CN106527176A - MFOA (modified fruit fly optimization algorithm)-SVM (support vector machine)-based boiler thermal efficiency and NOX modeling method - Google Patents

Abstract

The present invention discloses an MFOA-SVM MFOA (modified fruit fly optimization algorithm)-SVM (support vector machine)-based boiler thermal efficiency and NOX modeling method. According to the method, in order to improve the thermal efficiency of a boiler and reduce NOX emissions, a boiler modeling theory according to which a modified fruit fly optimization algorithm (MFOA) is adopted to optimize a support vector machine (SVM) is put forward; according to the problems of low optimization precision and low convergence speed of a fruit fly optimization algorithm (FOA), a three-dimensional search and adaptive variable step size strategy is adopted to modify the FOA, and the three parameters, namely, penalty factor C, nuclear parameter g and insensitive loss coefficient epsilon of the SVM are optimized , so that the prediction of the SVM for boiler combustion is more accurate.

Description

A kind of boiler thermal output and NO based on MFOA-SVMXModeling

Technical field

The present invention relates to control technology field, specifically a kind of boiler thermal output and NO based on MFOA-SVM_XModeling side Method.

Background technology

The coal burning of thermal power plant generates electricity and accounts for the 70% or so of national generating total amount, and its burning produces black smoke, NO_XIn air Pollutant, not only heavy damage environment more threaten the life and health of people, while in order to support national energy-saving reduce discharging advocate, reduce Station boiler operating cost and NO_XIt is extremely urgent Deng discharging.

As boiler running process is extremely complex, for given station boiler, rational proportion operational factor becomes control Boiler efficiency and NO_XThe Main Means of discharge, but the non-linear relation of complexity between parameters, is there is, it is invisible to increased Boiler combustion models difficulty, and a large amount of scholars have carried out many explorations to solve this problem, such as utilizes neutral net and loses Propagation algorithm is combined, and sets up boiler combustion model.But neural net model establishing needs substantial amounts of sample data, and training time Long, capability of fitting is poor, is not suitable for line modeling.Obtain larger achievement at present is mutually tied with genetic algorithm using support vector machine Close.In order to obtain more effectively, more accurately combustion model, the present invention carrys out comprehensive modeling with fruit bat algorithm and support vector machine. Vector machine (SVM) its Nonlinear Processing ability is held, the training study of small sample, efficient generalization ability is especially suitable for, it is ensured that The Global Optimality of good fitness and solution, its good characteristic are applied in terms of many modeling and forecastings, although The modeling for boiler is put into practice, but has all been the penalty factor and kernel functional parameter g using Swarm Intelligence Algorithm to model Optimizing is carried out, and then insensitive loss coefficient ε is substituted into into SVM models as fixed value.Original fruit bat algorithm is also deposited in addition Low optimization accuracy it is low, convergence rate is slow the problems such as.

The content of the invention

It is an object of the invention to provide a kind of boiler thermal output and NO based on MFOA-SVM_XModeling method, to solve The problem proposed in above-mentioned background technology；For achieving the above object, the present invention provides following technical scheme：It is a kind of to be based on MFOA- The boiler thermal output and NO of SVM_XModeling method, comprises the following steps：

Given sample set { (x₁,y₁),(x₂,y₂)…,(x_l,y_l), wherein x_i∈Rⁿ, y_i∈ R, i=1,2 ..., l.x_i,y_i Input quantity and output are corresponded to respectively, l is sample number.The linear fit function of optimum is obtained by training study：

In formula：ω is weight vector, and b is amount of bias.The essence that SVM is returned is exactly to find parameter ω and b so that for instruction Input variable x beyond white silk sample_iHave | y_i-f(x_i) |≤ε, i.e. insensitive loss function are 0, and cause 1/2 | | ω | |²=1/ 2ω^Tω is minimum.Thus it is summarized as solving-optimizing problem under constraints：

In formula：C is penalty factor, and its value is normal number；ξ_i ^*And ξ_iIt is slack variable；The size of C represents sample data Higher than the punishment degree of accuracy rating ε；ε is insensitive loss coefficient.

