CN101859102B - Gas concentration concluding method of boiler equipment and gas concentration concluding apparatus - Google Patents

Abstract

This invention provides a gas concentration concluding method of a boiler equipment, making use of a neural network to conclude the concentration of the gas component exhausted from the boiler equipment, and making use of the neural network used by a plurality of neural networks prepared for each coal category and the neural network for judging the coal category proportion to conclude. The invention further provides a gas concentration concluding apparatus of the boiler equipment, making use of the neural network to conclude the concentration of the gas component exhausted from the boiler equipment, a, comprising a summate which makes use of the neural network used by a plurality of neural networks prepared for each coal category, the neural network for judging the coal category proportion and the mixed proportion concluded by the neural network for judging the coal category proportion to do weighted average to the concluded value of the gas concentration exported a plurality of neural networks prepared for each coal category.

Description

The gas concentration estimating method of coal-burning boiler and gas concentration apparatus for predicting

The application is that application number is 2007101964803, the applying date is on Dec 5th, 2007, denomination of invention is divided an application for the application of " control device of boiler plant and gas concentration apparatus for predicting ".

Technical field

The present invention relates to the control device of boiler plant (boiler plant).In addition, the present invention relates to utilize neural network (neural net), infer concentration, particularly CO concentration, the NO of gas componant of the coal-burning boiler of one of inscape as firepower equipment _XThe method and apparatus of concentration.

Background technology

In the field of equipment control, be controlled to be basic steering logic with PID is main flow all the time.In addition, also propose to have a lot by the learning functionality with teaching take neural network as representative, can tackle flexibly the technology of the characteristic of equipment.

In order to utilize the learning functionality with teaching to consist of control device, owing to be necessary to prepare in advance the successful example that becomes the teaching data, therefore the learning method without teaching has been proposed also.

As the example without teaching study, the intensified learning method is arranged, but this intensified learning method is the framework of study control, by the repeatedly experimental interaction with environment such as control object, generation makes the measuring-signal that obtains from environment reach expectation value to the operation signal of environment.

This learning method can't be prepared successful example even have in advance, but if pre-defined expectation state, the advantage of the Expectation of Learning action that just can automatically conform.

In intensified learning, have learning functionality, the evaluation of estimate (intensified learning, being called as remuneration) of the scale that the measuring-signal that obtains from environment with utilization calculates is clue, generation is to the operation signal of environment, and the expected value of the evaluation of estimate that obtains till making from the present condition to future reaches maximum.

As the method for this kind of actual installation learning functionality, have Actor-Critic, Q study, the real time Dynamic Programming scheduling algorithm of for example middle narration of the intensified learning of technical literature (Reinforcement Learning).

In addition, as the framework of the intensified learning that develops said method, introducing in the above-mentioned technical literature has the framework that is called the Dyna-structure.This be the model of simulating control object be object, what kind of operation signal in advance study should generate, and utilizes this learning outcome to determine the method for operation signal that control object is applied.

In addition, as the technology of applicable intensified learning, enumerated the described technology of TOHKEMY 2000-35956 communique.This technology is such technology: possessing in advance a plurality of groups with system of model and learning functionality is the intensified learning module, model in each intensified learning module and the predicated error of control object are less, try to achieve the responsibility signal of getting larger value, be weighted for the operation signal to control object that generates from each intensified learning module pro rata with this responsibility signal, determine the operation signal that control object is applied.

Coal-burning boiler uses coal to make fuel.In having the steam power plant of coal-burning boiler, the environmental pollutants CO, the NO that from boiler, discharge _XConcentration be necessary for below the setting.CO and NO _XThe generation the relationship between quantities is opposite, if air (or oxygen) surplus, then NO _XThe generation quantitative change many; If opposite lack of air then the generation quantitative change of CO is many.In recent coal-burning boiler, in order to reduce CO and NO _XBoth and raising burning efficiency carry out sending into two sections burnings of air interimly.In this burning control, carry out the adjustment of air capacity, the selection of burner combustion pattern etc., make only fired state.Be used for the adjustment of optimization burning control, for example, the plan of the adjustment of ride gain, burner combustion pattern etc. are implemented when off-line in advance.

But however, the burning condition of adjusting in advance is only for representational operation mode, only the operating plan of cardinal principle.With respect to this, from the viewpoint of economy, require corresponding to operating conditions such as the burden requirement value that constantly changes, atmospheric conditions, fuel characteristics, equipment operation should reach optimization, that is, and control CO, NO _XConcentration in tolerance band, make simultaneously burning efficiency maximization.

In order to realize these, need to be take present operating condition as the basis, can online simulation CO, NO _XRelative concentration requires the variation of change in control.That is, with respect to the present operating condition that obtains from measurement data, the burning efficiency when evaluation change control requires and the discharge rate of carrying capacity of environment material need with good grounds 2 functions of exploring optimum control points.

As CO, NO _XThe estimating method of concentration has based on the learning-oriented algorithm such as neural network and so on, utilizes the real machine data each operating condition and gas concentration to be changed the method for the relationship modeling between the tendency.As long as when using the method the real machine data are arranged, just can generate simply CO, NO corresponding to this practical characteristic _XTherefore the Inference Model of concentration is one of often applicable method.In addition, utilize in addition mode inference CO, NO _XThe method of concentration (for example, referring to Patent Document 2).

Non-patent literature 1: intensified learning (Reinforcement Learning), loyal virtue refined Zhang Gongyi in river all on three, Co., Ltd. is published in gloomy north, and on Dec 20th, 2000 published

Patent documentation 1: TOHKEMY 2000-35956 communique

Patent documentation 2: TOHKEMY 2006-132902 communique (summary)

But, carry out manufacturing course or as automobile, train, aircraft etc., be difficult to carry out in the situation of repeatedly experimental running with the control object of reality being difficult to, for example TOHKEMY 2000-35956 communique is described, take the simulation control object model as object, the method for in advance learning is effective.

But, in the described technology such as TOHKEMY 2000-35956 communique, be difficult to the movement of the control object of complete simulation reconstruction reality with model, between model and working control object, there is model error generally speaking.And, have movement, its process of actual control object more complicated, the tendency that model error is larger.

When model error becomes large, owing to model and the movement of the control object of reality are different results, carried out the in advance result of study even therefore adopt take this model as object, also might can not get envisioning such effect of improving even carry out control operation.That is, exist and become the danger that learning outcome exists with ... the locally optimal solution of model characteristics.

Also have, the described technology such as TOHKEMY 2000-35956 communique for security and stability running control control object, are obtained the not open and hint of optimum solution simultaneously.

Summary of the invention

The object of the present invention is to provide a kind of control device of boiler plant, even between the movement of the model of simulating control object and working control object, exist in the situation of error, also the danger that is absorbed in locally optimal solution can be got rid of as far as possible, in security and stability running control control object, optimum solution can be obtained.

In addition, the CO, the NO that discharge from boiler _XConcentration is even under identical burning condition, also because the difference of coal class (hereinafter referred to as coal) and difference.As the utilization form of generating plant, just at last in the running, the situation of the coal type change that is sent to grinding machine (comminutor) is arranged also, at this moment, in the short time behind the coal type change, be in the state that mixes with coal after changing before changing.That is the transition state when, existence switches to coal B from coal A.

Owing to the optimization that also requires for this kind operating condition to turn round, so CO, NO _XThe Inference Model of concentration, not only for whole coals, the condition of mixing for multiple coal also has high requirement of inferring precision.Therefore, infer CO, NO _XThe neural network of concentration is not only wanted corresponding whole coals, also will consider the situation of the transition state that changes coal, even also must be able to tackle under the state that at least two kinds of coals are mixed.

But, making the neural network precision tackle well these all various states and have limit, its result exists the problem of the deduction precise decreasing of model.

