CN115618506A

CN115618506A - Method for predicting power of single-shaft combined cycle gas turbine

Info

Publication number: CN115618506A
Application number: CN202211123316.0A
Authority: CN
Inventors: 张宝凯; 庄义飞; 曲晓荷; 郭宝
Original assignee: China Datang Corp Science and Technology Research Institute Co Ltd; Datang Boiler Pressure Vessel Examination Center Co Ltd; East China Electric Power Test Institute of China Datang Corp Science and Technology Research Institute Co Ltd
Current assignee: China Datang Corp Science and Technology Research Institute Co Ltd; Datang Boiler Pressure Vessel Examination Center Co Ltd; East China Electric Power Test Institute of China Datang Corp Science and Technology Research Institute Co Ltd
Priority date: 2022-09-15
Filing date: 2022-09-15
Publication date: 2023-01-17

Abstract

The invention discloses a method for predicting power of a single-shaft combined cycle gas turbine, which comprises the following steps: acquiring optimal variable data sets, wherein each optimal variable data set comprises a training set, a test set and a verification set; acquiring a QM (quality metric model) based on a quantum particle swarm algorithm according to the optimal variable data set, acquiring a DM (model management) model and an EM (effective component analysis) model based on a deep neural network and a recurrent neural network algorithm, and respectively performing modeling training on a training set; obtaining an absolute error between a prediction result and an actual value of the model, selecting the model corresponding to the minimum absolute value error as a prediction optimal algorithm under a test set, and obtaining a sample test set; setting a deep neural network classification model, carrying out optimization based on a quantum particle swarm algorithm to obtain an HM model, training the HM model through a sample test set, verifying the trained HM model on a verification set, and adaptively selecting a sub-model according to different working conditions to carry out power prediction on the gas turbine. The prediction accuracy can be improved by the method and the device.

Description

Method for predicting power of single-shaft combined cycle gas turbine

Technical Field

The invention relates to the technical field of power generation of gas turbines, in particular to a method for predicting power of a single-shaft combined cycle gas turbine.

Background

The single-shaft gas-steam combined cycle has the advantages of long start-stop time of the unit, strong environmental adaptability, high load response capability, high thermodynamic cycle efficiency and the like, and further obtains large-scale development of domestic markets compared with other forms of energy power generation. The calculated power of the gas turbine directly influences the coordination and optimization of the whole machine cycle, the gas turbine and the steam turbine of the existing single-shaft combined cycle unit share one generator, the power measuring device of the generator can monitor the power value of the gas turbine into 2 parts, namely, the power measuring device is judged by the state of a clutch engaging signal (KUPE) of a shaft clutch of an SSS steam turbine, and when the engaging signal KUPE =0, the power measuring value of the gas turbine is the actual value of a dynamometer. And when the combined cycle operates normally and the meshing signal KUPE =1, the measured value of the power of the gas turbine is the calculated value of the power of the gas turbine. The calculated value is mathematically solved through a logic mechanism mathematical relation and related operation parameters, although the calculation method is accurate, the calculated value of the power of the combustion engine cannot be solved in advance, and hysteresis exists. Therefore, the reliable gas turbine performance prediction model is established, advanced accurate and perfect controllable parameters are directly provided for the combined cycle, and the economy and the safety of the unit are guaranteed.

The core of constructing an accurate prediction model is to find a suitable modeling method aiming at research problems, and in the existing research, the modeling method adopts three methods, namely a traditional physical analysis method, a statistical method and a data-driven advanced intelligent method. In the prior art, caoJun et al published the document "development of high-precision full-range simulation system of F-class gas-steam combined cycle unit", and use APROS software to analyze the mass and momentum of the thermodynamic process of a gas turbine, and perform real-time dynamic simulation of the full-range process of the F-class gas-steam unit according to the law of conservation of energy. However, the mechanistic approach has room for improvement in modeling accuracy. As a subsystem for intelligent power plant control, a plant-level data monitoring System (SIS) realizes monitoring recording and real-time access of a large number of parameter data states in unit process control, and provides guarantee for an advanced intelligent modeling control strategy. The Lijingxuan and the like issue 'gas turbine mechanism-data hybrid modeling method research', a hybrid model controller method which takes an intelligent algorithm as mechanism model error compensation is designed, verification experiments based on Distributed Control System (DCS) data are carried out on the design of different combination modes, and the prediction precision is improved. The RS-RBF-based gas turbine control system sensor fault diagnosis research is published by the Yunshao et al, the sensor fault symptom attributes are processed by an improved equal-frequency discrete method to construct a rough set, and the next step of establishing an RBF network is carried out to reduce the misjudgment rate of the sensor fault. Although the existing research is successful in modeling the parameters of the power station, the above researches are all shallow machine learning methods, and deep useful information hidden in a data bottom layer cannot be captured, so that the prediction accuracy is not high enough.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the problem that the power prediction precision of the single-shaft combined cycle gas turbine is not high is solved.

In order to solve the technical problems, the invention provides the following technical scheme:

acquiring a sample data vector set according to original data of the power monitoring variable of the gas turbine;

abnormal point rejection and smooth wave processing are carried out on the sample data vector set, important monitoring variables are selected by combining a partial mutual information algorithm, an optimal variable data set is obtained, the optimal variable data set is divided into a first optimal variable data set, a second optimal variable data set and a third optimal variable data set, and each optimal variable data set comprises a training set, a testing set and a verification set;

identifying a power multivariable system model structure of the gas turbine based on a quantum particle swarm algorithm according to the optimal variable data set, acquiring a QM (quality metric model), acquiring a DM (model division multiplexing) model and an EM (effective matrix) model based on a deep neural network and a recurrent neural network algorithm, and respectively performing modeling training on the training set;

verifying the QM model, the DM model and the EM model through the test set, obtaining absolute errors between prediction results and actual values of the QM model, the DM model and the EM model, selecting the model corresponding to the minimum absolute value error as a prediction optimal algorithm under the test set, performing label classification on the test set at the moment, adding the test set with the finished label classification into the test set again, and obtaining a sample test set; setting a deep neural network classification model structure, optimizing based on a quantum particle swarm algorithm to obtain an HM model, training the HM model through the sample test set, verifying the trained HM model on the verification set, and adaptively selecting a sub-model according to different working conditions to predict the power of the gas turbine.

