CN101504152B - Plant control method and plant controller - Google Patents

Abstract

A control system includes a basic control command operating unit, a fuel data storage unit, a running results database for storing past running results values of a control subject, a data creating unit configured to calculate a distance between data of the past running results values and the data sets and determining data set in which a distance between data becomes minimum, a modeling unit configured to model a relationship between operation parameters of a combustion apparatus and components in combustion gas of the combustion apparatus by using the data set determined by the data creating unit and a correcting unit for calculating combustion apparatus operation parameters with which components having a better condition than that of the components in a current gas are provided by using a model of the modeling unit and correcting operation command values of the basic control command operating unit by calculated operation parameters.

Description

The control method of complete set of equipments and control device

The application is that application number is 200710084770.9, and the applying date is on February 28th, 2007, denomination of invention for " have burner the control object control device and have the control device of the complete set of equipments of boiler " the dividing an application of application for a patent for invention.

Technical field

The present invention relates to have the control method and the control device of the complete set of equipments of boiler.

Background technology

In the past, being controlled to be basic control logic in complete set of equipments control field with PID is main flow.In addition, also propose utilizing with the neutral net is the band teacher learning function of representative, can with the characteristic of complete set of equipments corresponding multiple technologies scheme flexibly.

In order to use band teacher learning function to constitute control device,, therefore teacherless learning's method has been proposed also owing to need prepare the successful case that becomes teacher's data in advance.

As teacherless learning's example, the intensified learning method is arranged.

The intensified learning method be for by with the interaction of the repetition test of environment such as control object etc., the instrumentation signal that will obtain from environment is as desirable signal, generates the framework to the study control of the operation signal of environment.Following advantage is arranged thus, that is, even in the occasion that can not prepare successful case in advance, as long as by pre-defined desirable state, just can be own according to the desirable action of environment learning.

In intensified learning, has following learning functionality, promptly, be clue with the evaluation of estimate (intensified learning, being called as remuneration) of using the scalar that the instrumentation signal that obtains from environment calculated, generation is to the operation signal of environment, makes to reach maximum from standing state to the desired value of resulting evaluation of estimate in the future.As the method that realizes this learning functionality, for example non-patent literature 1 (intensified learning (Reinforcement Learning) is arranged, loyal virtue, all refined Zhang Gongtong translations in river on three, Co., Ltd. is published in gloomy north, and on December 20th, 2000 published) described Actor-Critic, Q learn, real-time DYNAMIC PROGRAMMING (real Time Inter Dynamic Programming) scheduling algorithm.

In addition, as the framework of the intensified learning that improves said method, above-mentioned document has been introduced the framework that is called as dynamic-architecture (Dyna-ア one キ テ Network チ ヤ).It is relatively good that this is that a kind of model with simulation control object is that object learns to generate what kind of operation signal in advance, uses the method for the operation signal that this learning outcome decision applies the control object then.And the model with the error that reduces to control object and model is adjusted function.

In addition, there are the following problems in the control device of the complete set of equipments that possesses burner, that is, in for example indefinite occasion of fuel performance or changed occasion of coal class etc. as coal fuel, the combustion characteristics or the thermal conduction characteristic of complete set of equipments change.As to method that should problem, can enumerate for example described technology of patent documentation 1-TOHKEMY 2004-190913 communique.

This be a kind of in coal fired boiler the method from the deviation computing heat output of fuel ratio of the measured value of main steam pressure and setting value.

In addition, the control device of patent documentation 2-Japanese kokai publication hei 8-200604 communique comprises: calculate based on the fluid measuring data relevant with the temperature of coal fired boiler on burner hearth, pressure, flow etc. and first evaluation unit of the absorption heat estimated value of the burner hearth of estimating; Calculate based on the fluid measuring data relevant and second evaluation unit of the absorption heat estimated value of the final secondary burner of estimating with the temperature of final secondary burner, pressure, flow etc.; The unit of the ratio of the absorption heat estimated value of the final secondary burner of obtaining the absorption heat estimated value of the burner hearth of calculating and calculating with above-mentioned second evaluation unit with above-mentioned first evaluation unit; And, grasp the combustion characteristics of boiler based on the ratio of the absorption heat estimated value that utilizes this unit to obtain, output gas distributes the revolution setting value of damper setting value, gas recycling fan and the arithmetic element that setting value is quickened in the boiler input.

In addition, in complete set of equipments such as boiler control field, in the past, being controlled to be basic control logic with PID became main flow.In addition, also propose utilizing with the neutral net is the band teacher learning function of representative, can with the characteristic of complete set of equipments corresponding multiple technologies scheme flexibly.And, constitute control device in order to use band teacher learning function, owing to need prepare the successful case that becomes teacher's data in advance, therefore teacherless learning's method of intensified learning method etc. has been proposed also.

This intensified learning method be for by with the interaction of the repetition test of environment such as control object etc., the instrumentation signal that will obtain from environment is as desirable signal, generates the framework to the study control of the operation signal of environment.Following advantage is arranged thus, that is, even in the occasion that can not prepare successful case in advance, as long as by pre-defined desirable state, just can be own according to the desirable action of environment learning.

And, in this intensified learning, has following learning functionality, promptly, be clue with the evaluation of estimate (intensified learning, being called as remuneration) of using the scalar that the instrumentation signal that obtains from environment calculated, generation is to the operation signal of environment, makes to reach maximum from standing state to the desired value of resulting evaluation of estimate in the future.In the method that realizes this learning functionality, Actor-Critic, Q study, real-time DYNAMIC PROGRAMMING (Actual Time Inter Dynamic Programming were arranged in the past) scheduling algorithm.

In addition, as the framework of the intensified learning that improves said method, the framework that is called as dynamic-architecture (Dyna-ア one キ テ Network チ ヤ) is arranged.It is relatively good that this is that a kind of model with simulation control object is that object learns to generate what kind of operation signal in advance, use the method for the operation signal that the decision of this learning outcome applies the control object then, at this moment, the model with the error that reduces to control object and model is adjusted function.

On the other hand, follow the progress of numerical value analytical technology, combustion reaction also can be reproduced its result by calculating to a certain extent, can utilize the simulator that is modeled to complete equipment to construct model (for example, with reference to patent documentation 3-TOHKEMY 2003-281462 communique).

Above-mentioned document is to carry out thermally equilibrated calculating etc. according to the variation of generated output value, temperature, pressure etc. to estimate the variation of heat output of fuel and the method for controlling, and has considered the influence to heat conductivility.But the variation of fuel performance not only influences heat conductivility, also influences the composition of burnt gas.

If NO _X, increase such as CO, then might bring influence to external environment condition, perhaps make the load rising of emission-control equipment etc., but in above-mentioned document record consider the method for influence that burnt gas is formed.

In addition, because combustion phenomena is the compound phenomenon of the complexity of the flowing of fuel and air (gas), heat transfer and combustion reaction, therefore control the problem that its movement is the comparison difficulty.

Especially form the variation of the variation generation of relative fuel performance about waste gas, it is difficult deriving suitable method of operating.Both made and utilized above-mentioned intensified learning theory, in test repeatedly, also needed long-term learning time for pluralities of fuel character is learnt its method of operating, and the possibility that also exists exhaust gas properties to worsen between the learning period.

Summary of the invention

The object of the invention is to provide a kind of variation of relative fuel composition also can suitably control the control device of burning gases composition.