Further, the problem that also has the coal switching that is difficult in the in-situ measurement equipment running.The information that obtains as the moment of switching about coal is owing to just stored the coal of other kinds, so the information of analogizing that can switch and so on of general horse back just in coal storage equipment.Be not only the blending ratio of measuring the coal under the transition state that described such multiple coal mixes, even judge that when to switch to new coal all very difficult.

Therefore, utilizing the real machine online data to infer CO, NO _XDuring concentration, the problem of existence is: be difficult to obtain online the information relevant with the coal of the fuel that drops into.

Another object of the present invention provides a kind of gas concentration estimating method and device of coal-burning boiler, for each coal, comprises that transition state can infer CO, NO accurately _XConcentration.

The control device of boiler plant of the present invention, the measurement data of the quantity of state of its input boiler plant is service data, device model based on the equipment operation characteristic of simulating boiler plant, calculating is to the running operation command value of boiler plant, turn round, it is characterized in that described control device has: the running experience database, it stores the service data in the past of boiler plant; The operating condition decision maker, it is according to the service data that stores in this running experience database, and at least one party's variation is inclined in calculating operation amount or the process value, and changes the state of trend information and load variations command value judgement boiler plant from this; The model error evaluating apparatus, it calculates the error by the measured value of the process value of device model calculating and boiler plant; The service data modeling device, it will be included in the process value of data in past of running experience database storage and the mutual relationship modelling between operational order value or the operational ton; Explore the some determination device, its utilize the error amount that calculated by the model error evaluating apparatus, with the result of calculation of the process value of pass through the calculating of service data model that is generated by the service data modeling device at least one party, calculating operation instruction candidate value or operational ton candidate value; The operational order determination device, it utilizes by the operating condition information of exploring an operational order candidate value of determination device calculating or operational ton candidate value and being judged by the operating condition decision maker, determines to control object to be the operational order value of the manipulater output of boiler plant.

In addition, the invention provides a kind of gas concentration estimating method of coal-burning boiler, it utilizes neural network to infer from the concentration of the gas componant of coal-burning boiler discharge, it is characterized in that, utilizes in a plurality of neural networks of each coal preparation and the neural network of coal ratio judgement usefulness and infers.

In addition, the invention provides a kind of gas concentration apparatus for predicting of coal-burning boiler, utilize neural network to infer from the concentration of the gas componant of coal-burning boiler discharge, it is characterized in that, it has totalizer, described totalizer utilization is judged the blending ratio that the neural network of usefulness and the neural network that described coal ratio is judged usefulness are inferred by a plurality of neural networks of preparing in each coal, coal ratio, and the inferred value of the gas concentration of described a plurality of neural networks outputs of preparing by each coal is weighted on average.

In the present invention, prepare neural network by each coal, only become privileged to carry out CO, NO for this coal _XConcentration is inferred the simulation of usefulness, improves and infers precision.And consider the admixture of coal, also prepare simultaneously the neural network for the blending ratio of inferring each coal of now using.The coal ratio of inferring based on this neural network is to CO, the NO of the neural network output of each coal _XThe inferred value of concentration is weighted on average, as final inferred value.

According to the present invention, even between the movement of control object of the model of simulation control object and reality, exist in the situation of error, also can get rid of the danger that is absorbed in locally optimal solution as far as possible, realize with security and stability running control control object, simultaneously realization can be explored the control device of the boiler plant of optimum solution.

In addition in the present invention, neural network is as long as become privileged and simulate just passablely each coal, so the deduction precision of each neural network improves, finally CO, NO _XThe deduction precision of concentration also can improve.In addition, the information that is difficult to the coal that obtains from measured value also can be analogized by neural network, therefore also can obtain correct CO, NO when changing coal _XConcentration.

Description of drawings

Fig. 1 is the pie graph of control device of the boiler plant of expression one embodiment of the present of invention;

Fig. 2 is the skeleton diagram of structure of steam power plant of the boiler plant of the expression embodiment of the invention that possesses Fig. 1;

Fig. 3 is the skeleton diagram of intensified learning concept of control device of the embodiment of the invention of key diagram 1;

Fig. 4 is the detailed structure view of the intensified learning device that possesses of the correcting device in the control device of the embodiment of the invention of presentation graphs 1;

Fig. 5 is the process flow diagram of the operating conditions exploration order in the control device of the embodiment of the invention of presentation graphs 1;

Fig. 6 is expression CO, NO _XThe measurement data of gas concentration and the key diagram of the relation of model characteristics;

Fig. 7 is that the operating parameter during the expression operating conditions is explored is set and the picture example of the picture that permissive condition is set;

Fig. 8 is that CO, the NO in the boiler plant measured in explanation _XThe skeleton diagram of gas concentration measuring position of gas concentration;

Fig. 9 is the display frame example of the example of the example that distributes of the gas concentration of expression boiler plant and measurement data;

Figure 10 is the example of the gas concentration distribution trend expression of expression boiler plant, and the key diagram of the example of the situation of the burning of boiler plant control;

Figure 11 is the structural drawing of relevant online gas concentration apparatus for predicting of the present invention;

Figure 12 is the structural drawing of the gas concentration deduction section in the gas concentration apparatus for predicting of the present invention;

Figure 13 is that the coal ratio in the gas concentration apparatus for predicting of the present invention is judged the structural drawing with neural network;

Figure 14 is the figure of the display case of expression display device.

Among the figure,

The equipment of 100-control object; The 101-boiler; The 102-burner; Air scoop behind the 103-; The 104-air heater; The 200-control device; The 220-IO interface; The 221-input-output unit; 230-basic controlling Command Arithmetic Unit; 240-running experience database; The 250-correcting device; 260-model error evaluating apparatus; 263-service data modeling device; 265-operating condition decision maker; 270-explores the some determination device; 280-operational order determination device; 290-intensified learning device; The 291-modeling device; The 292-learning device; 296-numeric value analysis database; 400-numeric value analysis calculation element; The online gas concentration apparatus for predicting of 701-; The 702-display device; 711-process value database; 712-gas concentration deduction section; 713-operating condition configuration part; The 714-display control unit; 721-coal ratio judgement neural network; 722-coal A infers neural network with gas concentration; 723-coal B infers neural network with gas concentration; 731-coal ratio is judged the output unit with neural network; The 741-totalizer.

Embodiment

About the control device of the boiler plant of the embodiment of the invention, referring to description of drawings.

Embodiment

Fig. 1 is the control block diagram of control device of the boiler plant of expression one embodiment of the present of invention;

In Fig. 1, the control device 200 take the boiler plant of steam power plant as this equipment 100 of control object for example, receive the metrical information 205 of various process value from the equipment 100 of control object, use these metrical informations 205, the calculating of programming in advance in control device 200 is to equipment 100 transmit operation command signals (control signal) 285.

Equipment 100 is according to the operation instruction signal 285 that receives from control device 200, by making operating machines that equipment 100 possesses, the actuator actions such as the aperture of valve of fluid flow of device interior or the vibroshock aperture of flowing through such as adjusting come the state of opertaing device 100.

In addition, control device 200 is according to the load command signals 51 from 50 receptions of centre feed instruction institute, calculate by basic controlling Command Arithmetic Unit 230, will deliver to equipment 100 by this operation instruction signal that calculates (control signal) 285, the generating output of opertaing device 100.

The control device of the boiler plant of present embodiment is the example of burning control that is applicable to consist of the boiler plant of steam power plant.In the present embodiment, special instruction is applicable to reduce CO and NO in the waste gas _XConcentration is the example of control function of control device of the boiler plant of purpose.

Fig. 2 represents to consist of the schematic configuration as the boiler plant of the steam power plant of control object.

In Fig. 2, bunker coal is pulverized in comminutor 110, and the auxiliary air as coal dust is adjusted usefulness with primary air and the burning of coal carrying usefulness drops into boiler 101 by burner 102, at the stove internal-combustion bunker coal of boiler 101.