The advantages are that: the method comprises the steps of processing original data acquired from a TCS data acquisition system through a mechanism mathematical analysis and PMI screening method to obtain an optimal variable data set, dividing the optimal variable data set into a first optimal variable data set, a second optimal variable data set and a third optimal variable data set, wherein each optimal variable data set comprises a training set, a testing set and a verification set. The method comprises the steps of identifying a power multivariable system model structure of the gas turbine based on a quantum particle swarm algorithm, obtaining a QM model, obtaining a DM model and an EM model based on a deep neural network and a recurrent neural network algorithm, carrying out modeling prediction on a first optimal variable data set, and constructing an adaptive selection algorithm model by adopting a QPSO-DBN algorithm, namely, carrying out gas turbine computing power adaptive prediction on the HM model, wherein the prediction process is not a single algorithm prediction result, but according to a sample working condition, the QPSO-DBN algorithm is used for carrying out adaptive selection on sub-models, namely, the QM model, the DM model and the EM model to obtain a more accurate prediction result, and the prediction precision is higher.

In an embodiment of the present invention, the obtaining the optimal variable data set includes the following steps:

s210, taking the sample data vector set as an initial data set, obtaining mutual information of the ii th element feature and the (N + 1) th element feature in the initial data set, and moving the corresponding element feature into the initial set when the maximum mutual information is selected;

s220, in the ii th element feature and the N +1 th element feature of the initial data set, eliminating each feature influenced by the initial set in the ii th element feature to obtain an input residual error, and eliminating each feature influenced by the initial set in the N +1 th element feature to obtain an output residual error;

s230, obtaining residual mutual information according to the input residual and the output residual, and finding out a residual variable when the residual mutual information is maximum;

s240, putting the residual variable into the initial set to obtain an updated initial set, obtaining a variable set according to the ii-th element characteristic and the residual variable, and returning to execute the steps S220-S240 according to the updated initial set and the variable set until the variable set is empty to obtain the akage information amount criteria corresponding to all variables in the variable set;

and S250, acquiring the optimal variable data set according to the variable sequence of all the variables moving into the initial set in the maximum mutual information and the size increase and decrease condition of the corresponding hematid pool information criterion value.

In an embodiment of the present invention, the mutual information of the ii th element feature and the N +1 th element feature in the initial data set is described by the following formula:

in the formula (I), the compound is shown in the specification,

expressed as the ii th element feature in the initial dataset

And N +1 th element characteristics

F (-) is expressed as an estimated density function of m samples, x _hl Expressed as samples of the h row and l column in the output matrix, y _l The corresponding output characteristic represented as the h row, l column;

the input residual error is obtained by the following formula:

the output residual is obtained by the following formula:

in the formula, V is an input residual error, U is an output residual error, E (-) is a condition expectation, and S is an initial set;

the set of variables is obtained by the following formula:

the akage pool information amount criterion is obtained through the following formula:

in the formula, C ₁ Expressed as a set of variables, AIC expressed as the Red pool information content criterion, u ₂ Regression residual, p, expressed as output residual of a variable selected in a set of variables ₁ Expressed as the number of variables selected in the variable set, and m as a sample.

In an embodiment of the present invention, identifying a gas turbine power multivariable system model structure based on a quantum-behaved particle swarm algorithm to obtain a QM model, includes the following steps:

s311, initializing positions of particles in a first search space, mapping the particles into a group of identification parameters for identifying a power model structure of the gas turbine to form an initial particle swarm, and initializing nine-dimensional optimal positions and nine-dimensional global optimal positions of the particles;

s312, obtaining an average optimal position of the nine-dimensional particle swarm according to the initial particle swarm and the optimal positions of the initial particles;

s313, bringing part of the data sets in the first optimal variable data set, the second optimal variable data set and the third optimal variable data set into the structure of the model for identifying the computational power of the gas turbine for identification training, obtaining a fitness value, comparing the current fitness value with the fitness of the previous iteration, updating the current particle position to a nine-dimensional optimal position if the fitness value of the current position is smaller than the fitness of the optimal position of the previous iteration, and otherwise, keeping the nine-dimensional optimal position of the previous iteration;

s314, acquiring the updated nine-dimensional global optimal position of the current particle according to the latest nine-dimensional optimal position of the current particle and the fitness of the current particle, comparing the nine-dimensional global optimal position with the nine-dimensional global optimal position of the previous iteration, if the current nine-dimensional global optimal position is smaller than the nine-dimensional global optimal position of the previous iteration, updating the nine-dimensional global optimal position of the current particle into the nine-dimensional global optimal position, and otherwise, keeping the nine-dimensional global optimal position of the previous iteration;

s315, acquiring a nine-dimensional attractor according to the latest nine-dimensional optimal position and the latest nine-dimensional global optimal position of the current particle, determining a nine-dimensional particle convergence region, and updating the nine-dimensional particle position;

s316, if the iteration times meet a first termination condition, outputting the optimal particles, and if the iteration times do not meet the first termination condition, returning to the step S312;

and S317, reversely bringing the output optimal particles back to the structure of the gas turbine calculation power model, and acquiring the QM model.