In addition, as mentioned above, though the intensified learning method is effective in control Properties of Objects, control method in occasion that in advance can not regularization, but in the running control of complete set of equipments, the time that need before model is set up, compile the actual effect of repetition test operation, thereby product quality descends during having this, and the problem that loss increases, and, by repetition test operation at this moment, though also can consider to change the influence that causes, not consider to the relatively more effective model construction method of these problems to environment by the character of complete set of equipments effluent.

In addition, in patent documentation 3, need the refinement computing grid in order to improve computational accuracy, but huge in the occasion amount of calculation of large-scale plants such as boiler, more owing to operating condition also changes continuously, so computing time is long, and it is difficult constructing model with the time of practicality.

As mentioned above, though the intensified learning method is an effective method for the control of the complete set of equipments of control method regularization in advance, but need a lot of times carry out data accumulation in order to construct model, can not get desirable control performance between break-in period sometimes according to the test running of complete set of equipments.

Second purpose of the present invention is to provide a kind of short and complete set of equipments control method and device with intensified learning method of good performance during utilizing model to construct.

Control device of the present invention possesses: import the basic controlling ordering calculation unit of the instrumentation data union of the control object with burner to the operational order value of above-mentioned control object; The fuel data memory cell of the data group of composition in the operating parameter of above-mentioned burner and the above-mentioned gas is stored in relatively a plurality of fuel compositions of the fuel of supplying with to above-mentioned burner; Store the running actual database of running actual value in the past of above-mentioned control object; The data pitch of calculating the running actual value in past of above-mentioned control object and above-mentioned data group from, decision makes data pitch from the data generating unit for the shortest data group; Use is by the data group of above-mentioned data generating unit decision, with the modelling unit of the relational modelization of the composition in the burning gases of the operating parameter of above-mentioned burner and above-mentioned burner; And the model that uses above-mentioned modelling unit calculates the operating parameter than the better burner of member condition in the current gas, proofreaies and correct the correcting unit of the operational order value of above-mentioned basic controlling ordering calculation unit with the operating parameter of calculating.

Above-mentioned second purpose realizes by the following method: in the learning functionality of the relation with learning manipulation amount and complete set of equipments state, and have and utilize this learning functionality to calculate in the control method corresponding to the complete set of equipments control device of the function of the operational order value of complete set of equipments state, the numerical analysis that utilizes flowing of a plurality of operating conditions to reach reacting phenomenon comes the process values of the state of computational rules complete set of equipments, the process values of each operating condition is generated continuous model approx, make the variation of its relative operating condition parameter become continuous relation, use generates continuous model once more by the service data of numerical analysis calculation process value and complete set of equipments actual machine, uses the above-mentioned continuous model that generates once more to learn.

Effect of the present invention is that the present invention is made of said units, even fuel composition (character) changes, also can automatically suitably control exhaust gas constituents, so can reduce the NO in the waste gas _X, harmful substance such as CO generation.

The present invention is owing to possess the method that realizes above-mentioned second purpose, so can be from the complete set of equipments test running time, use the result of numerical analysis to utilize the intensified learning method to be controlled to complete equipment, thereby can shorten between the introduction period of control device.

Description of drawings

Fig. 1 is the structure chart of the control device of explanation first embodiment of the present invention.

Fig. 2 is the structure chart of the control device of explanation second embodiment of the present invention.

Fig. 3 is the structure chart of the control device of explanation the 3rd embodiment of the present invention.

Fig. 4 is the figure of the calculating process of explanation data generating unit.

Fig. 5 is the display frame example figure of explanation model error.

Fig. 6 is the display frame example figure of explanation model error.

Fig. 7 is the structure chart of explanation thermal power generation complete equipment.

Fig. 8 is the figure of the correspondence of explanation power plant group and analysis center.

Fig. 9 is the concept map of explanation intensified learning.

Figure 10 is that explanation fuel data storehouse shows and fuel performance input picture example figure.

Figure 11 is the figure of explanation corrective loop.

Figure 12 is the figure of the processing procedure of description status evaluation unit 300 and data group switch unit 310.

Figure 13 is the figure of explanation corrective loop.

Figure 14 is the square frame pie graph of an embodiment of expression complete set of equipments control device of the present invention.

Figure 15 is the key diagram of an example of expression running actual database of an embodiment of the invention or numerical analysis database.

Figure 16 is the key diagram of an example of the learning outcome database of expression an embodiment of the invention.

Figure 17 is the flow chart of expression according to the processing procedure of an embodiment of the invention.

Figure 18 is the key diagram of the continuous model of one embodiment of the present invention.

Figure 19 is the flow chart of expression according to the numerical analysis data supplementing processing procedure of an embodiment of the invention.

Figure 20 is the key diagram of the error assessment of an embodiment of the invention.

Figure 21 is the key diagram of the numerical analysis data supplementing of an embodiment of the invention.

Figure 22 is the key diagram of the continuous model correction of an embodiment of the invention.

Among the figure:

The 100-complete set of equipments; The 200-control device; The 210-data generating unit; The 220-input/output interface; The 221-input-output unit; 230-basic controlling ordering calculation unit; The 240-actual database that turns round; 250-modelling unit; The 260-correcting unit; 270-fuel data memory cell.

1001-complete set of equipments (for example boiler plant); The outside output interface of 1002-; The 1003-outer input interface; 1004-basic controlling ordering calculation unit; The 1005-actual database that turns round; 1006-numerical analysis unit; 1007-numerical analysis database; 1008-continuous model unit; The 1009-continuous model; 1010-error assessment unit; The 1011-parameter is appended the unit; 1012-continuous model amending unit; 1013-control method unit; 1014-learning outcome database; 1015-operational ton arithmetic element; The 1016-subtracter; The 1017-adder; The 1018-switch.

The specific embodiment

Below, with reference to the description of drawings preferred forms.Fig. 1 represents first embodiment.Control device 200 of the present invention is from the instrumentation information 205 of complete set of equipments 100 receiving course values as the control object, uses it to carry out the computing of pre-sequencing and operation instruction signal (control signal) is sent to complete set of equipments 100 in control device 200.Complete set of equipments 100 is according to the operation instruction signal 285 that receives, and makes the driver action of the aperture that for example is called valve or damper aperture and is controlled to the state of complete equipment.

Present embodiment is the example that is applied to the burning control of thermal power generation complete equipment.In this example, special in to be applied to reduce the NO in the waste gas _XAnd the concentration of CO is that the center describes as the example on the control function of purpose.

Fig. 7 represents the structure as the thermal power generation complete equipment of control object.By burner 102 coal that acts as a fuel, the primary air of coal carrying usefulness and the auxiliary air that usefulness is adjusted in burning are dropped in the boiler 101, with boiler 101 coal combustion.Coal and primary air are drawn from pipe arrangement 134, and auxiliary air is drawn from pipe arrangement 141.In addition, the follow-up air with two sections burning usefulness drops in the boiler 101 by follow-up air scoop 103.This follow-up air is drawn from pipe arrangement 142.

The high-temperature gas that produces by burning of coal is after the path flow of boiler 101, by air heater 104.Afterwards, after removing harmful substance, be discharged to the atmosphere from chimney with emission-control equipment.

The water supply of circulation is incorporated in the boiler 101 by feed pump 105 in boiler 101, by gas superheat, becomes the steam of HTHP in heat exchanger 106.In addition, the quantity with heat exchanger is set at 1 in the present embodiment, but also can dispose a plurality of heat exchangers.

Passed through the steam of the HTHP of heat exchanger 106, introduced in the steam turbine 108 by turbo-regulator 107.The energy drives steam turbine 108 that utilizes steam to have is with engine 109 generatings.