Bunker coal and primary air are imported into burner 102 from pipe arrangement 134, and auxiliary air is imported into burner 102 by pipe arrangement 141.And, the rear air (after air) of two sections burning usefulness is dropped into boiler 101 by rear air scoop 103.Should be imported into rear air scoop 103 from pipe arrangement 142 by rear air.

Bunker coal is at the high-temperature combustion gas of the stove internal-combustion generation of boiler 101, in the stove of boiler 101 along route shown by arrows side flow downstream, discharge from boiler 101 after becoming burnt gas, flow down to the air heater 104 in boiler 101 outer setting.

Passed through the burnt gas of air heater 104, in not shown emission-control equipment, removed afterwards the objectionable impurities that contains in the burnt gas, then discharged to atmosphere by chimney.

In addition, the part of the burnt gas that flows down from the stove of boiler 101 is directed by gas recirculation system 701 from the downstream of stove, flows in the stove from the stove bottom of boiler 101, is formed in boiler 101 and carries out recycle.The recirculation volume of this burnt gas is regulated by the flow rate regulating valve that arranges at gas recirculation system 701.

Water supply in boiler 101 circulations, be imported into boiler 101 from the not shown condenser that arranges at turbine 108 through make-up pump 105, heat exchanger 106 places on the stove that is arranged at boiler 101, the burning gases that flowed down by the stove inside at boiler 101 heating and become the steam of High Temperature High Pressure.

And in the present embodiment, although the quantity of illustrated heat exchanger 106 is 1, also can dispose a plurality of heat exchangers.

The steam of the High Temperature High Pressure that produces at heat exchanger 106 places is directed to steam turbine 108 by turbo-regulator 107, and the energy drives steam turbine 108 that utilizes steam to have makes generator 109 rotary electrifications.

Then, the path that the stove of boiler 101 of primary air and auxiliary air drop in to(for) the burner 102 that arranges from the stove at boiler 101, the rear air scoop 103 that arranges from the stove at boiler 101 are put into the rear air in the stove of boiler 101 describes.

Primary air is directed to pipe arrangement 130 from fan 120, branch into halfway by the pipe arrangement 132 of air heater 104 and the pipe arrangement 131 by air heater 104 not, the primary air that flows down at these pipe arrangements 132 and pipe arrangement 131 collaborates to be directed to comminutor 110 in pipe arrangement 133 again.

By the air of air heater 104, the burnt gas heating of being discharged from the stove of boiler 101.

Utilize this primary air that the coal (coal dust) that comminutor 110 generates is carried to burner 102 by pipe arrangement 133.

Auxiliary air and rear air are directed to pipe arrangement 140 from fan 121, after by air heater 104 heating, branch into the pipe arrangement 142 that pipe arrangement 141 that auxiliary air uses and rear air are used, be directed at respectively the burner 102 and the rear air scoop 103 that arrange on the stove of boiler 101.

The control device of the boiler plant of present embodiment is in order to reduce CO and NO _XConcentration has to adjust from burner 102 and drops into the air capacity of boiler 101 and the function of putting into the air capacity of boiler 101 from rear air scoop 103.

As shown in Figure 1, the control device 200 of boiler plant is made of following part: basic operation Command Arithmetic Unit 230; Correcting device 250, its change or correction are from the basic operation command value 235 of basic operation Command Arithmetic Unit 230 outputs; Running experience database 240, its storage are taken in the running experience data that consist of by process measurement value 205, operator's input signal, from command signal of upper control system etc.; IO interface (input and output I/F) 220, it is used for carrying out the transmitting-receiving of data with control object equipment 100 or operator etc.; Input-output unit 221, it is used for the operator and watches various data, the operational order when perhaps being used for input setting value or operation mode, manual operation etc.

Basic operation Command Arithmetic Unit 230 is with PID (ratio, integration, differential) controller is as basic constituent element, according to from the centre feed instruction the 50 load command signals 51 that receive, will be by equipment 100 detected process measurement values 205 through input and output I/F220, operator's input signal, calculate as input from command signal of upper control system etc., calculate and output basic operation command value 235, this basic operation command value 235 is for arranging at equipment 100, the valve that the fluid flow of the device interior of flowing through is regulated, vibroshock, the command value of the exercises machines such as motor.

In the control device of the boiler plant of present embodiment, as shown in Figure 1, possess the correcting device 250 that changes or revise basic operation command value 235 at control device 200, the below describes this correcting device 250.

Correcting device 250 is made of following: intensified learning device 290, model error evaluating apparatus 260, service data modeling device 263, operating condition decision maker 265, exploration point determination device 270 and operational order determination device 280.

The intensified learning device 290 that correcting device 250 possesses has following function: utilize and be stored in service data 245, the measurement data 205 in the running experience database 240 and the numeric value analysis result 401 that is calculated by numeric value analysis device 400, by the intensified learning theory study suitable method of operating corresponding with equipment state.

Illustrated that numeric value analysis device 400 is located at the structure of control device 200, but when as grid computing machine (grid computer), supercomputer, requiring to have large calculation of capacity ability, also numeric value analysis device 400 and control device can be arranged in 200 minutes.

The detailed description of intensified learning theory, owing to for example in the described technical literature " intensified learning (Reinforcement Learning); loyal virtue refined Zhang Gongyi in river all on three; Co., Ltd. is published in gloomy north; on Dec 20th, 2000 published " detailed narration is arranged, only the concept of intensified learning is described herein.

Fig. 3 represents the concept of the control of above-mentioned intensified learning theory.

The 600 output function instructions 630 of 610 pairs of control objects of control device.Control object 600 is moved according to steering order 630.At this moment, by the action based on steering order 630, the state of control object 600 changes.

Control device 610 receives remuneration 620 from control object 600, and remuneration 620 is whether the state of expression variation is preferred for control device 610, and these are amounts of what degree.

In fact the information that receives from control object 600 is the quantity of state of control object, generally is that control device 610 calculates remuneration based on this quantity of state.Generally be set as, more larger close to the expectation state remuneration, the more approaching state remuneration of not expecting is less.

Control device 610 is by repeatedly carrying out experimental operation, and the method for operating that the study remuneration reaches maximum (that is, as far as possible close to expectation state) is constructed suitable operation (control) logic automatically corresponding to the state of control object 600.

The band teaching theories of learning take neural network as representative are necessary that Provision in advance success example as the teaching data, is not suitable for the situation that new equipment does not have service data, and because of the complicated situation that can not prepare in advance successful example of phenomenon.

With respect to this, described intensified learning theory is classified as without teaching study, on the viewpoint that has automatically repeatedly the ability of experimental generation preferred operations, has and also goes for the not advantage of very clear and definite situation of control object characteristic.

But, in order only to learn by the service data of equipment, owing to needing wait until store the required service data of enough study, so need long period competence exertion effect.And because be repeatedly experimental study, running that also might equipment becomes the state that is not supposed to, and is not only not to be supposed to, even sometimes also can brings danger to the device security aspect.

Therefore, in the control device 200 of the control device of the boiler plant that consists of present embodiment, the model of learning in advance to simulate control object is that what kind of operation signal object should generate for well.

The numeric value analysis device 400 that control device 200 possesses based on the boiler construction analog machine characteristic of equipment 100, utilizes the numeric value analysis gimmicks such as method of difference, finite volume method, Finite element method, calculates its burning (reaction), gas flow, conduction process.

The numeric value analysis gimmick of present embodiment does not have feature, because do not exist with ... the parsing gimmick, so omission is about the explanation of Numerical method.

Numeric value analysis by numeric value analysis device 400 calculates the phenomenon under the various operating conditionss, about as CO concentration or NO in the waste gas of the device characteristics of equipment 100 _XConcentration is calculated respectively both concentration in these measuring positions.