In an embodiment of the present invention, the gas turbine computational power model structure is:

wherein Y(s) is the output variable of the gas turbine power model gas turbine power, U ₁ (s),U ₂ (s),U ₃ (s) input variables, K, respectively, in the optimal variable data set ₁ 、K ₂ 、K ₃ 、T ₁ 、T ₂ 、T ₃ 、τ ¹ 、τ ² 、τ ³ Expressed as the parameter to be identified, e ^-s Expressed as a delay factor, and s is expressed as a base projection variable symbol of the real space function in the complex frequency domain space.

In one embodiment of the present invention, wherein,

the search space is [50, 500], the particle dimension is j, (j =1,2, \ 8230; 9), the number of particles is i, (i =1,2, \ 8230; 50);

the average optimal position of the nine-dimensional particle swarm is obtained through the following formula:

in the formula, C _j (t) represents the average optimal position of the nine-dimensional particle swarm, M is a constant and takes the value of the particle number, M =50, t represents the iteration number,

in the j dimension expressed as the ith particleThe optimal position of (a);

the nine-dimensional optimal particle position updating formula is as follows:

in the formula (I), the compound is shown in the specification,

the nine-dimensional optimal position of the current particle,

expressed as the nine-dimensional optimal position of the particle of the previous iteration, E _rr () expressed as the fitness of the current particle;

the nine-dimensional global optimal position is obtained by the following formula:

in the formula (I), the compound is shown in the specification,

representing that the current particle is the latest nine-dimensional global optimal position;

the nine-dimensional attractor is obtained by the following formula:

updating the nine-dimensional particle position is by the following equation:

in the formula, L _i Expressed as a nine-dimensional attractor, [ phi ] expressed as [0,1]]Random number within interval, x _i,j (t) particles denoted as the t-th iteration, x _i,j (t + 1) is expressed as particles for the t +1 th iteration, and u is expressed as being [0,1]]The number of random distributions within the interval, α, is expressed as the compression expansion factor.

Wherein the compression expansion factor is obtained by the following formula:

in an embodiment of the present invention, obtaining the HM model comprises the following steps:

s410, sending the test set to a trained QM model, a trained DM model and a trained EM model to obtain a prediction result under the test set;

s420, obtaining an absolute value of an error between the prediction result and an actual value, selecting a model corresponding to the minimum absolute value of the error as a prediction optimal algorithm under a test set, and performing label classification on the test set at the moment, wherein the label of the QM model, the label of the DM model and the label of the EM model are respectively set as a first class, a second class and a third class;

s430, adding the test set with the marked type into the original test set again to form a new sample test set;

s440, setting a deep neural network classification model structure, initializing positions of particles in a second search space, mapping the particles into a group of parameters of the number of nodes of a four-layer hidden layer to form an initial particle swarm, and initializing a four-dimensional optimal position and a four-dimensional global optimal position of the particle swarm;

s450, obtaining an average optimal position of the four-dimensional particle swarm according to the four-dimensional optimal position of the initialization particle;

s460, bringing a new sample test set into the deep neural network classification model structure for network training, obtaining the fitness of 10 times of cross validation average precision of each particle, comparing the fitness of the current average precision with the fitness of the average precision of the previous iteration, updating the current particle position to the optimal position if the fitness of the current average precision is smaller than the fitness of the average precision of the second optimal position of the previous iteration, otherwise, keeping the optimal position of the previous iteration;

s470, obtaining the four-dimensional global optimal position of the current particle according to the four-dimensional optimal position of the current particle and the fitness of the average precision of the current particle, comparing the four-dimensional global optimal position with the four-dimensional global optimal position of the previous iteration, if the current four-dimensional global optimal position is smaller than the four-dimensional global optimal position of the previous iteration, updating the four-dimensional global optimal position of the current particle to the four-dimensional global optimal position, otherwise, keeping the four-dimensional global optimal position of the previous iteration;

s480, acquiring a four-dimensional attractor according to the four-dimensional optimal position and the four-dimensional global optimal position of the current particle, determining a four-dimensional particle convergence region, and updating the position of the four-dimensional particle;

s490, if the iteration times meet a second termination condition, outputting an optimal four-dimensional particle to obtain an HM model, and if the iteration times do not meet the second termination condition, returning to the step S450; carrying back the output optimal four-dimensional particles to the model structure to obtain the HM model;

s4100, sending the verification set to the HM model trained in step S490 for classification, and then selecting the best model of the verification results from the QM model, the DM model and the EM model for prediction.

In one embodiment of the present invention, wherein,

the second search space is [10, 300], the particle dimension is j, (j =1,2, \82304), the particle number is i, (i =1,2, \823050);

the average optimal position of the four-dimensional particle swarm is obtained through the following formula:

in the formula, C _j2 (t) is expressed as the average optimal position of the four-dimensional particle swarm, and the numerical value of M is 50;

the fitness of the average precision is obtained by the following formula:

in the formula, A _rr Fitness expressed as mean accuracy, f _L (x _i,j (t)) is expressed in a new sample test set

Prediction label obtained by prediction of extreme learning machine under training sample, V _{tran_label} Expressed as in a new sample test set

Actual labels under training samples, F _k (X) is a real number, based on the predicted label being equal to the actual label, wherein the regularization coefficient of the extreme learning machine is set to 0.001;

the particle four-bit optimal position updating formula is as follows:

in the formula (I), the compound is shown in the specification,

expressed as the four-dimensional optimal position of the current particle,

expressed as the four-dimensional optimal position of the particle of the previous iteration, A _rr () fitness expressed as the average accuracy of the current particle;

the formula for obtaining the latest four-dimensional global optimal position is as follows:

in the formula (I), the compound is shown in the specification,

representing the latest four-dimensional global optimal position of the current particle;

the four-dimensional attractor is obtained through the following formula:

in the formula, L _2i Expressed as a four-dimensional attractor, [ phi ] expressed as [0,1]]Random numbers within the interval;

the updated four-dimensional particle position is obtained by the following formula:

x 'in the formula' _i,j (t) is represented as the four-dimensional particle, x 'of the t iteration' _i,j (t + 1) represents the four-dimensional particle for the t +1 th iteration.