As ω is probably Infinite-dimensional, according to the principle of duality, increase Lagrange multiplier a_i ^*And a_i, constitute Lagrange letters Number, solves the output function that its optimization problem can obtain SVM：,

In formula：K(x_i, it is x) kernel function, Jing analyses use RBF, such as formula (4).

In formula：G is kernel functional parameter, and its size determines the relation of the multiformity and mapping function of sample data distribution, choosing The value for selecting optimum could obtain best supporting vector machine model.

As further scheme of the invention：

According to the ultimate principle of support vector machine, penalty factor, insensitive loss coefficient and kernel functional parameter affect mould Type convergence rate and precision of prediction.To obtain optimal supporting vector machine model parameter, FOA algorithms are combined with the problem, As FOA is to find globally optimal solution in two-dimensional space, it is impossible to accurately search out the optimal solution in three dimensions；Pass through in addition Many experiments find that fruit bat flight step-length affects larger to convergence rate and low optimization accuracy.It is therefore proposed that three dimensions search and Adaptive step stragetic innovation fruit bat algorithm, using three parameters of improved fruit bat algorithm optimization SVM regression models, optimum ginseng Number [C, g, ε] is the maximum fruit bat position of flavor concentration, as follows to relative parameters setting：

1) initialize the control parameter in FOA.It is 30 to arrange population scale m, and iterationses t is 100, random initializtion fruit Fly position (X₁,Y₁,Z₁), interval is set to [0,1].

2) X, Y, Z are the matrix variables of the row of i rows 2, calculate the range formula of each fruit bat flight：

In formula：Number of times of the t for current iteration；K is the step-length coefficient of expansion, takes k=1.7, therefore, the distance of fruit bat flight was both Consider the fruit bat position of prior-generation, it is contemplated that the evolution of iteration, makes flying distance carry out with the size of flavor concentration certainly Adapt to change.

Calculate fruit bat population and origin apart from D (i,:)：

Calculate flavor concentration decision content S (i), make variable [S (and i, 1), S (i, 2), S (i, 3)] represent parameter [C, g, ε], be Accelerate the calculating speed of model, change simultaneously each coefficient of decision content.

3) concentration decision function, carries out 3 folding cross validation model generalization abilities to sample data, and concentration decision function is such as Under：

In formula：L is the population of each training subset in cross validation；f(x_i) for actual value；y_iPredictive value.

4) fly up to position calculation concentration decision function value to every fruit bat first, retain maximum position, fruit bat colony whole Fly to the position, into iterative cycles, the position or iterationses for having been look for Cmax convergence reaches maximum (100), then stop optimizing and export MFOA-SVM models.

As further scheme of the invention：

Control parameter (total generated output, feeder coal-supplying amount, flue gas oxygen content, the smoke evacuation temperature run using station boiler Degree, fire box temperature, secondary air damper aperture, total fuel quantity, total blast volume, economizer temperature, air preheater outlet cigarette temperature, generating Unit load) and 9 Elemental analysis datas being obtained using crane coal combustion of the power plant；It is possible to additionally incorporate boiler slag carbon content and flying dust contains Totally 36 parameters obtain boiler thermal output model to carbon amounts.Method according to back balance solves thermal efficiency value such as formula (9).

η=100- (q₂+q₃+q₄+q₅+q₆) (%) (9)

In formula：q₂For heat loss due to exhaust gas, q₃For fuel gas heat loss due to incomplete combustion, q₄For the imperfect combustion heat of solid Loss, q₅For radiation loss, q₆For heat loss due to sensible heat in slag.