At the CO of measuring position concentration or NO _XThe result of calculation of concentration is calculated owing to calculate grid (grid) in each of its cross section, therefore with it in the measuring position be boiler export flowing path section about, front and back, be divided into three kinds of whole zones, calculate respectively its mean concentration.

As the service performance of the numeric value analysis that passes through described numeric value analysis device 400 with the equipment of the modeling of described equipment, with CO concentration or the NO in the waste gas _XConcentration is that above-mentioned explanation has been carried out in representative.But except this CO or NO _XIn addition, also can be for the C0 in the waste gas ₂, SO _X, the particulate subclass that consists of of Hg (amount of mercury), fluorine, coal smoke or mist, at least a concentration among the VOC (volatile organic compounds) simulate.

Fig. 4 represents the structure of the intensified learning device 290 that the correcting device 250 of control device 200 of the boiler plant of present embodiment possesses.Result of calculation 401 by numeric value analysis device 400 calculates is stored in the numeric value analysis database 296 that consists of intensified learning device 290.

Consist of the modeling device 291 of intensified learning device 290, read in necessary data 297 from numeric value analysis database 296, calculate CO concentration, NO _XThe mean value of concentration.

This modeling device 291, with the mean concentration calculated as the teaching signal, will be at that time operating conditions as input signal, in the neural network that is consisted of by input layer, middle layer, output layer, utilize error Back-Propagation method (back propagation method) study input/output relation.

The formation of described neural network and learning method are general methods, and these methods also can be other methods, because present embodiment does not exist with ... formation or the learning method of neural network, so detailed herein.

Consist of the learning device 292 of intensified learning device 290, before equipment 100 runnings (before the test working), take based on by the device characteristics of the numeric value analysis of this neural network learning as object, study suppresses CO, NO _XThe method of operating of generation.

The operation signal 293 suitable with the operational ton of exporting from learning device 292 is burner 102 and afterwards air mass flow, the air mass flow of each burner and the generator output of generator 109 of each position of air scoop 103 that arranges at boiler 101.

In the present embodiment, to described fuel flow rate, air mass flow, generating output and CO and NO _XThe relation of concentration has been carried out modelling, but the cuit of present embodiment and output project are not limited to this.

In addition in the present embodiment, modeling method also is not limited to neural network, also can use other statistical models such as regression model.

Consist of the learning device 292 of intensified learning device 290, to the model that is generated by modeling device 291, the input data 293 that the air mass flow of the burner 102 that output arranges at boiler 101 and each position of rear air scoop 103, the fuel flow rate of each burner consist of.

From the input data 293 of the learning device 292 output operating conditions corresponding to equipment 100, set respectively upper lower limit value, amplitude of variation (scale amplitude), the maximum changing amplitude that single job can obtain.Each amount of these input data 293 determines each numerical value at random in can the scope of the value of obtaining.

In the modeling device 291 of intensified learning device 290, to generating complete mode input input data 293, in this modeling device 291, calculate NO _XConcentration and CO concentration are as 294 outputs of output data.

The learning device 292 of intensified learning device 290, the counter that possesses constitutional formula (1) calculating formula of calculating consideration value, by receiving the output data 294 from modeling device 291, carry out the calculating of formula (1) by described counter, calculate consideration value.

The remuneration of the consideration value that the counter of learning device 292 calculates is defined by formula (1).

Herein, R is consideration value, O _NOXBe NO _XValue, O _COBe CO value, S _NOXAnd S _COBe NO _XAnd the goal-setting value of CO, k ₁, k ₂, k ₃, k ₄Be positive constant.

Formula 1

R＝R ₁+R ₂+R ₃+R ₄

${\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HSA00000126737300091.png"\; state="\{\{state\}\}">R</figure-callout>}}_1=\left\{\begin{array}{cc}{k}_1& (O_{\mathrm{NOx}}\&le;S_{\mathrm{NOx}})\\ 0& (O_{\mathrm{NOx}}>S_{\mathrm{NOx}})\end{array}\right.$

${\mathrm{<figure-callout\; id="2"\; label="R"\; filenames="HSA00000126737300091.png"\; state="\{\{state\}\}">R</figure-callout>}}_2=\left\{\begin{array}{cc}{k}_2& (O_{\mathrm{CO}}\&le;S_{\mathrm{CO}})\\ 0& (O_{\mathrm{CO}}>S_{\mathrm{CO}})\end{array}\right.$

${\mathrm{<figure-callout\; id="3"\; label="R"\; filenames="HSA00000126737300091.png"\; state="\{\{state\}\}">R</figure-callout>}}_3=\left\{\begin{array}{cc}{k}_3(S_{\mathrm{NOx}}-O_{\mathrm{NOx}})& (O_{\mathrm{NOx}}\&le;S_{\mathrm{NOx}})\\ 0& (O_{\mathrm{NOx}}>S_{\mathrm{NOx}})\end{array}\right.$

$R_4=\left\{\begin{array}{cc}{k}_4(S_{\mathrm{CO}}-O_{\mathrm{CO}})& (O_{\mathrm{CO}}\&le;S_{\mathrm{CO}})\\ 0& (O_{\mathrm{CO}}>S_{\mathrm{CO}})\end{array}\right.....\left(1\right)$

Shown in (1), at NO _X, the CO value is lower than respectively the S of goal-setting value _NOXAnd S _COThe time, remuneration R is provided ₁And R ₂And, at NO _X, the CO value more is lower than respectively the S of described goal-setting value _NOXAnd S _COThe time, provide and the proportional remuneration R of its deviation ₁And R ₂

And the define method of remuneration also can be considered other several different methods, is not limited to formula (1) method.

In the counter of learning device 292, reach maximum in order to make the remuneration of being calculated by formula (1), the combination of study input data 293 is operational ton, thus on the result corresponding to present situation, can learn to reduce NO _X, CO the combination of operational ton.

Learning device 292 reads in the service data 205 of the equipment 100 of present moment under the state that study finishes, according to learning outcome, calculate and export the remuneration that makes formula (1) by counter and reach maximum operational ton 295.

The numeric value analysis that described consideration value is undertaken by described numeric value analysis device 400, utilize by the equipment operation characteristic of the modeling of equipment namely with waste gas in CO, NO _X, CO ₂, SO _X, at least a corresponding measurement data is calculated in the particulate subclass that consists of of Hg (amount of mercury), fluorine, coal smoke or mist, VOC (volatile organic compounds); The object of described operational order value is the fuel flow rate of burner 102 supplies that arrange to the stove at boiler 101, to the air mass flow of burner 102 supplies, to the air mass flow of air scoop 103 supplies, angle being set, being supplied at least one in the supply air themperature of boiler 101 to the amount of exhaust gas recirculation of boiler 101 recycle, burner 101.

But present embodiment does not exist with ... the supplying method of this remuneration, can use additive method yet.

According to more than because before equipment 100 running, even if namely under the state of the running experience data that do not have equipment 100, also can be by intensified learning be inhibited CO, NO _XThe running operation method, therefore applicable when the control device 200 of present embodiment can be from the test working of equipment 100, and can bring into play effect.

For example, in situation about only learning by the running experience data of equipment 100, because accumulation data needs a few week～some months, so might can not get sufficient control performance during this period.CO, the NO of therefore possible boiler plant _XConcentration uprises and the environment deterioration, or the consumption of the medicine (ammonia) that uses at denitrification apparatus increases.

In addition, owing to carry out repeatedly experimental run until continue to have enough running experience data, so from the aspect of safe handling, also might become the state of not expecting.

With respect to this, according to the running operation method of present embodiment, owing to just do not have at last under the state of the running experience data before the running of equipment 100, also can be by intensified learning be inhibited CO concentration, NO _XTherefore the running operation method of concentration can avoid described these risks by establishment.