In one embodiment of the present invention, F is the actual tag when the predicted tag is equal to the actual tag _k (X) has a value of 1, and F is the value of 1 when the predicted tag is not equal to the actual tag _k The value of (X) is 0.

In an embodiment of the present invention, the first termination condition is t =1000.

Compared with the prior art, the invention has the beneficial effects that: the performance of the calculated power of the gas turbine is predicted by a mixed form adaptive selection algorithm model, namely an HM model, the calculated power of the gas turbine can be predicted by adaptively selecting submodels according to different working conditions, the accuracy of the calculated power prediction of the single-shaft gas turbine is obviously higher than that of other algorithms, and the prediction accuracy is improved.

Drawings

FIG. 1 is a flow chart of a method for predicting power of a single shaft combined cycle gas turbine in accordance with an embodiment of the present invention.

Fig. 2 is a flowchart of obtaining an optimal variable data set according to an embodiment of the present invention.

Fig. 3 is a flowchart of acquiring a QM model according to an embodiment of the present invention.

FIG. 4 is a flow chart for obtaining the HM model, according to an embodiment of the present invention.

Fig. 5 is a schematic diagram of a PMI variable selection process according to an embodiment of the present invention.

FIG. 6 is a comparison diagram of the optimization results of the PSO algorithm and the QPSO algorithm according to the embodiment of the invention.

Fig. 7 (a) - (b) are schematic diagrams of prediction results of different prediction algorithms in the first optimal variable data set according to the embodiment of the present invention.

Fig. 7 (c) - (d) are schematic diagrams illustrating prediction results of different prediction algorithms in the second optimal variable data set according to the embodiment of the present invention.

Fig. 7 (e) to (f) are schematic diagrams of prediction results of different prediction algorithms in the third optimal variable data set according to the embodiment of the present invention.

Detailed Description

In order to facilitate the understanding of the technical solutions of the present invention for those skilled in the art, the technical solutions of the present invention will be further described with reference to the drawings attached to the specification.

The terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the present application, "a plurality" means two or more unless specifically limited otherwise.

Referring to fig. 1 and 2, the present invention provides a method for predicting power of a single-shaft combined cycle gas turbine, comprising the steps of:

s100, acquiring a sample data vector set according to original data of the power monitoring variable of the gas turbine;

s200, carrying out exception point elimination and smooth wave processing on the sample data vector set, selecting important monitoring variables by combining a partial mutual information algorithm to obtain an optimal variable data set, and dividing the optimal variable data set into a first optimal variable data set, a second optimal variable data set and a third optimal variable data set, wherein each optimal variable data set comprises a training set, a test set and a verification set;

s300, identifying a power multivariable system model structure of the gas turbine based on a quantum-behaved particle swarm algorithm according to the optimal variable data set, acquiring a QM (quality metric model), acquiring a DM (model division multiplexing) model and an EM (expectation maximization) model based on a deep neural network and a recurrent neural network algorithm, and respectively performing modeling training on the training set;

s400, verifying the QM model, the DM model and the EM model through the test set, obtaining absolute errors between prediction results and actual values of the QM model, the DM model and the EM model, selecting the model corresponding to the minimum absolute value error as a prediction optimal algorithm under the test set, performing label classification on the test set at the moment, adding the test set with the label classification completed into the test set again, and obtaining a sample test set; setting a deep neural network classification model, optimizing based on a quantum particle swarm algorithm to obtain an HM model, training the HM model through the sample test set, verifying the trained HM model on the verification set, and adaptively selecting a sub-model according to different working conditions to predict the power of the gas turbine.

Referring to FIG. 1, in one embodiment of the present invention, before predicting the power of a single-shaft combined cycle gas turbine in step S100, a parameter value related to the calculated power of the gas turbine is determined according to a mathematical relation of internal logic mechanisms of the gas turbine. The gas turbine and the steam turbine of the single-shaft combined cycle unit share one generator, and the monitoring of the power value of the gas turbine by a power measuring device can be divided into 2 parts, namely the measured value of a dynamometer and the calculated value of the power of the gas turbine. Wherein, the calculated gas turbine power value can be represented by the following formula:

P _GT ＝min(P _EL ,max(10.5984，f ₄ (T ₁ )×PGT ₇₀ ))， (1)；

in the formula, P _GT Power, P, expressed as a measurement of gas turbine power _EL Expressed as measured power value, PGT, of the dynamometer ₇₀ Expressed as calculated gas turbine power, T ₁ Expressed as compressor inlet temperature, f ₄ (T ₁ ) The calculated value of the power of the gas turbine is represented as a conversion function of the inlet temperature of the compressorThe power value corresponding to the pressure of the outlet of the compressor is obtained by the following formula:

in the formula, p ₁ Expressed as compressor inlet pressure, f ₁ (x)、f ₂ (x) And f ₃ (x) Expressed as gas turbine exhaust temperature T _otc Pressure ratio u to compressor ₁ A transformation function obtained by the following formula:

f ₃ (x)＝f ₃ (T _otc -579×1.015+a _x )×0.7， (5)；

in the formula, a _x The meaning of no actual representation is only 0 or-4, and the specific value is judged according to the axial displacement of the rotor. Wherein the gas turbine exhaust temperature T _otc Obtained by the following formula:

T _otc ＝y ₁ +y ₂ +y ₃ ， (6)；

y ₁ ＝0.5B+0.5C， (7)；

y ₂ ＝-(T ₁ ×0.01625+0.0115)×T ₁ ， (8)；

in the formula, B and C are respectively expressed as the average value of the two-branch thermocouple of the exhaust temperature of the 24-branch gas turbine, n is expressed as the rotating speed of the gas turbine, when n is more than 47.5Hz,

when y is ₃ And (5). It y is ₁ ～y ₃ Without practical implications, only for obtaining the gas turbine exhaust temperature T _otc The intermediate variable of (1). f. of ₅ (x) Expressed as gas turbine exhaust temperature T _otc Pressure ratio u to compressor ₁ And f is a conversion function of ₁ (x)～f ₅ (x) Respectively corresponding to linear transformation functions in different value ranges.