The path of the follow-up air that then, the primary air that drops into from burner 102 and auxiliary air are described, drops into from follow-up air scoop 103.

Primary air is introduced in the pipe arrangement 130 from fan 120, branches in the way by the pipe arrangement 132 of air heater with not by its 131, collaborates once more at pipe arrangement 133, is incorporated into grinding mill 110.Air by air heater utilizes gas superheat.The coal (fine breeze) that uses this primary air to generate in grinding mill 110 is blown in the burner 102.

Auxiliary air and follow-up air are incorporated in the pipe arrangement 140 from fan 121, after overheated with air heater 104, branch into the pipe arrangement 142 that pipe arrangement 141 that auxiliary air uses and follow-up air are used, and are incorporated into burner 102 and follow-up air scoop 103 respectively.

Control device 200 is in order to reduce NO _XAnd the concentration of CO, have air capacity that adjustment drops into from burner and the function of the air capacity that drops into from follow-up air scoop.In addition, though not expression in Fig. 7, have the situation of the part of burnt gas being returned the gas recirculation device of burner hearth but exist, perhaps have the situation of the ejection angle of burner being made variable up and down device, also can be with these objects as control operation.The air mass flow that becomes the fuel flow rate of supplying with to burner, burner air flow of control object, supplies with to air scoop, the operational ton of gas recirculation device, the ejection angle of burner etc. all are the operating parameters to boiler.

Control device 200 comprises: basic controlling ordering calculation unit 230; The correcting unit 260 of change or correction basic operation command value 235 of 230 outputs from basic controlling ordering calculation unit; The running actual database 240 of the running real data that accumulation, storage constitute by operation instrumentation value 205, operator's input signal, from the command signal of upper level control system etc.; Be used for and control object complete set of equipments 100 or operator's etc. the reception of carrying out data and the input/output interface 220 of transmission; And the input-output unit 221 of the operational order when being used for the operator and seeing various data or input setting value or operation mode, manual operation etc.

Basic controlling ordering calculation unit 230 with PID (ratio, integration, differential) controller as basic structural unit, with operation instrumentation value 205, operator's input signal, from the command signal of upper level control system etc. as input, computing is to being arranged on the basic operation command value 235 and the output of various Work machines such as valve on the complete set of equipments 100, damper, motor.

The function of basic operation command value 235 and structure are owing to identical with the control device in existing thermal power plant, so in this description will be omitted.

The invention is characterized in possess data generating unit 210, modelling unit 250, correcting unit 260 and fuel data memory cell 270.Below, function separately is described.

Modelling unit 250 has the function of the model of the relation that generates the fuel flow rate simulated as operating parameter, air mass flow etc. and certain specific component concentration in the waste gas of this operating parameter.

Read in data 275 from fuel data memory cell 270, use the neutral net that constitutes by input layer, intermediate layer, output layer to utilize error Back-Propagation method (back-propagating method) study input/output relation.The structure of neutral net and learning method are general methods, and these methods also can be other methods, because the present invention and do not rely on the structure or the learning method of neutral net, so in this detailed.

The input data are air mass flow, the fuel flow rate of each burner, the engine power of each position of burner and follow-up air scoop, and the output data are NO _XAnd the concentration of CO.

Though with the relational modelization of fuel flow rate, air mass flow, generated output and NOx and CO, the present invention only is not defined in these with cuit and output project in this example.In addition, modeling method also is not limited to neutral net, also can use other statistical method generation models such as regression model.

In the fuel data memory cell 270 each coal (coal is also different according to place of production heterogeneity) is all stored the input data of a plurality of modellings unit 250 and the data group of output data.

The data group has from actual operation database 240 to be extracted the data group of real data in the past and carries out the interior burning numerical analysis of boiler and the result of calculation of calculated in advance.

Only with regard to the running real data, generation model not before the accumulation operational data.And, there is not the coal class of use experience certainly can not generation model.Therefore, in the present invention,, its result is stored in the fuel data memory cell 270 with the numerical analysis of burning of the counting system of the design of simulated object complete set of equipments realistically and utilization condition.

Combustion phenomena is owing to be the compound phenomenon of the complexity of the flowing of fuel and air (gas), heat transfer and combustion reaction etc., so hold generally difficulty relatively of its movement.But, by having changed for example burning infrastest of conditions such as fuel performance (composition, particle diameter), combustion atmosphere, according to its result with the key element phenomenon modelization, thereby also can analyze large-scale as the boiler in thermal power plant and have the phenomenon of complicated inside movement with the precision of practicality.

In addition, though the numerical analysis technology just had in the past,, need huge calculating grid (grid) quantity for analyzing the especially large-scale plant as boiler to a certain degree accurately, because it wants the computing time of overspending in industrial calculating, in fact be impossible therefore.

But, by using for example described numerical analysis technology of TOHKEMY 2003-281462 communique, can keep precision also to analyze at high speed, and, the performance of computer in recent years also has raising, can realize the numerical analysis of the detailed phenomenon of large sized complete sets of equipment.

In the burning numerical analysis, multiple coal composition is calculated.Coal is formed by representational coal brand (kind) being carried out the composition analysis decision.Because coal is natural resources, so even its composition of the identical place of production is often also incomplete same.Therefore, analyze the sample of multiple situation, use its average composition.

In addition, it is more coal to be stored at outdoor situation in the power plant and since weather influence water content and diurnal variation, so even the coal of identical type also has the situation of characteristic variations.

So, make water content that the numerical analysis of burning of multiple situation of change be arranged for the coal of same kind, also store its result.Therefore, can estimate water content to NO _XAnd the degree of influence of CO concentration.

Then, use Fig. 4 that data generating unit 210 is described.

In step 500, read in a reference value (data pitch is from feasible value) that is used for judging whether to change the 250 data groups of using in the modelling unit.A reference value can just be stored in case import this value from keyboard 222 inputs.And, also can change a reference value afterwards.

In step 510, read in from specified time limit (for example one month) burner till current and air mass flow, the fuel flow rate of each burner, generator power, the NO of each position of follow-up air scoop from running actual database 240 _XThe actual value 245 of concentration, CO concentration.

In step 520, read in the numerical analysis result data group that the current model that is just using generates usefulness from fuel data memory cell 270.

In step 530, distinguish identical in order to make the running real data 245 of reading in the air mass flow of each position of burner and follow-up air scoop, the fuel flow rate of each burner, the value of generator power in step 510, the numerical analysis data group interpolation that to read in step 520 is calculated the NO of this moment between data _XThe interpolation value of concentration, CO concentration.

Though use the cubic spline interpolation as interpolating method, also can use other interpolating method.

Numerical analysis data group owing to the situation that can not become with the identical condition of running real data is in the majority, is therefore carried out this data interpolation owing to be the discrete data of calculating with predetermined condition, can make itself and the real data term harmonization that turns round.

In step 540, obtain the data point of the actual value of reading in step 510 and the data pitch of the interpolation value of the numerical analysis data calculated in step 530 from.

Data pitch is from the Euclidean distance that is defined as formula (1) expression.Coordinate at two data point P, Q be (Xp1, Xp2, Xp3 ..., Xpn), (Xq1, Xq2, Xq3 ..., in the time of Xqn), point-to-point transmission apart from d _JkSquare can obtain with formula 1.Here, Xpi, the Xqi as coordinate is air mass flow, the fuel flow rate of each burner, generator power, the NO of each position of burner and follow-up air scoop _XConcentration, CO concentration.In addition, j is the fuel data sets of numbers, and k is the NO in the fuel data group of j number _XAnd the instrumentation data number of CO.