But, even if in the present embodiment, be zero owing to also be difficult to make the error (with the deviation of the measured value of real machine equipment) of numeric value analysis result, therefore after equipment 100 runnings, utilize the characteristic of the running experience data correction neural network model of equipment 100.

For this reason, the modeling device 291 that consists of intensified learning device 290 has following function: add aforesaid operations amount and CO concentration, NO from the service data 245 that is stored in running experience database 240 in the numeric value analysis data _XThe relation data of concentration appends study.

By the study of appending at modeling device 291, the characteristic of service data 245 is reflected in the characteristic model of neural network.

In addition, take by these service data 245 revised models as object, the learning device 292 of intensified learning device 290 has the function that relearns once again, thus, can learn more accurately the good method of operating of control performance.

Certainly, owing to can only from obtaining service data for the operating conditions after the running, therefore might also exist other the preferred state of operating conditions.Namely might sink into locally optimal solution.

For fear of this point, although obtain data as long as utilize the combination of considering for all operational tons to enlist the services of running, in physical device, for safety and keep stable running, this point can not be accomplished.

So in the present embodiment, in order to alleviate the risk that is absorbed in locally optimal solution, possess in the correcting device 250 that control device 200 has: operating condition decision maker 265, model error evaluating apparatus 260, service data modeling device 263 and an exploration point determination device 270.

When equipment 100 is in rated condition, if variation in stable condition then the operational order value diminishes, or be almost certain value, repeat equal state.

If such state is locally optimal solution, then generally speaking owing to can not change into other states, therefore become and to find better operating conditions, but in the present embodiment, by the exploration point determination device 270 that arranges in the correcting device 250 that possesses at control device 200, have the operating conditions that decision is deliberately changed, explore the function of best operating conditions.

Because the change of operating conditions probably causes the deterioration of state, so determine that the operating conditions of exploration usefulness is very difficult.

So in the present embodiment, the neural network model that the described modeling device 291 that consideration arranges by the intensified learning device 290 that possesses at control device 200 generates and error and the device measuring value of device measuring value, be the running experience data tendency the two, determine operating conditions.

Below utilize the process flow diagram of Fig. 5, order and the function of the operating conditions that determines exploration usefulness is described.

In Fig. 5, at first in the step 500 of condition discrimination, by the operating condition decision maker 265 that arranges at the correcting device 250 that consists of control device 200 shown in Figure 1, determining apparatus 100 is in rated condition.

That is, judge at operating condition decision maker 265: the load instruction 52 of inputting from the basic controlling Command Arithmetic Unit 230 that arranges at control device 200 has unchanged.When predefined reference value was following, it was unchanged to be judged as load at the absolute value of load instruction 52 and the deviation of upper sub-value.

Be judged as load when unchanged at operating condition decision maker 265, for the operational ton 295 from 290 outputs of intensified learning device, utilize operating condition decision maker 265 further to judge: whether the absolute value of the deviation of itself and upper sub-value is below predefined reference value, and whether the maximal value of the operational ton 295 that begins before the official hour and minimum value be below predefined reference value.

In the step 501 that takes a decision as to whether rated condition, entirely be in reference value when following in above-mentioned various values, determining apparatus 100 is in rated condition, enters to judge that the data number is whether greater than the step 502 of lower limit.

In the step 501 that determines whether rated condition, judge the condition judgement result 266 who exports by operating condition decision maker 265, respectively input model error assessment device 260 and service data modeling device 263.

And, be judged as by operating condition decision maker 265 when not being in rated condition, do not carry out the exploration of operating conditions and stop.

Then judging that the data number is whether in the step 502 greater than lower limit, by operating condition decision maker 265, from the present fuel supplying flow of each burner 102 of to the stove of boiler shown in Figure 2 101, arranging and arrange to the air mass flow of each burner 102, in the stove of boiler 101 each after the value of air mass flow of air scoop 103, preset the scope of the departure of regulation, retrieval enters the running experience data of the scope of this departure, and its number and reference value (data amount check lower limit) are compared.

By the above-mentioned retrieval of operating condition decision maker 265, judging that the data amount check that is judged as retrieval during the data number is whether greater than the step 502 of lower limit during greater than reference value, namely when being judged as "Yes", enters the step 505 that model error calculates; Be judged as the data number below reference value, when namely being judged as "No", entering the step 503 that shows warning.

The step 505 that model error calculates is functions of the model error evaluating apparatus 260 that possesses of correcting device 250.This model error evaluating apparatus 260, for at the full point of judging the running experience data whether the data number extracts greater than the step 502 of lower limit, calculate the error of the neural network model that generates with modeling device 291, enter the step 506 that next correlation model generates.

In the model error evaluating apparatus 260 of the step 505 that model error calculates, the counter that possesses the calculating formula of the formula (2) that consists of error of calculation value, by described counter, in each running experience data point, the error of the neural network model that will generate with the modeling device 291 of setting on the intensified learning device 290, as error value E 1, calculate according to formula (2).

(formula 2)

E1=(running experience data)-(model calculated value) ... (2)

In addition, model error evaluating apparatus 260 possesses the counter of the calculating formula of the formula (3) that consists of the average calculation error, by described counter basis (3) formula the average calculation error E2.K is the running experience number of data points herein.

(formula 3)

$\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HSA00000126737300091.png"\; state="\{\{state\}\}">E</figure-callout>}2=\frac{\mathrm{\&Sigma;}\left|\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HSA00000126737300091.png"\; state="\{\{state\}\}">E</figure-callout>}1\right|}{k}...\left(3\right)$

Judging the data number whether greater than the step 506 of lower limit, is the function of the service data modeling device 263 that arranges on the correcting device 250.By this service data modeling device 263, utilize at the full point of judging the running experience data whether the data number extracts greater than the step 502 of lower limit generating run amount and from CO and the NO of boiler plant combustion gas _XThe correlation model of concentration.

Correlation model can be multiple regression formula, neural network, other statistical model, is if input operation amount parameter then export CO, the NO of the waste gas of discharging from boiler plant _XThe model of concentration.

Determining the step 507 of operating parameter and the step 508 of operating parameter alteration command, is the function of exploring some determination device 270.

Determining the step 507 of operating parameter, by the exploration point determination device 270 that correcting device 250 possesses, at first read the error amount 261 of model error evaluating apparatus 260 calculating of the step 505 by determining operating parameter.

Fig. 6 is the figure of the relation between the characteristic of the explanation device characteristics of boiler plant and its model.The longitudinal axis of Fig. 6 (a)～Fig. 6 (b) is got from CO, the NO of the waste gas of boiler plant discharge _XConcentration, transverse axis extract operation parameter A, B, C, D, E represent measurement data with point, dot the device characteristics of real machine equipment, with the characteristic of solid line representation model.

In addition, the longitudinal axis of Fig. 6 (c) is got CO, the NO of the waste gas of boiler plant _XConcentration, transverse axis extract operation parameter A, B, C, D, E represent measurement data with point, the relation between expression correlation model and the running experience data.

And, the longitudinal axis delivery type error of Fig. 6 (d), transverse axis extract operation parameter A, B, C, D represent measurement data to represent correlation curve with solid line with point.

In Fig. 6 (a), when the model characteristics of the device characteristics of dotted line and solid line is in concerning shown in Fig. 6 (a), CO, the NO of the waste gas of the boiler plant of trying to achieve from model characteristics _XIt is B that concentration reaches minimum operating point; But in the device characteristics of the equipment 100 of reality, the CO of waste gas, NO _XIt is that the C operating point is best that concentration reaches minimum point.

If at operating point B running device 100, then can obtain representing with point among Fig. 6 (a) the running experience data of the equipment 100 of measurement data.Although deviation is arranged in the measurement data, the error value E 1 between device characteristics and the model characteristics is the correlation curve shown in Fig. 6 (d).