According to the TCS internal logic mechanism mathematical quantitative relation, the calculated power of the gas turbine is mainly related to the inlet temperature of the gas compressor, the inlet pressure of the gas compressor, the pressure ratio of the gas compressor, the exhaust temperature of the gas turbine, the axial displacement of a rotor and the strong rotating speed of the gas turbine, so that the 6 variables are selected as modeling candidate input variables for model modeling variables, and the calculated power of the gas turbine is selected as output variables.

The method comprises the steps of obtaining original DATA of modeling candidate input variables and output variables from a TCS DATA acquisition system, and establishing a sample DATA set DATA = X U Y, wherein X represents characteristics, and Y represents output characteristics. Wherein the content of the first and second substances,

n represents the number of features, and in the present embodiment, N =6.

The features contain m numbers, i.e. m samples. Output characteristic Y = { Y = ₁ ,y ₂ ,…,y _m Each of which corresponds to a numerical value

The power value is calculated for each sample of gas turbines. Constructing input-output matrices

And rewriting the input/output matrix D into a sample data vector set composed of N +1 column vectors

Wherein the content of the first and second substances,

representing the first elementary feature in the sample data vector set, which is actually the first column vector in the input-output matrix D, corresponding to the first feature of the sample data set.

Referring to fig. 1 and fig. 2, in an embodiment of the present invention, in step S200, obtaining an optimal variable data set includes the following steps:

s210, taking the sample data vector set as an initial data set, obtaining mutual information of the ii th element feature and the (N + 1) th element feature in the initial data set, and moving the corresponding element feature into the initial set when the maximum mutual information is selected.

Wherein, mutual information of the ii th element characteristic and the N +1 th element characteristic in the initial data set is obtained through the following formula:

in the formula (I), the compound is shown in the specification,

expressed as the ii th element feature in the initial dataset

And the (N + 1) th element feature

F (-) is expressed as an estimated density function of m samples, x _hl Expressed as samples of the h row and l column in the output matrix, y _l Represented as the corresponding output characteristic of row h and column l. The initial set is denoted by S, and at the very beginning, the initial set is an empty set.

S220, in the ii th element feature and the N +1 th element feature of the initial data set, eliminating each feature influenced by the initial set in the ii th element feature to obtain an input residual error, and eliminating each feature influenced by the initial set in the N +1 th element feature to obtain an output residual error.

Wherein the input residual is obtained by the following formula:

the output residual is obtained by the following formula:

in the formula, V is expressed as an input residual, i.e. as the remaining ii th element feature in the sample data vector set D

Each feature in (a) rejects the input residual affected by the initial set information. U is expressed as output residual error, namely the characteristic of the (N + 1) th element in the data vector set D

The output residual error after the influence of the initial set information is removed from each feature in the image. E (-) is expressed as a conditional expectation.

And S230, obtaining residual mutual information according to the input residual and the output residual, and finding out a residual variable when the residual mutual information is maximum.

Obtaining residual mutual information I (V, U) according to formula (10), wherein the residual variables are used

And (4) showing.

S240, putting the residual variable into the initial set to obtain an updated initial set, obtaining a variable set according to the ii-th element characteristic and the residual variable, and returning to execute the steps S220-S240 according to the updated initial set and the variable set until the variable set is empty to obtain the akage information amount criterion corresponding to all variables in the variable set.

Wherein the set of variables is obtained by the following formula:

the akabane information content criterion is obtained through the following formula:

in the formula, C ₁ Represented as a set of variables, which is a non-empty set. The AIC is expressed as a chi-chi information amount criterion and is used for expressing the residual information of the output residual corresponding to the residual variable. u. of ₂ Regression residual, p, expressed as output residual of a variable selected in a set of variables ₁ Expressed as the number of variables selected in the variable set. The Chichi pool information quantity criterion is a measurement standard for judging the regression residual error and the number of input model variables.

S250: and acquiring the optimal variable data set according to the variable sequence of all the variables moving into the initial set when the variables are in the maximum mutual information and the size increase and decrease conditions of the corresponding hematid pool information quantity criterion values.

Wherein, all the variables are variables in a variable set, and the optimal variable data set is:

wherein D is ^* Expressed as an optimal variable data set,

as indicated by the L-th variable,

expressed as the (N + 1) th variable, and L ≦ N. Dividing the optimal variable data set to include a first optimal variable data set

Second optimal variable data set

And a third optimal variable data set

And each optimal variable data set comprises a training set S ₁ Test set S ₂ And a verification set S ₃ 。

In the steps, a DBSCAN abnormal point detection method and a Savitzky-Golay smoothing filtering combined method are adopted to carry out data preprocessing on a sample data vector set, processed data are mapped to a [0,1] interval, and then a partial mutual information PMI algorithm is combined to select important modeling variables to obtain a processed optimal variable data set.