[formula 1]

${d_{\mathrm{jk}}}^2={\mathrm{\Σ}}_{i=1}^n{(\mathrm{Xpi}-\mathrm{Xqi})}^2$ ... formula (1)

In step 550,, running real data 245 and distance each data point have been judged whether to calculate to being stored in the whole fuel composition data groups in the fuel data memory cell 270.

The occasion of calculating advances to step 560 in that whole fuel composition data groups are through with.Occasion in the remaining fuel composition data group of not calculating turns back to step 520, and change is as the fuel composition data group of calculating object and with distance between same process calculated data.

In step 560,, obtain and the average distance d of the real data 245 that turns round with formula (2) at first to each fuel data group _{J_ave}

[formula 2]

$d_{j\_\mathrm{ave}}=\frac{{\mathrm{\Σ}}_{m=1}^k{d}_{\mathrm{jm}}}{k}$ ... formula (2)

Then, select d _{J_ave}Be the minimal data group.

In step 570, compare a reference value (feasible value) and the d that read in step 500 _{J_ave}, if d _{J_ave}Below feasible value, then send d to modelling unit 250 _{J_ave}The information 215 of minimal data sets of numbers and finishing.

In addition, at d _{J_ave}Surpass the occasion of feasible value, advance to step 580.

In step 580, generate new data set, model revision directive signal is input in the modelling unit 250.At d _{J_ave}The occasion that surpasses feasible value means that original fuel composition data group and nearest running real data are inconsistent.Thereby, at d _{J_ave}Add the data group of having added nearest running real data on the minimum fuel composition data group and generate new data set.At this moment, Data Labels is set, so that can distinguish numerical analysis data and running real data.

The newly-generated data sets of numbers and the information 215 of model revision directive signal are exported in modelling unit 250.

Modelling unit 250 1 receives model revision directive information 215, just with reference to the new data set number, uses the data of this data group to regenerate model.Though model generating method is with above-mentioned same,, pay attention to the importance of running real data and generation model with Data Labels identification running real data.Specifically, by increasing the input number of times of running real data, it also is reflected on the model characteristics consumingly than other numerical analysis data.

Thus, at least near the data point that has the running actual result since can with the approaching characteristic modelization of running real data, thereby reduced the error of model.

Data generating unit 210 is for example carried out with one-week interval.At d _{J_ave}The occasion that surpasses feasible value is appended the running real data of the new accumulation of week age part, and is regenerated model once more, and the ratio that therefore is used for the running real data of the data group that model generates increases, and model characteristics moves closer to the real-world operation characteristic.

Then, correcting unit 260 is described.Correcting unit 260 is equivalent to the simulated operation command signal 265 of operating parameter to 250 outputs of modelling unit.Modelling unit 250 calculates the NO as the output valve of model to simulated operation command signals 265 such as the fuel flow rate of the air mass flow of each position of model input burner that generates and follow-up air scoop, each burner, generator powers _XReach CO concentration and these gas componant information are outputed to correcting unit 260.

In correcting unit 260, read in and respectively operate value as current complete set of equipments state from service data 205, with this state is that benchmark changes in the scope of air mass flow in transformable amplitude of variation of each position that made burner and follow-up air scoop before the time of operation next time, as these simulated operation instruction 265 outputs.

The variable amplitude of operational ton is obtained from the responsiveness and operation (control) interval of the driver of the damper of login in advance or valve etc.In addition, the variable amplitude of each operational ton is divided into specified quantity, their all combinations are changed operational ton.

Each simulated operation command signal 265 usefulness modelling unit that carried out like this changing is calculated NO _XAnd CO concentration.From the evaluation of estimate J that wherein extracts with formula (3) definition is minimum simulated operation command signal.Here, C _NOx, C _COBe respectively NO _XAnd the calculated value of CO concentration, A ₁, A ₂It is coefficient.

[formula 3]

J=A ₁C _NOx+ A ₂C _CO... formula (3)

To formula (3) input NO _XAnd the current instrumentation value of CO concentration, with the evaluation of estimate that calculates as J _RRelatively J that estimates with the calculated value of model and the J that estimates with current instrumentation value _R, in the occasion that the condition of formula (4) is set up, interpolation is proofreaied and correct and is exported as operation instruction signal 285 on basic operation command value 235.

[formula 4]

J＜J _R... formula (4)

In the occasion that formula (4) is set up, expression according to composition in the gas of the simulated operation command signal 265 of model liken to into member condition in the gas of current instrumentation value better.

Use Figure 11 that the method for correction is described.

With subtracter 281 calculating basic operation command value 235 and evaluation of estimate J is the deviation signal 287 of the simulated operation command signal 265 of minimum, it is added to adder 284 generates correct operation command value 288 on the basic operation command value 235.

If, owing to the unusual NO that makes as the output valve of modelling unit 250 in the unusual exclusive disjunction loop of importing data _XReach the occasion that CO concentration calculated value becomes unusual, owing to by being made as zero correct operation command value 288 is equated with basic operation command value 235, can reduce the danger of wrong output abnormality signal thus with the coefficient that multiplier 283 and deviation signal 287 multiply each other.

NO as the output valve of modelling unit 250 _XAnd whether CO concentration calculated value 255 unusual, to judge to the input data of modelling unit 250 and the upper lower limit value check of output data and the bound check of rate of change.Even at least one exceeds the occasion of predefined upper lower limit value, also be made as 0 output that prevents the simulated operation command signal 265 under the state that unusual possibility is arranged, estimated by output signal with switch 282.Switch 282 is set at 1 in occasion in addition with output signal.

Switch 286 receives the result of determination of formula (4), and one in selection basic operation command value 235 and the correct operation command value 288 as operation instruction signal 285 outputs.

According to above content, estimate operational ton and NO based on the result's of burning numerical analysis model by use _XAnd the relation of CO generation, can reduce NO _XAnd the operation of the generation of CO.

In addition, for the occasion of fuel performance (coal class) variation, owing to also can select suitable numerical analysis data group, so can keep the high accuracy of model.Therefore, the occasion in the change of coal class also can suppress NO _XAnd the control performance of CO descends.

Therefore, in the past by the information of operator to the variation of control device input fuel performance, the experience and knowledge that perhaps depends on the operator changes the mode of control parameter by automation.Therefore, do not rely on operator's technical merit and can carry out high-performance and stable running, can reduce operator's working load.

Have again, in the big occasion of deviation of the pre-prepd numerical analysis result data group and the real data that turns round, owing to generate the new data set of having appended the running real data, therefore can be to changing model gradually automatically with the approaching model of real data that turns round.

In addition, be decided to be NO though in the example of present embodiment, will control the object process values _XAnd CO concentration, but the present invention is not limited thereto, and also can be with the CO in the gas ₂, SO _X, Hg (mercury) amount, fluorine, the particulate subclass that is made of coal dust or mist, VOC (VOC) be as object.

Then, use Fig. 2 that second embodiment is described.

Be with the difference of above-mentioned first embodiment, use 290 study of intensified learning unit to reduce NO _XAnd the method for operating of CO.

Intensified learning unit 290 has following function,, uses the running number be accumulated in the running actual database 240 that is, according to according to the suitable method of operating of intensified learning theory study corresponding to the complete set of equipments state.

The detailed explanation of intensified learning theory, because existing narration in for example " Co., Ltd. is published in gloomy north; on December 20th, 2000 published for intensified learning (ReinforcementLearning), loyal virtue, all refined Zhang Gongtong translations in river on three " is so only illustrate the notion of intensified learning at this.

