Summary of the invention
Because the above-mentioned defect of prior art, technical matters to be solved by this invention is to provide a kind of coal fired power plant superheater model parameter identification method based on data-driven, by actual operating data relevant with model parameter in unit, set up the neural network identification model between required identified parameters, thereby can carry out according to the real-time change situation of steam generator system the on-line optimization identification of model parameter, having data easily obtains, to advantages such as the normal influence on system operation of unit are little, and avoid adopting in model the shortcoming of preset parameter, make model can follow the tracks of preferably the ruuning situation of real system, for the control optimization of system is laid a solid foundation.
For achieving the above object, the invention provides a kind of coal-fired power station boiler superheater model parameter identification method, comprise the following steps:
A) set up the non-linear lumped parameter model of boiler superheater, determine known parameters and need the parameter of identification;
B) known parameters described in step a) is carried out to bad point successively and process and data smoothing processing, known parameters after treatment is set up to known parameters real-time data base;
C) real-time data base of known parameters described in step b) is carried out to n section and divide processing, set up the training sample database of n different load section, the training sample database of each load section, according to being divided into uniformly some training sample subsets between loading zone, and is normalized the parameter in each training sample subset;
Wherein, n >=1.
D) set up a corresponding radial basis function neural network model for the each described training sample database in step c), and make n radial basis function (Radial Basis Function, RBF) neural network model compose in parallel hybrid network;
E) actual measured value of extraction current time superheater system, carries out on-line parameter identification to radial basis function neural network model, obtains the numerical value that current time needs the parameter of identification;
F) every time interval t, radial basis function neural network model is upgraded to processing, by new hybrid network, the parameter that needs identification is carried out to on-line parameter identification.
Further, in described step b), described known parameters is carried out to bad point processing and refer to: judge bad point by polynomial expression slip approximating method, and adopt and push away difference formula before 7 second orders bad point is rejected.
Further, in described step b), it is to adopt 7 weighted filtering methods to realize that described known parameters is carried out to data smoothing processing.
Further, in described step c), described known parameters real-time data base carries out n section division processing and refers to: determine that segmentation number is n, guarantee every segments database number equate or approximately equalised situation under, known parameters real-time data base is divided into n section according to payload, obtains the training sample database of n different load section.
Further, in described step c), described normalized refers to: the normalized value that obtains the parameter in training sample subset by calculating.Being calculated as follows of described normalized value:
Wherein: be the parameter in training sample subset, X
maxthe maximum parameter in the training sample subset of X place, X
minthe minimum parameter in the training sample subset of X place, X
*it is the normalized value of X.
Further, in described step d), describedly set up a corresponding radial basis function neural network model and comprise the following steps:
The first step, determine the output quantity of radial basis function neural network model;
Second step, from the non-linear lumped parameter model of described boiler superheater, search and the related known parameters of described output quantity, choose the input quantity of radial basis function neural network model;
The 3rd step, employing radial basis function neural network training method are trained described input quantity and described output quantity, obtain radial basis function neural network model.
Wherein, the output quantity in the first step is the parameter that needs identification.
Wherein, in the 3rd step, from known parameters, remove dependent parameter and the parameter that is difficult to obtain in known parameters, can obtain the input quantity of radial basis function neural network model.
Further, in described step e), described on-line parameter identification refers to: the actual measured value of the boiler superheater to current time is carried out bad point processing successively, data smoothing is processed and normalized, and according in the radial basis function neural network model of the corresponding load section of actual measured value input of processing as the large young pathbreaker of preload, thereby obtain the numerical value that current time needs the parameter of identification.
Further, in described step f), described renewal processing, comprises the following steps:
The first step, the actual measurement data of the superheater system in the past t time is carried out to bad point processing, data smoothing successively process and normalized, the actual measurement data after treatment of obtaining in the t time;
Second step, according to the load in each moment in the past t time, obtain over the training sample database under each moment actual measurement data after treatment in the t time, and further obtain over the affiliated training sample data subset of each moment actual measurement data after treatment in the t time;
The 3rd step, by the actual measurement data after treatment that belongs to same training sample data subset by order from front to back, replace successively original sample data, the training sample database after being upgraded;
The 4th step, for upgrade after training sample database set up a corresponding new radial basis function neural network model.