Referring to fig. 1 and 3, in an embodiment of the invention, in step S300, a first optimal variable data set

Second optimal variable data set

And a third optimal variable data set

Respectively identifying a gas turbine power multivariable system model based on quantum particle swarm optimization QPSO and DBN based on deep neural network ₁ And a recursive neural network Elman algorithm model. For the convenience of viewing the drawings in the specification, in the embodiment, the QPSO identification model is expressed by QM, and the deep neural network DBN ₁ The model is denoted DM and the recurrent neural network Elman model is denoted EM.

Acquiring a QM model, a DM model and an EM model, and carrying out modeling training on the QM model, the DM model and the EM model, wherein the method mainly comprises the following steps:

s310, acquiring the power multivariable system model structure of the gas turbine, identifying the power multivariable system model structure of the gas turbine, and acquiring the QM model.

The method is characterized in that a QM model structure of a power multivariable system of the gas turbine is identified based on a QPSO (quantum particle swarm optimization), the model parameter estimation has great difficulty due to the uncertain parameters and the structure of a transfer function, the identification model structure is selected to be the most common transfer function form with second-order object characteristics in a control system for the convenience of simplicity and robustness of the model, and the structure is as follows:

wherein G(s) is expressed as a transfer function, K is expressed as an amplification factor, _τ expressed as the delay time, T is expressed as the time constant, e ^-s Expressed as a delay factor, and s is expressed as a base projection variable symbol of the real space function in the complex frequency domain space.

Bonding of

The output variable is quantitatively identified by utilizing QPSO algorithm

And input argument

The structure of the gas turbine power calculation model is as follows:

where Y(s) is the output variable of gas turbine calculated power modulo gas turbine calculated power, U ₁ (s),U ₂ (s),U ₃ (s) are respectively input variables in the optimal variable data set, and the QPSO algorithm needs to identify 9 parameters, namely K ₁ 、K ₂ 、K ₃ 、T ₁ 、T ₂ 、T ₃ 、τ ¹ 、τ ² 、τ ³ And obtaining a group of optimal solutions by searching 9-dimensional space to minimize the fitness value, wherein the fitness function is a mean square error function and is obtained by the following formula:

in the formula, E _rr Expressed as a fitness function, H is expressed as the total number of samples in the optimal variable dataset, with H ≦ m. k is not expressed, the value range is k =1 \8230, H, y (k) is expressed as the actual value of the power of the gas turbine, and y' (k) is expressed as the predicted value of the calculated power model of the gas turbine.

After determining the structure of the power multivariable system model of the gas turbine, the power multivariable system model is determined

And (5) performing identification. The identification process comprises the following steps:

s311, initializing particle positions in a first search space, mapping the particles into a group of identification parameters for identifying a gas turbine calculation power model structure to form an initial particle swarm, and initializing a nine-dimensional optimal position and a nine-dimensional global optimal position of the particles.

Wherein the search space is [50, 500]]The initialization particle is x _i,j (t) particle dimension j, (j =1,2, \8230; 9, particle number i, (i =1,2, \8230; 50), nine-dimensional optimal position of the initialized particle is

The nine-dimensional global optimum position of the initialization particle is g _best ＝[0…0] _1×j 。

S312, obtaining the average optimal position of the nine-dimensional particle swarm according to the initial particle swarm and the optimal positions of the initial particles.

in the formula, C _j (t) represents the average optimal position of the particle group, M is a constant and takes the particle number, i.e., in the present embodiment, M =50, t represents the number of iterations,

expressed as the optimal position in the j-th dimension for the ith particle.

And S313, bringing part of the data sets in the first optimal variable data set, the second optimal variable data set and the third optimal variable data set into the structure of the calculation power model of the identified gas turbine for identification training, acquiring a fitness value, comparing the current fitness value with the fitness of the previous iteration, updating the current particle position to a nine-dimensional optimal position if the fitness value of the current position is smaller than the fitness of the optimal position of the previous iteration, and otherwise, keeping the nine-dimensional optimal position of the previous iteration.

The nine-dimensional optimal particle position updating formula is as follows:

in the formula (I), the compound is shown in the specification,

the nine-dimensional optimal position of the current particle,

expressed as the nine-dimensional optimal position of the particle of the previous iteration, E _rr And (b) as the fitness of the current particle.

Wherein, the fitness value is 55% of each of the first optimal variable data set, the second optimal variable data set and the third optimal variable data set, and then the sum of the mean square deviations is taken.

And S314, acquiring the nine-dimensional global optimal position of the current particle after updating according to the latest nine-dimensional optimal position of the current particle and the fitness of the current particle, comparing the nine-dimensional global optimal position with the nine-dimensional global optimal position of the previous iteration, updating the nine-dimensional global optimal position of the current particle to be the nine-dimensional global optimal position if the current nine-dimensional global optimal position is smaller than the nine-dimensional global optimal position of the previous iteration, and otherwise, keeping the nine-dimensional global optimal position of the previous iteration.

in the formula (I), the compound is shown in the specification,

expressed as the current particle being the latest nine-dimensional global optimum position

S315, acquiring a nine-dimensional attractor according to the latest nine-dimensional optimal position and the nine-dimensional global optimal position of the current particle, determining a nine-dimensional particle convergence region, and updating the nine-dimensional particle position.

Wherein the nine-dimensional attractor is obtained by the following formula:

updating the nine-dimensional particle position is by the following equation:

in the formula, L _i Expressed as a nine-dimensional attractor, [ phi ] expressed as [0,1]]Random number within interval, x _i,j (t) particles denoted as the t-th iteration, x _i,j (t + 1) denotes the particles for the t +1 th iteration, and u is denoted at [0,1]]The number of random distributions within the interval, α, is expressed as the compression expansion factor.

Wherein the compression expansion factor is obtained by the following formula:

wherein, M =50, and the compression expansion factor adopts a linear reduction strategy.