Fig. 9 represents the notion according to the control of intensified learning theory.610 pairs of control of control device object, 600 output control instructions 630.Control object 600 is according to control instruction 630 actions.At this moment, by the action according to control instruction 630, the state of control object 600 changes.Receive remuneration 620 from control object 600, this remuneration 620 is that the state that expression changes is that wish or undesirable for control device 610, and they are amounts of what degree.

In fact, the information that receives from the control object is the quantity of state of control object, and general control device 610 is based on this information calculations remuneration.It is big more generally to be set at approaching more desirable state remuneration, and it is more little to become undesirable state remuneration more.

Operate to control device 610 repetition tests, make remuneration reach the method for operating of maximum (that is, trying one's best) by study, thereby construct suitable operation (control) logic automatically according to the state of control object 600 near desirable state.

Band teacher learning theoretical needs that with the neutral net is representative provide successful case in advance as teacher's data, complicated in new complete set of equipments or phenomenon and occasion that can not prepare successful case in advance is not suitable for using.

Corresponding therewith, the intensified learning theory is categorized as the teacherless learning, self have this one side of ability that repetition test ground generates desired procedure, the advantage that has is also to be applied to control the not necessarily clear and definite occasion of Properties of Objects.

In this second embodiment, utilized this intensified learning theory.

Though intensified learning is the study of repetition test ground, but occasion in complete set of equipments control, in the danger or the aspects such as damage of complete set of equipments to manufacturing a product of running, it is difficult directly actual complete set of equipments is carried out as object that the operation of repetition test ground implements.So in the present invention, the running actual result generation service performance model from complete set of equipments adopts the mode of learning as object with this model.

The models that 290 pairs of intensified learning unit are generated by modelling unit 250, the simulated operation command signal 265 that output is made of the fuel flow rate of the air mass flow of each position of burner and follow-up air scoop, each burner.Simulated operation command signal 265 is corresponding to the operating condition of complete set of equipments, is set with upper lower limit value, amplitude of variation (pitch width) respectively, with the adoptable maximum changing amplitude of once-through operation.But each amount of simulated operation command signal 265 determines each numerical value at random in the scope of adopted value.

The model input simulated operation command signal 265 that 250 pairs of modelling unit have generated is calculated the NO as output data 255 _XAnd CO concentration.

Intensified learning unit 290 receives the output data 255 of modelling unit 250, calculates consideration value.

Consideration value defines with formula (5).Here, R is a consideration value, O _NOxBe NO _XValue, O _COBe CO value, S _NOxAnd S _COBe NO _XAnd the goal-setting value of CO, k1, k2, k3, k3 are positive constant.

[formula 5]

R=R ₁+ R ₂+ R ₃+ R ₄... formula (5)

${\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="H2009101264839E00031.png,H2009101264839E00061.png"\; state="\{\{state\}\}">R</figure-callout>}}_1=\left\{\begin{array}{c}{k}_1(O_{\mathrm{NOx}}\&le;S_{\mathrm{NOx}})\\ 0(O_{\mathrm{NOx}}>S_{\mathrm{NOx}})\end{array}\right.$

$R_2=\left\{\begin{array}{c}{k}_2(O_{\mathrm{CO}}\&le;S_{\mathrm{CO}})\\ 0(O_{\mathrm{CO}}>S_{\mathrm{CO}})\end{array}\right.$

$R_3=\left\{\begin{array}{c}{k}_3(S_{\mathrm{NOx}}-O_{\mathrm{NOx}})(O_{\mathrm{NOx}}\&le;S_{\mathrm{NOx}})\\ \begin{array}{cc}& 0(O_{\mathrm{NOx}}>S_{\mathrm{NOx}})\end{array}\end{array}\right.$

$R_4=\left\{\begin{array}{c}{k}_4(S_{\mathrm{CO}}-O_{\mathrm{CO}})(O_{\mathrm{CO}}\&le;S_{\mathrm{CO}})\\ \begin{array}{cc}& 0(O_{\mathrm{CO}}>S_{\mathrm{CO}})\end{array}\end{array}\right.$

As shown in Equation (5), NO _X, the CO value is than the occasion that goal-setting value has also descended, and gives remuneration R ₁And R ₂, having again, the occasion and this deviation that also descend than goal-setting value give remuneration pro rata.

In addition, the define method of remuneration it is also conceivable that other several different methods, is not limited to the method for formula (5).

Intensified learning unit 290 is because the combination of learning simulation operation instruction signal 265 is an operational ton, and makes the remuneration of calculating with formula (5) reach maximum, therefore can learn to reduce NO corresponding to present situation from the result _X, CO the combination of operational ton.

Intensified learning unit 290 reads in the service data 205 of current time under the state of study that is through with, output makes the remuneration of formula (5) be maximum operational ton 295 based on learning outcome.

260 pairs of basic operation command value 235 of correcting unit apply to be proofreaied and correct and exports as operation instruction signal 285.

Bearing calibration is identical with first embodiment basically.Corrective loop as shown in figure 13 is with the difference of first embodiment shown in Figure 11, uses the operational ton 295 that calculates with intensified learning unit 290 to replace on dummy instruction signal 265 these aspects.

Usually, the output of switch 282 is set at 1, with adder 284 deviation of basic operation command value 235 and operational ton 295 is added in the intensified learning command value 288 that becomes on the basic operation command value 235 basic operation command value 235.

Switch 286 selects intensified learning command value 288 as 285 outputs of operational order value usually.

But, in the input and output value of the input and output value of modelling unit 250 or intensified learning unit 290, there is any one to exceed the occasion of the limited field of upper lower limit value and rate of change, be set at as the output of switch 282 and select 0, and in switch 286, select basic operation command value 235 and output.

Thus, dual preventing because data exclusive disjunction loop unusual and the situation of output abnormality operational order value.

In addition, control result as intensified learning command value 288, the occasion that surpasses prescribed limit in control deviation, perhaps its frequency or duration surpass the occasion of prescribed limit, it is invalid to be judged as intensified learning command value 288, same selection (processing) in the time of can carrying out with data exception in switch 282 and 286 and stop the output of intensified learning command value 288.Even at this moment also can utilize basic operation command value 235 to continue into running, can not bring obstacle to the running of complete set of equipments.

According to above content, in second embodiment, can utilize intensified learning unit 290 to construct optimum manipulation method automatically.In addition, under the state of study that is through with, if the operating condition data 205 that input is current are because instantaneous output remuneration is maximum operational ton, therefore the combination that need not as first embodiment, regularly change the combination of simulated operation command signal 265 and obtain suitable operational ton in each control, the computer load in the time of can reducing control.Thus, owing to the stability raising of computer action, and the raising of the reliability of control device, therefore also have the completeness of complete set of equipments running, the stable also effect of raising.

Then, use Fig. 3 that the 3rd embodiment is described.

Be to possess state evaluation unit 300 and data group switch unit 310 with the difference of second embodiment.State evaluation unit 300 monitors the model error as the deviation of the calculated value 255 of the model that is generated by modelling unit 250 and the running real data 205 corresponding with it.

Use the processing procedure of Figure 12 description status evaluation unit 300 and data group switch unit 310.

In step 600, read in the setting value of the feasible value of relative model error.

In step 610, read in running real data 205 and operational ton actual value that will this moment and be input in the model and the calculated value 255 that obtains.

In step 620, calculate the deviation (model error) of the model calculated value 255 under the running real data 205 of reading in and the corresponding operating condition by step 610.

In step 630, to the time series data computation moving average in past of the model error that calculates by step 620, and moving average calculation at interval rate of change at the appointed time.