Shown in Fig. 6 (d), near the error amount of the model error the operating point B is for negative, and namely the running experience data of indication equipment 100 are less than CO, the NO by the waste gas of the boiler plant of model prediction _XConcentration.Might exist than the predicted value of model this moment further reduces CO, NO _XThe operating conditions of value.

When near the model error value of (present status) is beyond predefined error range near the operating point B shown in Fig. 6 (d), by the correlation model that is generated by the service data modeling device 263 that arranges at correcting device 250, CO, the NO of the waste gas of predicted operation quantitative change more the time _XConcentration.

The relation of Fig. 6 (c) expression correlation model and running experience data.According to the running experience data shown in Fig. 6 (c), CO, the NO of the waste gas of the boiler plant in the time of can predicting from correlation model change operating conditions _XConcentration.

Also have the change amplitude of operational ton, the maximum variable quantity when setting exploration from the input picture shown in Fig. 7 (a).The input field 322 of Fig. 7 (a) is inputted the value of maximum variable quantity by each operational ton.

Then, behind the maximum variable quantity of input field 322 inputs, click setting conclusion button 321 with mouse, then preserve input value.If press default setting button 320, automatically input maximum variable quantity with respect to the standard value of predefined each operational ton.

In the step 507 of the decision operating parameter of Fig. 5, shown in Fig. 7 (b), become license higher limit and the license lower limit of the condition of exploring to exploring 270 inputs of some determination device as described later.

Then explore a some determination device 270 by this, the combination of the value of operational ton is changed, calculate CO, the NO of the waste gas of the boiler plant that calculates with correlation model _XThe large operational ton of reduction effect, enter the step 508 of operating parameter alteration command output.

Then, in the step 508 of operating parameter alteration command output, will be by the CO, the NO that explore the waste gas that some determination device 270 calculates _XThe large described operational ton of reduction effect, as the output of change candidate 271.

And, CO, NO _XReduce the evaluation method of effect, identical with the definition of the consideration value of described intensified learning.

At CO, the NO than the waste gas of present situation _XWhen the operating conditions that concentration, consideration value become preferred value did not exist, namely the state shown in Fig. 6 (b) was such, prediction near present situation operating point B, the CO of waste gas, NO _XWhen concentration value increases, be in take this predicted value that this condition is the basis in the scope of License Value of setting, by the candidate of exploring some determination device 270 and select at random the alter operation amount.

Thus, might be able to find CO, the NO of waste gas of the boiler plant of the device characteristics shown in Fig. 6 (b) _XConcentration reaches minimum operating point D.

The tolerance band of the predicted value of the CO of waste gas, NO concentration is set by the input picture that the permissive condition of Fig. 7 (b) expression is set.In the picture example of Fig. 7 (b), the condition of the exploration of the optimal operating condition of the exploration point determination device 270 that arranges on the import admission correcting device 250.

In the input picture that the permissive condition of Fig. 7 (b) expression is set, can be chosen in the input field 306 license higher limit of prescribed concentration respectively of the NO concentration of the input field 305 of CO concentration of waste gas and waste gas; Perhaps specify respectively from the increasing degree of present situation concentration at the input field 307 of CO concentration and the input field 308 of NO concentration.Select one of them, to input field 305～308 inputs allowed values separately.

And, if select " solicit operation person's the license " of the input picture that permissive condition sets, then show the alter operation amount that some determination device 207 determines of exploring to the operator, what the operator can judge alter operation could.In addition, also can by operator's judgement, revise the alter operation amount and operate this moment.

Then after the described allowed values of input, set conclusion button 303 if press, then preserve input message.

And, if select the input picture of permissive condition setting " always forbidding exploring ", the model error evaluating apparatus 260, service data modeling device 263, operating condition decision maker 265, the exploration point determination device 270 that then are arranged on the correcting device 250 are not carried out these processing.

When near the error amount of the model error the operating point B shown in Fig. 6 (d) is in the predefined error range, by the model that the modeling device 291 that is arranged on the study intensifying device 290 generates, CO, the NO of the waste gas during prediction alter operation amount _XConcentration.At this moment, at CO, the NO of the waste gas that is calculated by model _XAdd on the value that error value E 2 parts are used as predicted value.

Service data during the operational ton change is reflected in the again study of correction of the model and the method for operating of utilizing the rear model of correction.

In the present embodiment, whether the error amount according near the model error the operating point B shown in Fig. 6 (d) is in the permissible error scope CO concentration or the NO of the waste gas when switching for prediction alter operation amount _XThe model of concentration, but also can make described exploration point determination device 270 have magnitude relationship corresponding to calculated value and the measured value of model, the function of the weighting of the calculated value of the model of using in the calculating of decision operational order candidate value or operational ton candidate value and the calculated value of service data model, it is average corresponding to the size of error amount the result of calculation of two models to be carried out loading.

In addition, the calculated value of the change amplitude of operational ton and model and the error amount of measured value determine pro rata.

As described above, under the in stable condition rated condition of equipment 100, by changing consciously operational ton, can from the locally optimal solution of model, free thus and can learn preferred operating conditions.

At this moment, the change amplitude of operational ton is because be waste gas CO, the NO that utilizes at the boiler plant of trying to achieve near the running experience data the present situation operating point _XThe predicted value of concentration or the predicted value that obtains after the predicted value of model adds model error decide, therefore, the danger of the outer state deteriorating of the prediction that the change of operational ton causes can suppress because can be implemented in the interior safety of running tolerance band and stable running that the operator sets.

Just in case as the result of operating conditions change, be CO, the NO that exceeds the waste gas of the boiler plant that the input picture set at the permissive condition of Fig. 7 (b) sets _XDuring the License Value scope of concentration, supspend the change (exploration) of operating conditions.At that time operational ton of storage shows whether again begin the information explored to the operator, obeys operator's input (judgement).If being input as of operator " termination " then ended a series of processing.

If being input as of operator " continuation ", although then continue the processing of operational ton change, the scope of the same operation value ± α of storage is as the selection of operational ton change candidate.α is the setting value that defines predefined environs by each operational ton.

Fig. 8 represents CO, NO from the waste gas that the boiler plant of present embodiment object is discharged _XThe mensuration example of concentration.As shown in Figure 8, the CO of waste gas and NO _XLocating of concentration is the gas flow path of boiler export position.

The gas concentration measurement face of boiler export as the right side part Watch with magnifier of Fig. 8 shows, is divided into 16 parts all around with gas flow path, is respectively arranged with CO and the NO that measures waste gas in this each cut zone of 16 parts _XThe concentration sensor of concentration.

So be disposed at the gas concentration measurement face by the concentration sensor with boiler export, CO and the NO of waste gas measured in the whole zone that can spread all over the gas flow path cross section of boiler plant _XThe CONCENTRATION DISTRIBUTION of concentration.

Fig. 9 and Fig. 8 are same, are illustrated in the gas flow path of boiler export position, measure CO and NO from the waste gas that the boiler plant of present embodiment object is discharged _XAn example of concentration.Fig. 9 (a) and Fig. 9 (b) divide into W, X, Y, Z with the gas flow path of boiler export position in front and back, left and right region is divided into 1,2,3,4 to carry out 16 and cuts apart, and is illustrated in the result that this 16 each cut zone after cutting apart is measured the CO concentration of waste gas.

Shown in Fig. 9 (a) and Fig. 9 (b), the gas flow path of boiler export position is divided into 16 parts, the CO concentration of the waste gas that will measure in each cut zone is corresponding to concentration, with color differentiating or with the deep or light concentration that represents of monochrome.

And in Fig. 9 (a) and Fig. 9 (b), because the reason of paper, only represented the measured value of CO concentration, certainly much less equally also can represent NO _XThe measured value of concentration.