And S316, outputting the optimal particles if the iteration times meet the first termination condition, and returning to the step S312 if the iteration times do not meet the first termination condition.

In this embodiment, the first termination condition of the number of iterations is t =1000. And reversely bringing the identified parameters back to the structure of the gas turbine computer power model to obtain a QM model.

S320, setting the number of input nerves of the deep neural network to be 3, setting the number of output neurons to be 1, and constructing the DM model, wherein the DM model is of a four-layer hidden layer structure, nodes of all hidden layers are set to be [64,71,11,51], the top layer of the DM model is a BP neural network, and the learning rate theta is set to be 0.09.

S330, setting the number of input neurons of the recurrent neural network to be 3, the number of output neurons of the recurrent neural network to be 3, setting the number of output neurons of the recurrent neural network to be 1, constructing the EM model, and setting hidden layer nodes of the recurrent neural network to be 4.

S340, sending the training set to the constructed QM model, the DM model and the EM model for model training.

Wherein, acquiring the DM model and the EM model is the prior art.

Referring to fig. 1 and 4, in an embodiment of the invention, in step S400, acquiring the HM model includes the following steps:

and S410, sending the test set to a trained QM model, a trained DM model and a trained EM model to obtain a prediction result under the test set.

And S420, obtaining an absolute value of an error between the prediction result and an actual value, selecting a model corresponding to the minimum absolute value of the error as a prediction optimal algorithm under a test set, and performing label classification on the test set at the moment, wherein the label of the QM model, the label of the DM model and the label of the EM model are respectively set as a first class, a second class and a third class.

And S430, adding the test set with the marked categories into the original test set again to form a new sample test set.

In this embodiment, 700 sets of samples are randomly sampled. And optimizing a DBN classification model network structure, namely an HM model, by a QPSO algorithm.

S440, setting a deep neural network classification model structure, initializing positions of particles in a second search space, mapping the particles into a group of parameters of the number of nodes of a four-layer hidden layer to form an initial particle swarm, and initializing a four-dimensional optimal position and a four-dimensional global optimal position of the particle swarm.

Wherein, the deep neural network classification model, i.e. the DBN classification model, is a typical structure with four hidden layer structures, the upper layer structure is the extreme learning machine ELM, and the second search space [10, 300] thereof]The particle dimension is j, (j =1,2, \ 8230; 4), the particle number is i, (i =1,2, \ 8230; 50), and the four-dimensional optimal position of the initialized particle is

The four-dimensional global optimum position of the initialization particle is g _best2 ＝[0…0] _1×j 。

S450, obtaining the average optimal position of the four-dimensional particle swarm according to the four-dimensional optimal position of the initialization particle.

Wherein, the average optimal position of the four-dimensional particle swarm is obtained by the following formula:

wherein, C _j2 (t) represents the average optimum position of the four-dimensional particle group, and the number of M is 50.

And S460, bringing a new sample test set into the deep neural network classification model structure model for network training, obtaining the fitness of 10 times of the average precision of cross validation of each particle, comparing the fitness of the current average precision with the fitness of the average precision of the previous iteration, updating the current particle position to the optimal position if the fitness of the current average precision is smaller than the fitness of the average precision of the second optimal position of the previous iteration, and otherwise, keeping the optimal position of the previous iteration.

Wherein, the fitness of the average precision is obtained by the following formula:

Training the actual label under the sample. Wherein the regularization coefficient of the extreme learning machine is set to 0.001.

When the predicted label equals the actual label, F _k (X) has a value of 1, and F is the value of F when the predicted tag is not equal to the actual tag _k The value of (X) is 0.

The particle four-bit optimal position updating formula is as follows:

in the formula (I), the compound is shown in the specification,

expressed as the four-dimensional optimal position of the current particle,

expressed as the four-dimensional optimal position of the particle of the previous iteration, A _rr And (t) is expressed as the fitness of the average precision of the current particle.

S470, according to the four-dimensional optimal position of the current particle and the fitness of the average precision of the current particle, obtaining the four-dimensional global optimal position of the current particle, comparing the four-dimensional global optimal position with the four-dimensional global optimal position of the previous iteration, if the current four-dimensional global optimal position is smaller than the optimal position of the previous iteration, updating the four-dimensional global optimal position of the current particle to the four-dimensional global optimal position, otherwise, keeping the four-dimensional global optimal position of the previous iteration.

Wherein the latest four-dimensional global optimum position is obtained by the following formula:

in the formula (I), the compound is shown in the specification,

represented as the latest four-dimensional global optimal position of the current particle.

And S480, acquiring a four-dimensional attractor according to the four-dimensional optimal position and the four-dimensional global optimal position of the current particle, determining a four-dimensional particle convergence region, and updating the four-dimensional particle position.

Wherein the four-dimensional attractor is obtained by the following formula:

in the formula, L _2i Expressed as a four-dimensional attractor, [ phi ] expressed as [0,1]]Random numbers within the interval.

in formula (II), x' _i,j (t) is expressed as a four-dimensional particle, x 'of the t iteration' _i,j (t + 1) is expressed as a four-dimensional particle for the t +1 th iteration, and u is expressed at [0.1 ]]The number of random distributions within the interval, α, is expressed as the compression expansion factor.

Wherein the compression expansion factor is obtained by the following formula:

S490, if the iteration number satisfies a second termination condition, outputting the optimal four-dimensional particle to obtain an HM model, and if the iteration number does not satisfy the second termination condition, returning to the step S450; and reversely bringing back the output optimal four-dimensional particles to the model structure to obtain the HM model. S4100, sending the verification set to the HM model trained in step S490 for classification, and then selecting the best model of the verification results from the QM model, the DM model and the EM model for prediction.