According to the model structure built, the parameter of acquisition must preferably show boiler combustion characteristic, and random selection is not It is used as training study with the data of time period.Parameters can be retained in boiler combustion and running in different time and not With the substantial amounts of historical data under operating mode, when data are chosen, to randomly select in the data for entirely being provided, so can body The ergodic of existing data, enables model preferably to react real system.Every machine of grand celebration Boiler Control of Cogeneration Plant system Group is configured with 2 number of units according to server S CADA, and this server can be transmission control at engineer station and CSI control centres two Signal and carry out monitor in real time, routing function can be referred to as, while and the collection and storage of data can be carried out, receive from The instruction of control station, sends various state parameters for control station again.The system sets scene using the mode of Dynamic cata exchange The standby hot data of whole for monitoring are analyzed and calculate and generate Excel file.The present invention is by means of which from generation 180 groups are randomly selected in Excel file for modeling analysis, as kernel function relies on the inner product of |input paramete vector, is quickening Training speed, by normalized data sample, such as formula (10)；

Property value after normalizationTraining set and test set make in a like fashion, then can use following formula again Convert back actual value：

Description of the drawings

The schematic flow sheet of the SVM models that Fig. 1 is optimized based on MFOA

Fig. 2 boiler structures

Fig. 3 MFOA-SVM model structures

Fig. 4 data samples

Specific embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Site preparation is described, it is clear that described embodiment is only a part of embodiment of the invention, rather than the embodiment of whole；It is based on Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under the premise of creative work is not made Embodiment, belongs to the scope of protection of the invention；

Fig. 1 is the schematic flow sheet of a kind of boiler thermal output based on MFOA-SVM in the present invention and NOX modeling methods, As shown in figure 1, in the present invention, a kind of boiler thermal output based on MFOA-SVM and NOX modeling methods are comprised the following steps：

Given sample set { (x₁,y₁),(x₂,y₂)…,(x_l,y_l), wherein x_i∈Rⁿ, y_i∈ R, i=1,2 ..., l.x_i,y_i Input quantity and output are corresponded to respectively, l is sample number.The linear fit function of optimum is obtained by training study：

In formula：ω is weight vector, and b is amount of bias.The essence that SVM is returned is exactly to find parameter ω and b so that for instruction Input variable x beyond white silk sample_iHave | y_i-f(x_i) |≤ε, i.e. insensitive loss function are 0, and cause 1/2 | | ω | |²=1/ 2ω^Tω is minimum.Thus it is summarized as solving-optimizing problem under constraints：

In formula：C is penalty factor, and its value is normal number；ξ_i ^*And ξ_iIt is slack variable；The size of C represents sample data Higher than the punishment degree of accuracy rating ε；ε is insensitive loss coefficient.

As ω is probably Infinite-dimensional, according to the principle of duality, increase Lagrange multiplier a_i ^*, and a_i, constitute Lagrange Function, solves the output function that its optimization problem can obtain SVM：

In formula：K(x_i, it is x) kernel function, Jing analyses use RBF, such as formula (4).

K(x_i, x)=exp (- g | | x_i-x_j||²) (4)

In formula：G is kernel functional parameter, and its size determines the relation of the multiformity and mapping function of sample data distribution, choosing The value for selecting optimum could obtain best supporting vector machine model.

According to the ultimate principle of support vector machine, penalty factor, insensitive loss coefficient and kernel functional parameter affect mould Type convergence rate and precision of prediction.To obtain optimal supporting vector machine model parameter, FOA algorithms are combined with the problem, As FOA is to find globally optimal solution in two-dimensional space, it is impossible to accurately search out the optimal solution in three dimensions；Pass through in addition Many experiments find that fruit bat flight step-length affects larger to convergence rate and low optimization accuracy.It is therefore proposed that three dimensions search and Adaptive step stragetic innovation fruit bat algorithm, using three parameters of improved fruit bat algorithm optimization SVM regression models, optimum ginseng Number [C, g, ε] is the maximum fruit bat position of flavor concentration, as follows to relative parameters setting：

1) initialize the control parameter in FOA.It is 30 to arrange population scale m, and iterationses t is 100, random initializtion fruit Fly position (X₁,Y₁,Z₁), interval is set to [0,1].