In step 640, relatively the rate of change value of the moving average of the model error that calculates by step 630, moving average and each constantly model error and to each feasible value of each value of reading in by step 600.

If end process then in permissible range, the occasion permissible range outside is switched indication 315 to data generating unit 210 output data groups, and indicates 316 to the change of modelling unit 250 output models.

Carry out from step 600 to step 630 by state evaluation unit 300, step 640 is undertaken by data group switch unit 310.

Fig. 5 represents the picture output example of the operation result of state evaluation unit 300.

In display frame 400, the time series preface chart of model error is presented in the chart zone 401.On chart, show each model error 408 and moving average 409 thereof constantly.

The longitudinal axis of chart is model error (%), can import the value range that shows at input field 403 (for example 0,100).Transverse axis is constantly, shows date on display field 404 or the moment.In addition, by mouse-based operation carriage release lever 402, transverse axis can change the demonstration time.

Selector button 405 during the demonstration be can use during the demonstration, a year unit, month unit, Zhou Danwei, day unit, hour unit selected.If any one button of selector button 405 during select showing with mouse, input window 410 during then showing, can specify demonstration during the zero hour.If just do not import whatever, then according to being to select demonstration zero hour during the demonstration selected of benchmark automatically with the current time of selecting by " OK " button.In addition, if press the Back button, then Shu Ru information is deleted.

Utilize this picture, can monitor the transition of model error by time series, which kind of state is the feasible value 407 that is easy to grasp the model relative model error in the use be in.The setting change of feasible value can be transferred to the setting picture by clicking " setting " button 406.

In addition, with the result of data group switch unit 310 computings, the occasion that surpasses permissible range is presented on the picture by way of caution.Have " model error surpasses feasible value ", " the display model error surpasses the duration of feasible value ", " the model error rate of change surpasses feasible value " in the warning.

If any warning occurs, then automatically show display frame 400, impel the operator to note.Meanwhile data group switch unit 310 is to data generating unit 210 output data groups switching commands 315, to modelling unit 250 output models change indication 316.

Especially in the occasion of " the model error rate of change surpasses feasible value ", " model error surpasses feasible value ", can think that the complete set of equipments characteristic sharply changes.In this occasion, the possibility that fuel performance changes with method change fuel composition data group that illustrates in first embodiment or generation new data set, is constructed model greatly once more.In addition, be that object carries out intensified learning with the model of constructing once more, automatically follow state variation.

Thus, can often monitor the tendency of model error automatically, because can be according to monitoring that the result carries out the study once more of model change and intensified learning, therefore always can keep stable control performance.

Fig. 6 is that expression is constructed model once more with data generating unit 210 change fuel data groups, and has estimated the result's of model error example.Shown model error about fuel composition A～D.As the benchmark of the data group selection of data generating unit 210, can calculating chart the mean value of 6 represented model errors, select this mean value to be minimum fuel composition data group.

Figure 10 is the display frame example 430 with the fuel composition data group of data generating unit 210 selections.On the epimere of picture, represent the composition of the data group selected with circular diagram 431 and table 433.

In addition, on the input picture of hypomere, the operator can import the fuel performance of actual use.In the occasion that multiple coal has been mixed, can on input field 435, import coal class title and cooperation ratio thereof.

This information can send to analysis center 30.As shown in Figure 8, analysis center 30 is connected with a plurality of power plants 50,51,52 by the dedicated communication line net, and the transmission that can carry out data mutually receives.

In the display frame example 430 of fuel composition data group, if by " sending to analysis center " button 434, then operator's information that will be input to input field 435 is sent to analysis center 30.And, will claim to send to analysis center 30 simultaneously by the data group name that data generating unit 210 is selected with nearest running real data (operational order value, process values).

Analysis center 30 is as if the above-mentioned information of 50～51 receptions from the power plant, the object complete set of equipments construction data of numerical analysis and the operational ton actual value that receives of just will burning is input to analytical model, to calculating NO when carrying out various changes as the fuel composition data of one of design conditions _XAnd CO concentration, the NO that selective reception is arrived _XAnd the error of CO concentration and measured value is minimum fuel composition data.

Sending the fuel composition data group of selecting like this and reach with this model from the model of constructing according to this data group to the object power plant by dedicated communication line net 40 is the learning outcome that object carries out intensified learning.

Control device 200 is if confirm this reception information, the fuel composition data group that just will newly receive is kept in the fuel data memory cell 270, the model that receives is placed modelling unit 250, the intensified learning result who receives is placed intensified learning unit 290.

The model construction method of analysis center 30 and intensified learning method are with identical in the method described in above-mentioned first and second embodiment.

According to above content,,, therefore can keep the high performance of control performance owing to can be updated to the higher data group of precision and use its model in the occasion that can obtain about the details of fuel change.

In this example, though carried out after receiving data from the power plant having been undertaken changing the numerical analysis of fuel composition data groups and model is constructed and intensified learning by analysis center 30, if but carry out the numerical analysis and the saving result that various changes are carried out in fuel composition of multiple situation in advance, the analysis result of preserving when then needing only the data that change according to the reception fuel performance is selected data group or model etc., can provide new model to the power plant immediately.Therefore, can shorten the time that control performance is descended owing to the fuel change.

In addition, before receiving the fuel performance data from the power plant, also can analyze various fuel performances in advance, and send new fuel composition data group, use model that it constructs and carry out the learning outcome of intensified learning with this model as object to the power plant successively, in the power plant, preserve these information in advance.

In this occasion, though might increase power plant one side's memory capacity, communication load, communications cost, but owing to when the coal class changes, need not to communicate by letter with analysis center 30, can change to new model or new intensified learning result rapidly, therefore can shorten the time that makes the risk of control performance decline owing to the change of coal class.

As effect, can be listed below according to the foregoing description.

Even fuel composition (character) changes, owing to can automatically suitably control exhaust gas composition, so can reduce the NO in the waste gas _X, harmful substance such as CO generation.

In addition, do not need between the long-term learning period, can just bring into play the effect of control device since the initial stage.Generally, for the running of the repetition test in the study,, owing to do not need the study of repetition test running, therefore can reduce the discharge rate of harmful substance in the present invention though might increase the harmful substance discharge rate.

Have again, owing to reduced the NO in the waste gas _XAmounts etc., the practical amounts such as ammonia use amount that therefore can cut down denitrification apparatus can also expect that the miniaturization of device or life of catalyst prolong.

In addition, owing to can follow the variation of fuel performance automatically, therefore except the adjustment working load that can reduce the operator, can also not rely on operator's experience or knowledge and realize suitable control, also have the advantage that improves the complete set of equipments maintainability.

In the above-described embodiments, though the control device of the complete set of equipments with boiler mainly has been described, this control device also can be used in the situation that control has the control object of burner.

Below, utilize illustrated embodiment to describe complete set of equipments control method and the device of realizing second purpose in detail.

Figure 14 is being applied in the present invention on the boiler plant of steam power plant, and the air mass flow of the boiler of the complete set of equipments 1001 that is made of boiler plant is supplied with in operation, study makes the CO concentration of a being discharged embodiment for the occasion of minimum control method, in this occasion, on complete set of equipments 1001, be provided with outside output interface 1002 and outer input interface 1003.

At first, outside output interface 1002 is from adder 1017 input signals, and to complete set of equipments 1001 outputs and the operation air mass flow.Also can possess as required and be used for keyboard or the display that the people operates.