The chart of the CO concentration of the waste gas that the right side of Fig. 9 (a) and Fig. 9 (b) represents, that gas flow path is divided into 16 zones as mentioned above all around, will come together at the mean value of the CO concentration of the waste gas that each cut zone is measured the front and back of gas flow path and gas flow path about and the structure example of dividing.

Fig. 9 (a) is illustrated in the boiler export position, be split in 16 parts the gas flow path about the concentration difference of waste gas CO concentration in zone little, but in the large situation of concentration difference of the waste gas CO in zone, front and back concentration; In addition, Fig. 9 (b) is illustrated in the boiler export position, and the concentration difference in the waste gas CO in zone, front and back concentration of gas flow path that is split into 16 parts is little, but about the large situation of concentration difference of waste gas CO concentration in zone.

In boiler plant, even apply operational order so that consisting of boiler 101 stove front and back or about become impartial to fuel flow rate or the air mass flow of stove supply situation under, because the difference of equipment, shown in Fig. 9 (a) and Fig. 9 (b), the CO concentration that gas flow path is divided into the waste gas in each cut zone of 16 parts in the boiler export position also might produce difference.

This is because to the action error of the actuators such as the foozle of the rotation of the difference in flow that configuration causes of the pipe arrangement of the stove of boiler 101 supply air or fuel or burner, device, valve, vibroshock etc., in most cases is difficult to it is carried out ex ante forecasting.

In the control device of the boiler plant of present embodiment, because in advance according to numeric value analysis result learning manipulation method, so the concentration difference of these reasons can not be predicted according to numeric value analysis in advance.Therefore, by the model error evaluating apparatus 260 that arranges at the correcting device 250 that consists of control device 200, the front and back shown in calculating chart 9 (a) and Fig. 9 (b) or about the zone CO concentration mean value measurement data, with the same operation condition under the error of numeric value analysis result.

The picture example that as gas concentration distribution show of both results shown in Fig. 9 (c) is such, and contrast represents measurement data and numeric value analysis result.

At zone 311 and the zone 313 of Fig. 9 (c), respectively mean value and the NO of the CO concentration shown in presentation graphs 9 (a) and Fig. 9 (b) _XThe measurement data of concentration is at regional 312 and regional 314 numeric value analysis results that represent respectively under the identical therewith operating conditions.

The measurement data that shows among Fig. 9 (c) is specified the picture that the gas concentration distribution trend of the Figure 10 (a) that can show from the real machine data selection button 315 of pressing Fig. 9 (c) bottom shows.

Figure 10 (a) is the picture that the gas concentration distribution trend of the waste gas of the boiler plant of expression in the present embodiment shows, represents respectively to have the trend map 330 of the CO concentration that changes along longitude and latitude constantly on the top of the picture of Figure 10 (a), and NO _XThe trend map 331 of concentration.

Trend map 330 and NO in these CO concentration _XIn the trend map 331 of concentration, can by with the mouse move left and right constantly specified line 332 specify moment of selection, the date of selection and the time be engraved in zone 310 and show.

In the picture bottom of this Figure 10 (a), show CO concentration and NO by the moment of mouse appointment _XThe CONCENTRATION DISTRIBUTION of concentration.

In addition, in the picture that the expression gas concentration of Fig. 9 (c) distributes, if press the numeric value analysis executive button 316 of bottom, the operating conditions in the moment that then will select in Figure 10 (a) sends to the numeric value analysis calculation element 400 that the control device 200 of Fig. 1 possesses, and sends the calculating fill order to this numeric value analysis calculation element 400.

If the calculating at the numeric value analysis of numeric value analysis calculation element 400 finishes, then read in its numeric value analysis result, in the zone 312 of Fig. 9 (c) and zone 314, show respectively CO and NO _XThe calculated value of concentration.

In addition, if press the basifacial model correction of the picture button 317 of Fig. 9 (c), then according to numeric value analysis result and the determination data of numeric value analysis calculation element 400, for the boiler export position by in 16 gas flow paths of cutting apart about or the zone, front and back with waste gas in CO and NO _XDeviation between concentration dependent numeric value analysis result and the determination data is calculated, before and after arranging in the operational ton condition of numeric value analysis with this departure is proportional or about fuel flow rate poor and calculate again.

The CO of described waste gas and NO _XThe deviation of about concentration or front and back is to result from the poor and deviation that produces of fuel flow rate, and wherein fuel flow rate is poor is fuel flow rate from a plurality of burners 102 of arranging before and after the stove of boiler 101 to stove that supply in from because certain is former thereby produce.

In the present embodiment, although during fuel flow rate is poor deviation is set, but as the operating conditions of change, the air mass flow of burner 102 supplies that also can be by in the stove that is arranged on boiler 101 or adjust to the spin intensity of the air of burner 102 supplies etc.

As previously mentioned, since front and back or about the concentration deviation can be reproduced by numeric value analysis, therefore can utilize this result, can again utilize the modeling device 291 that arranges on the intensified learning device 290 of the correcting device 250 that consists of control device 200, again construct characteristic model, can use again learning manipulation method of the model again constructed by this modeling device 291.

By so, can more meet the study of practical characteristic, improve the control performance of the boiler plant of control device 200.

The situation of the burning control of the boiler plant of Figure 10 (b) expression present embodiment.Picture example shown in Figure 10 (b), Figure 10 (a) represents the operational ton with the moment of mouse selection.

Antetheca and the burner 102 of rear wall and fuel flow rate, the air mass flow of air scoop 103 in that the stove of boiler 101 arranges represent by histogram and digital numerical value about picture.The operator can show data with reference to these, sets the change candidate 271 of the operational ton of exploring usefulness.

Explore the setting of change candidate 271 of the operational ton of usefulness, by the top that shows in digital value respectively input value and push quantitative change more set button 332, the numerical value of input is set to change candidate 271 thus.

In the step 503 that the warning of Fig. 5 shows, show that to the operator running experience data are less than the warning of setting value, allowing the operator input could continue.

Input determination step 504 decision persons.In the situation that is " automatically continuing ", enter the step 505 that model error calculates; In the situation that is " manually continuing ", enter the step 509 of input operation parameter modification value; In the situation that is " termination ", end processing and the end of its back.

In the step 509 of the input operation parameter modification value of Fig. 5, show the picture of the burning control shown in Figure 10 (b), the change value of operator's input operation amount, the step 508 that enters the operating parameter alteration command.

The operational order determination device 280 that arranges at the correcting device 250 of the control device 200 of pie graph 1, receive three kinds of operational ton signals, that is: basic operation command value 235, from the operational ton 295 of intensified learning device 290 output, from exploring the operation change candidate 271 of some determination device 270 outputs.

In operational order determination device 280, usually select the operational ton 295 as the intensified learning result, determine that scan instruction is as 285 outputs of operational order value, but produce in the unusual situation in the intensified learning function, or be not in desired result's the situation be judged as the control result because of certain reason operator, select basic operation command value 235 to export as operational order value 285.

In addition, rated conditions in the condition judgement result 266 based on operating condition decision maker 265, in the situation that allows automatically to explore, export as operational order value 285 from the operation change candidate 271 of exploring 270 outputs of some determination device by these operational order determination device 280 selections.

According to embodiments of the invention, even between the movement of model and actual control object, exist in the situation of error, also can do one's utmost to get rid of the risk that is limited to locally optimal solution, when can be implemented in safely and steadly running control control object, can explore the control device of the boiler plant of optimum solution.

Below, with reference to the description of drawings embodiments of the present invention.

Figure 11 is the figure of structure of the online gas concentration apparatus for predicting 701 of the relevant embodiments of the invention of expression.Online gas concentration apparatus for predicting 701 is computing machines.In process value database (process value DB) 711, take in the trend data of the device measuring value that the inference process of neural network uses.Operating condition configuration part 713 is in the inference process of neural network, in the input data of gas concentration deduction section 712 settings corresponding to operating condition.In gas concentration deduction section 712, take in to be useful on and infer CO and NO _XThe neural network of concentration.The output of gas concentration deduction section 712 is read into the control system of coal-burning boiler, is used for the optimization of burning control.In addition, the part of 2 outputs of gas concentration deduction section 71 also is read into display control unit 714, and is shown in display device 702.