In a preferred embodiment of the present invention, a kind of coal-fired power station boiler superheater model parameter identification method, comprises the following steps: set up the non-linear lumped parameter model of boiler superheater, determine known parameters and the parameter that needs identification; Known parameters is carried out to bad point and process and data smoothing processing, obtain known parameters real-time data base; Set up the training sample database of n different load section, and parameter is normalized; Set up a corresponding RBF neural network model for each tranining database, and make n RBF neural network model compose in parallel hybrid network; Extract the actual measured value of current time superheater system, RBF neural network model is carried out to on-line parameter identification; Every time interval t, the carrying out of RBF neural network model upgraded and processed.
As can be seen here, compared with prior art, the parameter identification method of superheater model that the present invention proposes, based on data-driven, has data and easily obtains, without the advantage such as unit being tested to.The present invention, by load dividing is set up by the composite nerve network of multiple RBF neural network model combinations and carried out parameter identification, has reduced the deficiency of single Neural generalization ability, has improved the identification precision of superheater model parameter under different load; In each load section, according to the multiple training sample data of step-length Further Division subset, guarantee the harmony that data sample is selected, further improved the generalization ability of model.In addition, the present invention has set up the online timing update mechanism of neural network identification model, to reflect the situation of change of superheater system its model parameter in long-time running, for model output provides good basis to the long-term on-line tracing of real system output.The present invention has avoided adopting in conventional superheater model the shortcoming of preset parameter, has realized the real-time identification of model parameter, has guaranteed the identification precision of model parameter.
Embodiment
In one embodiment of the invention, boiler superheater meets condition below:
A, outside hot-fluid distribute along pipe range and even circumferential;
B, tube wall metal are only considered radially to conduct heat, and do not consider axial heat conduction;
C, intraductal working medium are incompressible, and parameter uniformity on flow section;
D, ignore attemperator dynamic perfromance;
E, representation parameter take superheater outlet parameter as lumped parameter model, resistance concentrates on entrance.
For the boiler superheater of the present embodiment, a kind of coal-fired power station boiler superheater model parameter identification method, comprises the following steps:
The first step, sets up the non-linear lumped parameter model of boiler superheater, determines known parameters and needs the parameter of identification.
The non-linear lumped parameter model of the present embodiment, specifically:
Wherein: D
1it is superheater inlet steam flow; D
2it is superheater outlet steam flow; D
jwit is desuperheating water flow; H
1it is superheater inlet steam enthalpy; H
2it is superheater outlet steam enthalpy; H
jwit is desuperheating water enthalpy; V
git is superheater volume; ρ
2it is superheater outlet vapour density; Q
2that superheater metal pipe-wall is passed to steam working medium heat; Q
sthat flue gas is passed to superheater metal pipe-wall heat; C
mit is superheater density metal; M
git is superheater metal quality; T
git is superheater tube wall temperature; K is the coefficient of heat transfer; α is index; T
1it is superheater inlet steam temperature; T
2it is superheater outlet steam temperature; K is coefficient; β is index; P
1superheater inlet steam pressure; P
2it is superheater outlet vapor pressure.
In the present embodiment, need the parameter of identification to comprise: Coefficient K, coefficient k and index β; All the other known parameters are obtained from design data, or are recorded by on-the-spot surveying instrument.
In the present embodiment: specified steam flow D
2909.6t/h, specified desuperheating water flow D
jw15t/h, superheater volume V
g120m
3, superheater density metal C
m7900kg/m
3, index C
mbe 0.8, it is 460J/ (kg. ℃) that metal specific heat holds, nominal steam pressure C
m14.5Mpa.
Second step, carries out bad point to known parameters and processes and data smoothing processing, obtains known parameters after treatment, thereby obtains known parameters real-time data base.
Described bad point processing is to judge bad point by polynomial expression slip approximating method, and adopts and push away difference formula before 7 second orders bad point is rejected.
Before 7 described second orders, push away difference formula, specifically: when below k point meets when formula, by k some rejecting:
Wherein:
I=7,8..., y
ifor measured data,
for interpolated data,
Described data smoothing processing is to adopt 7 weighted filtering methods to realize, specifically:
Wherein:
for filtered numerical value, y
mfor the actual measured value of moment m.
For the needs of subsequent calculations, in the present embodiment by model formation:
:
Be transformed to:
Can, in the hope of superheater inlet steam flow, for the ease of identification, be taken the logarithm in superheater inlet steam flow formula both sides by above formula and service data:
Arrange:
Order
x=log (P
1), a=log (k), β=b has formula:
y=ax+b。
By above-mentioned discussion, can be in the hope of (x, y) data pair by actual operating data, recycling least square method can pick out a and b, and then tries to achieve k and β, and parameter K in like manner can obtain.
