ}\left(x_i\right)\·\mathrm{\φ}\left(x_j\right)=\exp \left|\frac{{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\σ}}^2}\right|$

φ (x) is a mapping function, establishes the object function of being asked to be: f (x _i)=w φ (x _i)+b, f (x _i) being the boiler combustion index prediction value of model output, w is the weight coefficient vector, b is an intercept; W, b by $\min R(w,\mathrm{\ξ},\mathrm{\ξ}*)=\frac{1}{2}w\·w+c\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\mathrm{\ξ}+\mathrm{\ξ}*$ In condition

$\left\{\begin{array}{c}{y}_i-w\·\mathrm{\φ}\left(x_i\right)-b\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ w\·\mathrm{\φ}\left(x_i\right)+b-y_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{*}\\ {\mathrm{\ξ}}_i\≥0\\ {\mathrm{\ξ}}_i^{*}\≥0\end{array}\right.i=1,\·\·\·,N$

Find the solution and get; Introduce relaxation factor ξ in the formula ^* _iAnd ξ _i, ξ ^* _i〉=0 and ξ _i〉=0, ε is for allowing error of fitting;

Step (3). determine the weight of new model and existing model; Data under the different operating conditions of collection boiler are used the master mould prediction and are predicted weighted-average method with new model as test samples, check data are predicted, i.e. y _i=α D _n+ β D _o, y wherein _iBe i group Optimization of operating target prediction value, D _nBe new model predicted value, D _oBe original model predication value, α is a new model predicted value weight coefficient, and β is the prediction weight coefficient of original model, and alpha+beta=1;

Definite employing particle cluster algorithm iteration optimizing of α and β determines, definition particle cluster algorithm initial position Z vector respectively tie up component, be respectively new model weight and original model weight beta, object function is: the minJ=∑ | y _i-Y _i|, Y wherein _iBe the actual motion value of i group operating mode, when J had reached minimum, reaches setting value or finished iterations, training was finished, and obtains the weight coefficient of new model and existing model;

Step (4). original model is combined with the weight proportion of new model by optimum, constitute new model, i.e. D=α D _n+ β D _o, wherein D realizes the renewal of boiler combustion optimization model for the model after upgrading.