Secondly, outer input interface 1003 inputs are from the signal of complete set of equipments 1001 outputs, output to running actual database 1005, basic controlling ordering calculation unit 1004 and operational ton arithmetic element 1015 as the service data of complete set of equipments.At this moment, also can possess as required and be used for keyboard or the display that the people operates.

Then, basic controlling ordering calculation unit 1004 is unit of exporting the basic controlling command signal of air mass flow of for example being used to be operable to complete equipment 1001 etc., is made of the control device that constitutes with general PID control logic.

In addition, in running actual database 1005, store from the service data of the complete set of equipments of outer input interface 1003 outputs.

Here, this service data refers to the process values of the state that is specified to complete equipment 1001, as for example shown in Figure 15, though be the NO of the relative air mass flow of expression _X, CO the data of relation of concentration, but except these NO that discharges from complete set of equipments _X, beyond the CO, also have CO ₂, SO _X, at least a amount or concentration in particulate subclass such as mercury, fluorine, coal dust or mist or the VOC.

Then, numerical analysis unit 1006 is the simulators that are modeled to the running of complete equipment, the analogy method that for example is to use patent documentation 3 to be put down in writing, operating conditions such as condition that is provided with the shape of the boiler of complete set of equipments 1001, the kind of coal etc. and air mass flow are that simulate on the basis, the CO concentration that produces when being calculated to be the complete equipment running etc.

As operating condition at this moment, except above-mentioned air mass flow, also have that air in the burner of fuel flow rate, air themperature, boiler distributes, at least a in the damper angle in parallel of the burner tilt of boiler or boiler.

And the analysis result of this numerical analysis unit 1006 provides as for example value of the CO concentration of relative air mass flow, and is stored in the numerical analysis database 1007.At this moment stored data are forms same with running actual database shown in Figure 15 1005.

Continuous model unit 1008 generates continuous model 1009 approx for being stored in the discrete air mass flow in the numerical analysis database 1007 and the relation of CO concentration, makes the variation of its relative parameter become continuous relation.Thereby this continuous model 1009 utilizes continuous model unit 1008 or continuous model amending unit 1012 to generate.

Then, the error of data that are stored in the numerical analysis database 1007 and the continuous model 1009 that is generated by continuous model unit 1008 is estimated in error assessment unit 1010.In addition, satisfy the occasion of certain condition in the error of utilizing error assessment unit 1010 to estimate, parameter is appended unit 1011 and is set the value of appending the air mass flow of carrying out numerical analysis, supplemental data point.And, utilize numerical analysis unit 1006 to carry out numerical analysis for the data point of appending.

On the other hand, continuous model amending unit 1012 uses the model of the service data correction continuous model 1009 in running actual database 1005.And control method unit 1013 serves as that the basis utilizes the intensified learning science of law to practise the method for operating of control flow with continuous model 1009.Learning outcome is stored in the learning outcome database 1014.At this moment, an example that is stored in the learning outcome in the learning outcome database 1014 is illustrated among Figure 16.

Then, operational ton arithmetic element 1015 is used from the service data and the learning outcome database 1014 of outer input interface 1003 outputs, the value of the air mass flow that calculating will be operated.For example, learning outcome is the occasion of Figure 16, if air mass flow is 0.45, then control air mass flow to be+0.05.The control signal of calculating outputs to subtracter 1016.

So the output signal of subtracter 1016 input basic controlling ordering calculation unit 1004 and the output signal of operational ton arithmetic element 1015 are calculated differing from and outputing to adder 1017 of these two kinds of signals.At this moment, be provided with switch 1018, utilize it that the output of subtracter 1016 is cut off from the input of adder 1017, can also carry out the only computing of the output of usefulness basic controlling ordering calculation unit 1004 as prior art.

In addition, the output signal of adder 1017 input basic controlling ordering calculation unit 1004 and the output signal of subtracter 1016, calculate these two kinds of signals and and output.Thereby, utilize these subtracters 1016 and adder 1017, can be used as the signal that the output signal of utilizing operational ton arithmetic element 1015 has been proofreaied and correct the output signal of basic controlling ordering calculation unit 1004 to the output signal of complete set of equipments.

Below, utilize the action of this embodiment of flowchart text of Figure 17.

In the embodiment of Figure 14, control CO concentration for the air mass flow of operating boiler, need know the variation of CO concentration of the variation of relative air mass flow.So, at first utilize numerical analysis to calculate the relation (step 1101) of air mass flow and CO concentration.Figure 18 represents an example of result of calculation at this moment.Point among the figure is that air mass flow is changed to 0.7 from 0.3, with 0.1 result calculated at interval.

At this moment analysis is preferably carried out with tiny interval as far as possible, but owing to the time that in the calculating of each point, will spend to a certain degree, so the data that in fact can only obtain dispersing.So, because to there not be part a little to carry out interpolation, thereby the continuous model that is similar to of generation, make the variation of its relative parameter become continuous relation (step 1102).

In the generation of continuous model at this moment, have the polynomial approximation utilized method, used neural network method etc.

Here, according to the characteristic of the dotted line of Figure 18, be the result who like this data point is generated approx continuous model, as long as just can be from estimating CO concentration continuously as the air mass flow that continuous model provided according to it.

Then, the data of utilizing the intensified learning method to construct to be kept in the learning outcome database 1014 (step 1103).

In this embodiment, owing to use learning outcome to be operable to complete equipment, so that control performance depends on the degree of the precision that is used for continuous model in study is bigger.

Here in order to improve precision, as long as will carry out tinyization of interval of the air mass flow of numerical analysis, the quantity that increases data point gets final product.But, as mentioned above, be difficult, so importantly select data point effectively to suppress the number of data point owing in the time that reality is used, will carry out the numerical analysis of many data points.For example in the occasion of Figure 18, can think near the importance height of the low air mass flow 0.5 of CO concentration data.So, by following procedure attachment data point.

Figure 19 is that expression utilizes numerical analysis to generate the detailed flow chart of continuous model, and the step 1201 here～step 1208 is equivalent to step 1101～step 1103 of Figure 17.

At first, in step 1201, set the point of the air mass flow of calculating with numerical analysis.Here the point of She Dinging is owing to be initial setting, so the quantity of the interval of point at this moment and point, as long as the precision of considering to analyze, required time etc., suitably setting gets final product in the excursion of air mass flow.

Then, in step 1202, use the numerical analysis unit to calculate the CO concentration of the point of each air mass flow.Result of calculation is stored in the numerical analysis database.

Then, in step 1203, discrete data are carried out interpolation, generate continuous model, afterwards in step 1204, the interim continuous model that calculates the occasion of having removed arbitrary data point, the error of the data point of in step 1205, having calculated and having removed.

Figure 20 is an example of result of calculation at this moment, in this occasion, the error of the continuous model that generates owing to the point of removing air mass flow 0.5 and the point of air mass flow 0.5 increases, distinguish that thus near the point this air mass flow 0.5 is bigger to the influence of continuous model, conclude in this part and need calculate with tiny interval.

So, after step 1205, this error and certain threshold value are compared (step 1206), surpass the occasion of this threshold value in error, be judged as near the importance height of the data of this point, make tinyization (step 1207) at interval to supplemental data point between the point of front and back.And, to this data point of having appended, use the numerical analysis unit to calculate CO concentration, calculate continuous model once more.

By carry out step 1202 so repeatedly～1207 diminishing supplemental data point effectively up to error.

Here, Figure 21 be expression by step 1207 supplemental data point, generated the data example figure of the occasion of continuous model once more, can generate the high continuous model of precision when being appreciated that the quantity of the point that can suppress to carry out numerical analysis by above process.And, afterwards, advance to step 1208 or step 1103, use the continuous model that is generated to carry out intensified learning.