2) X, Y, Z are the matrix variables of the row of i rows 2, calculate the range formula of each fruit bat flight：

In formula：Number of times of the t for current iteration；K is the step-length coefficient of expansion, takes k=1.7, therefore, the distance of fruit bat flight was both Consider the fruit bat position of prior-generation, it is contemplated that the evolution of iteration, makes flying distance carry out with the size of flavor concentration certainly Adapt to change.

Calculate fruit bat population and origin apart from D (i,:)：

Calculate flavor concentration decision content S (i), make variable [S (and i, 1), S (i, 2), S (i, 3)] represent parameter [C, g, ε], be Accelerate the calculating speed of model, change simultaneously each coefficient of decision content.

3) concentration decision function, carries out 3 folding cross validation model generalization abilities to sample data, and concentration decision function is such as Under：

In formula：L is the population of each training subset in cross validation；f(x_i) for actual value；y_iPredictive value.

4) fly up to position calculation concentration decision function value to every fruit bat first, retain maximum position, fruit bat colony whole Fly to the position, into iterative cycles, the position or iterationses for having been look for Cmax convergence reaches maximum (100), then stop optimizing and export MFOA-SVM models.

Control parameter (total generated output, feeder coal-supplying amount, flue gas oxygen content, the smoke evacuation temperature run using station boiler Degree, fire box temperature, secondary air damper aperture, total fuel quantity, total blast volume, economizer temperature, air preheater outlet cigarette temperature, generating Unit load) and 9 Elemental analysis datas being obtained using crane coal combustion of the power plant；It is possible to additionally incorporate boiler slag carbon content and flying dust contains Totally 36 parameters obtain boiler thermal output model to carbon amounts.Method according to back balance solves thermal efficiency value such as formula (9).

η=100- (q₂+q₃+q₄+q₅+q₆) (%) (9)

In formula：q₂For heat loss due to exhaust gas, q₃For fuel gas heat loss due to incomplete combustion, q₄For the imperfect combustion heat of solid Loss, q₅For radiation loss, q₆For heat loss due to sensible heat in slag.

According to the model structure built, the parameter of acquisition must preferably show boiler combustion characteristic, and random selection is not It is used as training study with the data of time period.Parameters can be retained in boiler combustion and running in different time and not With the substantial amounts of historical data under operating mode, when data are chosen, to randomly select in the data for entirely being provided, so can body The ergodic of existing data, enables model preferably to react real system.Every machine of grand celebration Boiler Control of Cogeneration Plant system Group is configured with 2 number of units according to server S CADA, and this server can be transmission control at engineer station and CSI control centres two Signal and carry out monitor in real time, routing function can be referred to as, while and the collection and storage of data can be carried out, receive from The instruction of control station, sends various state parameters for control station again.The system sets scene using the mode of Dynamic cata exchange The standby hot data of whole for monitoring are analyzed and calculate and generate Excel file.The present invention is by means of which from generation 180 groups are randomly selected in Excel file for modeling analysis, partial data is shown in Fig. 4.As kernel function relies on |input paramete vector Inner product, be accelerate training speed, by normalized data sample, such as formula (10).

Property value after normalizationTraining set and test set make in a like fashion, then can use following formula again Convert back actual value：

It is obvious to a person skilled in the art that the invention is not restricted to the details of above-mentioned one exemplary embodiment, Er Qie In the case of spirit or essential attributes without departing substantially from the present invention, the present invention can be realized in other specific forms；Therefore, no matter From the point of view of which point, embodiment all should be regarded as exemplary, and be nonrestrictive, the scope of the present invention is by appended power Profit is required rather than described above is limited, it is intended that all in the implication and scope of the equivalency of claim by falling Change is included in the present invention, should not be considered as any reference in claim and be limited involved claim；

Moreover, it will be appreciated that although this specification is been described by according to embodiment, not each embodiment is only wrapped Containing an independent technical scheme, this narrating mode of description is only that those skilled in the art should for clarity Using description as an entirety, the technical scheme in each embodiment can also Jing it is appropriately combined, form those skilled in the art Understandable other embodiment.