The 3rd step, known parameters real-time data base is carried out to n section and divide processing, set up the training sample database of n different load section, and the training sample database of each load section is divided into some training sample subsets uniformly according to S between loading zone, and the parameter in each training sample subset is normalized.
Described n section is divided and is processed, and refers to: determine that segmentation number is n, guaranteeing that every segments database number tries one's best equal in the situation that, known parameters real-time data base is divided into n section according to payload, thereby obtains the training sample database of n different load section.
N=4 in the present embodiment, is divided into more than 85% the training sample database of load section, the training sample database of 70%~85% load section, the training sample database of 55%~70% load section, the training sample database of load 55% below section.
Described normalized, is:
Wherein: be the parameter in training sample subset, X
maxthe maximum parameter in the training sample subset of X place, X
minthe minimum parameter in the training sample subset of X place, X
*it is the normalized value of X.
The present embodiment is further that step-length is divided into some training sample subsets uniformly by the training sample database of each load section according to 0.5% total load.
The 4th step, sets up a corresponding RBF neural network model for each tranining database, and makes n RBF neural network model compose in parallel hybrid network.
Described set up a corresponding RBF neural network model, comprise the following steps:
1) determine the output quantity of RBF neural network model, need the parameter of identification;
2) known parameters of searching from the non-linear lumped parameter model of boiler superheater and needing the relating to parameters of identification to be, removes dependent and the amount that is difficult to obtain in known parameters, thereby obtains the input quantity of RBF neural network model;
3) adopt existing RBF neural metwork training method to train input quantity and output quantity, obtain RBF neural network model.
In the present embodiment, the input quantity of RBF neural network model is: P
1, T
g, T
1, D
1and D
2, output quantity is: Coefficient K, coefficient k and index β.
The 5th step, the actual measured value of extraction current time superheater system, carries out on-line parameter identification to RBF neural network model, obtains the numerical value of the parameter that now needs identification.
Described on-line parameter identification, be: the actual measured value of current time superheater system is carried out to bad point processing, data smoothing processing and normalized successively, and in the RBF neural network model of the corresponding load section of actual measured value input of processing according to the large young pathbreaker when preload, thereby obtain the numerical value of the parameter that now needs identification.
The 6th step, every 6 hours, RBF neural network model is upgraded to processing, the RBF neural network model after being upgraded, and make the RBF neural network model after renewal compose in parallel new hybrid network, by new hybrid network, the parameter that needs identification is carried out to on-line parameter identification.
Described renewal processing, comprises the following steps:
1) actual measurement data of the superheater system in past 6 hours is carried out successively to bad point processing, data smoothing processing and normalized, the actual measurement data after treatment of obtaining in the t time;
2) according to the load in each moment in past 6 hours, obtain over the training sample database under each moment actual measurement data after treatment in 6 hours, and further obtain over the affiliated training sample data subset of each moment actual measurement data after treatment in 6 hours;
3) by the actual measurement data after treatment that belongs to same training sample data subset by order from front to back, replace successively original sample data, the training sample database after being upgraded;
4) set up a corresponding new RBF neural network model for the training sample database after upgrading.
The present embodiment, in order to pick out exactly the variation of different load operating mode drag parameter, carries out segmentation according to the load of system, and operating mode larger difference is set up respectively to multiple corresponding neural network identification models.Set up identification model by segmentation, the deficiency of generalization ability can reduce on the one hand the operation of single Neural simulation whole system Wide Range time, thus improve the identification precision of model parameter under different load; The scale and the sample size that have reduced again on the other hand each neural network, improved network training speed, for the real-time of on-line identification provides corresponding assurance.Setting up before each neural network model, need to carry out suitable selection and pre-service to the service data of actual acquisition, further to improve generalization ability and the identification precision of neural network.In addition, for each neural network identification model has been set up the mechanism of online updating training sample, and timing is carried out the training of neural network with legacy data in measured data replacement training sample database, to be reflected in better the situation of change of time statistical significance drag parameter, thereby model and real system can be matched, for control and the optimization of system provide basis.
More than describe preferred embodiment of the present invention in detail.Should be appreciated that those of ordinary skill in the art just can design according to the present invention make many modifications and variations without creative work.Therefore, all technician in the art, all should be in by the determined protection domain of claims under this invention's idea on the basis of existing technology by the available technical scheme of logical analysis, reasoning, or a limited experiment.