But, because the model that obtained by said process has used The numerical results, inevitably and the complete set of equipments of reality between have error, so preferably use the service data of complete set of equipments actual machine as far as possible.

So, if the initial continuous model that is generated by numerical analysis that uses carries out intensified learning, obtain the service data of complete set of equipments actual machine, then service firing data correction model is relatively good.

Thereby, in this embodiment,, be object learning manipulation method once more with this correction model by following process correction continuous model, therefore, be provided with the step 1104 of Figure 17, utilize continuous model amending unit correction continuous model here.For example, utilizing numerical analysis to generate the service data that comprises the complete set of equipments actual machine in the data of continuous model, generate continuous model once more.At this moment because preferential service firing data, therefore can be in advance in addition suitable attention.

Figure 22 appends service data and an example of revising the occasion of continuous model like this, compares with the continuous model of Figure 21, has reflected the tendency of the service data of complete set of equipments actual machine as can be known.For example, near the air mass flow that has obtained service data, become the continuous model of the tendency that has reflected service data.

In addition, even, also can generate continuous model, so can obtain the higher model of precision based on the result of numerical analysis owing to there be not the part of continuous data.

Then, in step 1105, use the continuous model revised to carry out intensified learning.In learning once more, understood the method that partly relearns, according to this method, can construct the control model with less load.

And the process of these steps 1104～1105 is all carried out when appending the complete set of equipments data at every turn.Thus, the continuous model based on the numerical analysis result can be modified to gradually the model consistent, and the model of control air mass flow can be modified to the model consistent with the characteristic of complete set of equipments actual machine with the characteristic of complete set of equipments actual machine.

So, use the result of this intensified learning to calculate the operational ton (step 1106) of air mass flow, then to complete set of equipments 1001 output function signals (step 1107), the air mass flow that control will be supplied with to the boiler of complete set of equipments 1001, and at this moment pass through said process, owing to can be that the basis is controlled with the analysis result according to the complete set of equipments test running, can shorten up to the time of introducing, can follow the accumulation of service data to be modified to the characteristic of complete set of equipments actual machine, suppress control so can obtain the good CO of performance.

Thereby, according to above-mentioned embodiment, before the complete set of equipments running, can utilize the numerical analysis result to construct the model that is used to learn, thereby can shorten between the introductory phase of complete set of equipments control, never the stage that fully rolls up complete equipment running real data begins to bring into play the specification specification performance, the control of the CO that can obtain being scheduled to.In a word, not only can be corresponding fuel between break-in period or the more consumption of raw material, but also can the corresponding influence that causes by effluent to environment from complete set of equipments.

In addition,, utilize numerical analysis result's error assessment generating run conditional parameter,, reduce the error of model, just can carry out the good control of performance by appending the numerical analysis data according to above-mentioned embodiment.At this moment, by using complete set of equipments service data correction model, carry out intensified learning once more, thereby always can carry out the good control of performance.

Therefore, be applied to by the invention with above-mentioned embodiment on the boiler of steam power plant, just can reduce increases NO _X, carrying capacity of environment material such as CO the risk of generation.

The present invention is owing to during from the complete set of equipments test running, just can utilize the numerical analysis result to be controlled to complete equipment, so can shorten between the introductory phase of control device.In addition, by utilizing the error assessment generating run conditional parameter of numerical analysis, append the numerical analysis data and reduce the error of model, and use complete set of equipments service data correction model, carry out intensified learning once more, thereby always can carry out the good control of performance.

In addition, on the boiler that applies the present invention to steam power plant, can reduce increases NO _X, carrying capacity of environment material such as CO the risk of generation.

Claims (7)

1. complete set of equipments control method, the employed control device of this complete set of equipments has the learning functionality of the relation of learning manipulation amount and complete set of equipments state, and have and utilize this learning functionality to calculate function corresponding to the operational order value of complete set of equipments state, it is characterized in that

Utilize flowing and the process values of the state of the numerical analysis of reacting phenomenon and computational rules complete set of equipments of a plurality of operating conditions,

The process values of each operating condition is generated continuous model approx, makes the variation of its relative operating condition parameter become continuous relation,

Use generates continuous model once more by the service data of numerical analysis calculation process value and complete set of equipments actual machine,

Use the continuous model that generates once more to learn.

2. complete set of equipments control method according to claim 1 is characterized in that,

Evaluation by above-mentioned numerical analysis calculation process value with utilize the error of the calculated value of above-mentioned continuous model,

Generate new operating condition parameter based on the numerical analysis result,

Use the new operating condition parameter that generates to carry out above-mentioned numerical analysis once more and calculate new process values,

Use generates continuous model once more by the old process values of above-mentioned numerical analysis calculating and new process values.

3. complete set of equipments control method according to claim 1 is characterized in that,

In the process values of the state of stipulating above-mentioned complete set of equipments, use from the particulate subclass of complete set of equipments discharge or at least a amount or the concentration of VOC,

In the aforesaid operations condition, use that air in the burner of air mass flow, fuel flow rate, air themperature, boiler distributes, at least a in the damper angle in parallel of the burner tilt of boiler or boiler.

4. complete set of equipments control method, the employed control device of this complete set of equipments has the successive value model of the characteristic that is modeled to complete equipment, and the method for operating of using the above-mentioned complete set of equipments of this successive value model learning, and based on the function of result's calculating of learning corresponding to the operational order value of complete set of equipments state with this learning functionality, it is characterized in that

Utilize flowing and the process values of the state of the numerical analysis of reacting phenomenon and computational rules complete set of equipments of a plurality of operating conditions,

The process values of each operating condition is generated continuous model approx, makes the variation of its relative operating condition parameter become continuous relation,

Use generates above-mentioned continuous model once more by the service data of numerical analysis calculation process value and complete set of equipments actual machine.

5. a complete set of equipments control device has the learning functionality of the relation of learning manipulation amount and complete set of equipments state, and has and utilize this learning functionality to calculate function corresponding to the operational order value of complete set of equipments state, it is characterized in that having:

Be stored as the running actual database of process values of the operating condition of complete equipment;

Storage is by the numerical analysis database of the numerical analysis calculation process value of the mobile and reacting phenomenon of the operating condition of complete set of equipments;

From above-mentioned numerical analysis database the process values of each operating condition is generated continuous model approx, make the variation of its relative operating condition parameter become the continuous model unit of continuous relation; And,

The process values that appends above-mentioned running actual database also generates the continuous model amending unit of continuous model once more,

The function that above-mentioned learning functionality is to use the continuous model that utilizes above-mentioned continuous model amending unit to generate once more to learn.

6. complete set of equipments control device according to claim 5 is characterized in that possessing:

Evaluation is by the unit of numerical analysis calculation process value with the error of the calculated value that utilizes above-mentioned continuous model;

Generate the unit of new operating condition parameter based on the error assessment result;

The new operating condition parameter that use has generated is carried out numerical analysis once more and is calculated the unit of new process values; And,

The unit that uses old process values that above-mentioned continuous model unit calculates by numerical analysis and new process values to generate continuous model once more.

7. complete set of equipments control device according to claim 5 is characterized in that,

In the process values of the state of stipulating above-mentioned complete set of equipments, use from the particulate subclass of complete set of equipments discharge or at least a amount or the concentration of VOC,

In the aforesaid operations condition, use that air in the burner of air mass flow, fuel flow rate, air themperature, boiler distributes, at least a in the damper angle in parallel of the burner tilt of boiler or boiler.

Images