Claims (3)

1. a kind of boiler thermal output and NO based on MFOA-SVM_XModeling method, it is characterised in that comprise the following steps：Given sample This collection { (x₁,y₁),(x₂,y₂)…,(x_l,y_l), wherein x_i∈Rⁿ, y_i∈ R, i=1,2 ..., l, x_i,y_iInput is corresponded to respectively Amount and output, l is sample number, obtains the linear fit function of optimum by training study：

In formula：ω is weight vector, and b is amount of bias, and the essence that SVM is returned is exactly to find parameter ω and b so that for training sample Input variable x beyond this_iHave | y_i-f(x_i)≤ε, i.e. insensitive loss function are 0, and cause 1/2 | | ω | |²=1/2 ω^Tω Minimum, thus it is summarized as solving-optimizing problem under constraints：

$\min R(\mathrm{\ω},b,{\mathrm{\ξ}}_i,{{\mathrm{\ξ}}_i}^{*})=\frac{1}{2}{\mathrm{\ω}}^2+C\underset{i=1}{\overset{l}{\Σ}}({\mathrm{\ξ}}_i+{{\mathrm{\ξ}}_i}^{*})$

In formula：C is penalty factor, and its value is normal number；And ξ_iIt is slack variable；The size of C represents sample data higher than essence The punishment degree of degree scope ε；ε is insensitive loss coefficient,

As ω is probably Infinite-dimensional, according to the principle of duality, increase Lagrange multiplierAnd a_i, Lagrange functions are constituted, Solve the output function that its optimization problem can obtain SVM：

$f\left(x\right)=\underset{i=1}{\overset{l}{\Σ}}({a_i}^{*}-a_i)K(x_i,x)+b---\left(3\right)$

In formula：K(x_i, it is x) kernel function, Jing analyses use RBF, such as formula (4),

K(x_i, x)=exp (- g | | x_i-x_j||²)(4)

In formula：G is kernel functional parameter, and its size determines the relation of the multiformity and mapping function of sample data distribution, selects most Excellent value could obtain best supporting vector machine model.

2. a kind of boiler thermal output and NO based on MFOA-SVM according to claim 1_XModeling method, it is characterised in that：

According to the ultimate principle of support vector machine, penalty factor, insensitive loss coefficient and kernel functional parameter affect model receipts Speed and precision of prediction are held back, to obtain optimal supporting vector machine model parameter, FOA algorithms is combined with the problem, due to FOA is to find globally optimal solution in two-dimensional space, it is impossible to accurately search out the optimal solution in three dimensions；In addition through multiple Experiment finds that fruit bat flight step-length affects larger to convergence rate and low optimization accuracy, it is therefore proposed that three dimensions search and adaptive Step-length stragetic innovation fruit bat algorithm is answered, using three parameters of improved fruit bat algorithm optimization SVM regression models, optimized parameter [C, g, ε] is the maximum fruit bat position of flavor concentration, as follows to relative parameters setting：

1) control parameter in FOA is initialized, it is 30 to arrange population scale m, and iterationses t is 100, random initializtion fruit bat position Put (X₁,Y₁,Z₁), interval is set to [0,1],

2) X, Y, Z are the matrix variables of the row of i rows 2, calculate the range formula of each fruit bat flight：

$\left\{\begin{array}{c}X(i,:)=X_1+k(X_t-X_{t-1})\\ Y(i,:)=Y_1+k(Y_t-Y_{t-1})\\ Z(i,:)=Z_1+k(Z_t-Z_{t-1})\end{array}\right.---\left(5\right)$