Figure 12 represents the structure of gas concentration deduction section 712.Gas concentration deduction section 712 is made of a plurality of neural networks.As previously mentioned, the neural network of having prepared to become privileged by each coal in the present invention.

In Figure 12, coal A infers that with gas concentration neural network 722, coal B infer neural network 722 with gas concentration, is the neural network corresponding to each coal, corresponding to input data output CO and NO _XThe inferred value of concentration.

The coal ratio is judged with neural network 721, has the function according to the blending ratio of input inferred from input data coal, the coal ratio is judged the output with neural network 721, and is corresponding with described each neural network by each coal preparation, exports the blending ratio of each coal.This output valve is set as the weighting coefficient of totalizer 741.That is, gas concentration infers that the neural network of usefulness is take each coal as inferred conditions CO, NO _XConcentration is carried out the weighted mean corresponding with blending ratio and is processed in totalizer 741.

More than be the general description about the gas concentration apparatus for predicting integral body of present embodiment, below, each neural network of using in this device is elaborated.

At first judge with neural network 721 for the coal ratio, utilize Figure 13 to describe.

Coal ratio judgement neural network, the measurement data of the equipment that utilization stores in process value DB also comprises the coal class that admixture is judged the coal micro mist of the input that acts as a fuel now.Judge the output unit 731 of using neural network in the coal ratio, with the blending ratio of numerical value output for each coal A, coal B, coal C.In Figure 13, although the value of each output unit output 0.80,0.15,0.00 is not directly these values to be used as blending ratio, standardizing to process to make adds up to 1.0.In this example, process to obtain 0.84,0.16,0.00 by normalization, judges that present input raw material is with the state mixing of 84% coal A, 16% coal B.

The coal ratio is judged with in the input of neural network 721, sets performance because of the process value of the different impacts that cause of coal.CO concentration, NO are for example arranged _XConcentration, grinding machine motor electric power, boiler absorbing heat (steam condition according to boiler inlet, outlet calculates), primary air flow (being primary air fan power etc. in the situation of non-measurement), secondary air flow, two sections combustion air flow ratios, load, burning flow etc.

The following describes be used to the coal ratio that realizes above-mentioned processing and judge the learning method of using neural network.

The coal ratio judges to process with the study of neural network and uses measured data, but except the process value as input value, also needs the value as the blending ratio of the coal of output valve.But, as previously mentioned, because less than the measured value about coal, so can not use measured data about output.Therefore, use be in coal storage device storage coal and through the fully long time, the fuel that prediction drops into now only be same coal during in measured data, carry out the study of neural network.For example, dropping into fuel in prediction only is in the situation of coal A, and in neural network learning, the output unit that will be equivalent to coal A is made as 1.0, and other output unit is made as 0.0, the correlativity of study and the process value of this moment.Equally, also learn coal B, coal C.

Utilize such method to carry out the coal ratio judgement neural network of study, the process value when being same coal for input fuel, only by the unit output 1.0 corresponding to this coal, other unit output 0.0.In addition, even have in the situation that a plurality of coals mix, the function of the interpolation that has by neural network, corresponding to the ratio that contains each coal, it is large that the output valve of unit becomes.

So, take process value as the basis, can the on-line evaluation information relevant with the coal that drops into fuel.

Below, gas concentration deduction neural network is described.

Neural network is used in the gas concentration deduction, will be equivalent to the process value of burning condition as input, output CO, NO at this moment _XConcentration.Input is learnt by each coal with the relation of output.

In study is processed, only use the measured value of equipment.But, same with the study of neural network 721 with described coal ratio judgement, utilize predict fuel only to learn for the data during the same coal.By constructing like this neural network of each coal.As the input of neural network, setting and CO, NO _XThe related process value of concentration, for example, set primary air flow (being primary air fan power etc. in the situation of non-measurement), secondary air flow, two sections combustion air flow ratios, load, burning flow, operation of combustors condition, atmospheric conditions (temperature, pressure, humidity) etc.Infer that in the gas concentration of preparing by each coal in the neural network, the data of using as input all share.

In inference process, be not only the device measuring value, also use the signal from control system of coal-fired boiler.This is because control online optimization for control system, reads in to change CO, the NO that controls when requiring _XThe inferred value of concentration.For example with respect to present operating condition, CO, NO when inferring burning conditions such as changing air mass flow _XConcentration according to the inferred value of these concentration and plant efficiency, is explored optimum reference mark.

Control in the optimized processing at the fuel that control system is carried out, operating condition configuration part 713 shown in Figure 11 requires to make the input value of neural network to change corresponding to control.That is, change process value in the process value that the input of neural network is set, corresponding with controlled condition corresponding to the order from control system.For example, CO, the NO to change primary air flow the time _XWhen the variation of concentration tendency was resolved, operating condition configuration part 713 was upgraded the value of the primary air flow in the input of neural network.

In above processing, the data of gas concentration deduction section 712 outputs are read into display control unit 714, and in display device 702 outputs.Figure 14 is the display case of display device 702.In Figure 14, CO, NO when having represented that as trend coal is switched _XThe ratio of the coal of concentration and the input that acts as a fuel.For CO and NO _XConcentration is for the discharge ratio of each coal also trending.

According to processing described above, realize the CO, the NO that use based on the fuel control optimization of control system _XThe inference process of concentration.

Industrial applicibility

The present invention goes for the control device of boiler plant.

Claims (5)

1. the gas concentration estimating method of a coal-burning boiler, it utilizes neural network to infer the concentration of the gas componant of discharging from coal-burning boiler, it is characterized in that,

A plurality of neural networks that utilization is prepared in each coal and coal ratio judge that the neural network of usefulness infers,

Utilize described coal ratio to judge that the neural network of usefulness infers the blending ratio of the coal of the coal that drops into of acting as a fuel, utilize this blending ratio, inferred value to the gas concentration of described a plurality of neural networks outputs of preparing by each coal is weighted on average, infers the concentration of the gas componant when coal changes.

2. the gas concentration estimating method of coal-burning boiler as claimed in claim 1 is characterized in that,

Judge in the study of neural network of usefulness in described coal ratio, the coal that using acts as a fuel drops into be identical coal during data, learn the tendency of the correlativity of this coal and process value.

3. the gas concentration apparatus for predicting of a coal-burning boiler utilizes neural network to infer from the concentration of the gas componant of coal-burning boiler discharge, it is characterized in that,

It has totalizer, described totalizer utilization is judged the blending ratio that the neural network of usefulness and the neural network that described coal ratio is judged usefulness are inferred by a plurality of neural networks of preparing in each coal, coal ratio, inferred value to the gas concentration of described a plurality of neural networks outputs of preparing by each coal is weighted on average

Utilize described coal ratio to judge and use neural network, the blending ratio of the coal of the coal that inferring acts as a fuel drops into, utilize this blending ratio, be weighted on average by the inferred value of described totalizer to the gas concentration of described a plurality of neural networks outputs of preparing by each coal, infer the concentration of the gas componant when coal changes.

4. the gas concentration apparatus for predicting of coal-burning boiler as claimed in claim 3 is characterized in that,

Judge in the study of neural network of usefulness that in described coal ratio the coal that using acts as a fuel drops into is the data during the identical coal, learn the tendency of the correlativity of this coal and process value.

5. the gas concentration apparatus for predicting of coal-burning boiler as claimed in claim 3 is characterized in that,

Possesses display device, the act as a fuel inferred value of composition of the inferred value of coal blending ratio of the coal that drops into or gas concentration of its demonstration.

Images