In formula：Number of times of the t for current iteration；K is the step-length coefficient of expansion, takes k=1.7, therefore, the distance of fruit bat flight both considered The fruit bat position of prior-generation, it is contemplated that the evolution of iteration, makes flying distance carry out self adaptation with the size of flavor concentration Change,

Calculate fruit bat population and origin apart from D (i,:)：

Calculate flavor concentration decision content S (i), make variable [S (and i, 1), S (i, 2), S (i, 3)] represent parameter [C, g, ε], be quickening The calculating speed of model, changes simultaneously each coefficient of decision content,

$\left\{\begin{array}{c}S(i,:)=1/D(i,:)\\ C=100\·S(i,1)\\ g=S\left(i,2\right)\\ \mathrm{\ϵ}=1/20\·S(i,3)\end{array}\right.---\left(7\right)$

3) concentration decision function, carries out 3 folding cross validation model generalization abilities to sample data, and concentration decision function is as follows：

$Smell\left(i\right)=\frac{1}{3}\underset{k=1}{\overset{3}{\Σ}}\left\{\frac{1}{l}\underset{i=1}{\overset{l}{\Σ}}{\[y_i-f\left(x_i\right)\]}^2\right\}---\left(8\right)$

In formula：L is the population of each training subset in cross validation；f(x_i) for actual value；y_iPredictive value, 4) to every fruit bat Fly up to position calculation concentration decision function value first, retain maximum position, fruit bat colony is all flown to the position, follows into iteration Ring, the position or iterationses for having been look for Cmax convergence reach maximum (100), then stop optimizing and export MFOA-SVM models.

3. a kind of boiler thermal output based on MFOA-SVM according to claim 1 and NOX modeling methods, its feature exist In control parameter (total generated output, feeder coal-supplying amount, flue gas oxygen content, exhaust gas temperature, the stove run using station boiler Bore temperature, secondary air damper aperture, total fuel quantity, total blast volume, economizer temperature, air preheater outlet cigarette temperature, generating set Load) and 9 Elemental analysis datas being obtained using crane coal combustion of the power plant；It is possible to additionally incorporate boiler slag carbon content and unburned carbon in flue dust Totally 36 parameters obtain boiler thermal output model, solve thermal efficiency value such as formula (9) according to the method for back balance,

η=100- (q₂+q₃+q₄+q₅+q₆) (%) (9)

In formula：q₂For heat loss due to exhaust gas, q₃For fuel gas heat loss due to incomplete combustion, q₄For solid-unburning hot loss, q₅For radiation loss, q₆For heat loss due to sensible heat in slag,

According to the model structure built, the parameter of acquisition must preferably show boiler combustion characteristic, and during random selection difference Between the data of section be used as training study, parameters can be retained in boiler combustion and running in different time and different works Substantial amounts of historical data under condition, when data are chosen, will randomly select in the data for entirely being provided, can so embody number According to ergodic, enable model preferably to react real system, every machine of grand celebration Boiler Control of Cogeneration Plant system is assembled 2 number of units are put according to server S CADA, this server can be that control signal is transmitted at engineer station and CSI control centres two With carry out monitor in real time, routing function can be referred to as, while and the collection of data and storage can be carried out, receive from control The instruction stood, sends various state parameters for control station again, and the system is supervised field apparatus using the mode of Dynamic cata exchange The hot data of whole controlled are analyzed and calculate and generate Excel file, and the present invention is by means of which from the Excel for generating 180 groups are randomly selected in file for modeling analysis, as kernel function relies on the inner product of |input paramete vector, is to accelerate training speed Degree, by normalized data sample, such as formula (10)

$\stackrel{\‾}{x_i}=\frac{x_i-min\left(x_i\right)}{max\left(x_i\right)-\min \left(x_i\right)}---\left(10\right)$

Property value after normalizationTraining set and test set make in a like fashion, then can use following formula to convert again Return actual value：

$x_i=\min \left(x_i\right)+\stackrel{\‾}{x_i}(max\left(x_i\right)-\min \left(x_i\right))---\left(11\right).$