In this embodiment, the second termination condition of the number of iterations is t =100. And the optimal four-dimensional particles output at the moment are the optimal number of neurons in each layer of the four hidden layers, and the optimal four-dimensional particles are brought back to the HM model structure, so that the HM model is obtained.

And sending the verification set to the HM model for classification, selecting a prediction model capable of obtaining the best gas turbine power prediction result, and then performing prediction by using the corresponding prediction model.

The prediction model comprises a QM model, a DM model and an EM model. The same method also performs validation tests on the second optimal variable data set and the third optimal variable data set.

Referring to FIG. 5, in an embodiment of the present invention, the data of the research is collected from TCS platform of 420MW single-shaft gas turbine combined cycle No. 1 unit of a certain power generation company in China, and the sampling frequency is setFor 1min, 3 groups of data continuously operated within 50 hours are respectively collected as a first optimal variable data set

Second optimal variable data set

Third optimal variable data set

Each group contains 3000 groups of samples, and the training set S1, the test set S2, and the verification set S3 are divided, and data information is shown in table 1.

Table 1 experimental data information

The PMI algorithm is adopted to screen the modeling candidate input variables, and the screening result is shown in FIG. 5. Index is represented as a variable label, labels 1-6 are rotor axial displacement, compressor inlet pressure, compressor inlet temperature, compressor pressure ratio, gas turbine exhaust temperature and gas turbine rotating speed in sequence, and the akali pool information criterion AIC calculated value represents that screening of independent variables is an optimal modeling independent variable set when the akali pool information criterion AIC value changes to the minimum value. As can be seen from FIG. 5, the Akabar information criterion AIC value reaches a minimum of-25869.9363 when the variables are screened to the gas turbine exhaust temperature. Selecting the maximum information entropy variable in the previous 3 screening processes: the compressor pressure ratio, the compressor inlet temperature, and the gas turbine exhaust temperature are the final modeling variables.

By mean relative error (MAPE), mean Square Error (MSE), correlation coefficient (R) ² ) Accuracy (AC) performance evaluation erythroid pool information criterion AIC by the following formula:

in the formula, m is the number of samples; x is the number of _p Calculating a power actual value for the combustion engine; x' _p Calculating a power prediction value for the combustion engine;

calculating the average value of the actual power values; a is the number of correct classifications; d is the number of error classifications. MAPE, MSE and R2 respectively represent the reliability, precision and fitting degree of the model prediction result, the smaller the MAPE and MSE values are, the better the MAPE and MSE values are, and the closer R2 is to 1, the better the model prediction result is.

Referring to fig. 6, in an embodiment of the present invention, in order to verify the performance of the prediction method of the single-shaft combined cycle gas turbine power, a QM model, a DM model and an EM model are used, where reference numeral 100 is a fitness error curve of a particle swarm optimization PSO and reference numeral 200 is a fitness error curve of a QPSO algorithm. Wherein, table 2 shows the calculated power of the combustion engine related to u identified by QPSO ₁ T1, totc model, the QM model estimation procedure will

The first 55% of data is used as a training set for parallel identification, and is verified through respective 300 groups of test sets, the fitness function is the sum of mean square deviations of 3 groups of data, the QPSO iteration frequency is set to be 1000, the population particles are 50, K ₁ 、K ₂ 、K ₃ ∈[-50,50],T ₁ 、T ₂ 、T ₃ ∈[0,1000],τ ₁ 、τ ₂ 、τ ₃ ∈[0,100]Based on the test set, the MSEs of the 3 groups of the identification performance results are respectively 0.9754,1.6911、2.0183。

TABLE 2 QM model identification parameters

Fig. 6 is a fitness error evolution curve of the QPSO algorithm and the particle swarm optimization algorithm PSO identification QM model, and it can be seen from fig. 6 that the optimization process of the QPSO algorithm on 9 parameters is better than the PSO algorithm, and in the 1500 iterations, the iteration speed and the optimization accuracy of the QPSO algorithm are obviously higher than those of the PSO algorithm.

Table 3 shows the performance parameter evaluation index results of the QM model, DM model, EM model, and HM model. As can be seen from Table 3, the HM model is in the first optimal variable data set

Second optimal variable data set

And a third optimal variable data set

In which a better prediction result is obtained, wherein the first optimal variable data set

MAPE, MSE and fitting coefficient R of medium HM model to EM model ² The values are respectively improved by 36.01 percent, 55.54 percent and 0.17 percent, and the prediction precision is obviously improved. At the second optimal variable data set

In the QM model, the DM model and the EM model, due to the expansion of the data change range and the acceleration of the data change frequency, the MSE value and the MAPE value are obviously enlarged in the model prediction result, and the HM model is in the MAPE, MSE and R states compared with the EM model ² The values are respectively improved by 40.18 percent, 51.04 percent and 2.15 percent. In the third optimal variable data set

Fitting coefficient R of middle, QM model, DM model and EM model ² All are less than 0.8, the fitting effect is poor, and the HM model is better than the DM models MAPE, MSE and R ² The values are respectively improved by 20.44%, 31.26% and 8.68% compared with the first optimal variable data set

And a second optimal variable data set

R ² The improvement range is obvious, and the adaptability of the adaptive selection prediction of the HM model is verified through the data set.

TABLE 3 evaluation index results for different algorithms

Fig. 7 (a) to 7 (f) further illustrate the absolute advantage of the HM model in the form of a line graph and a three-dimensional bar graph, and as can be seen from fig. 7 (b), 7 (d), and 7 (f), the number of absolute error values between the prediction result and the actual value of the HM model in the range of 1 or less is greater than that of other models, 222, 134, and 111 are provided, and the number of absolute error values falling in the range of more than 5 is also smaller than that of other models, 4, 23, and 24 are provided. As can be seen from fig. 7 (a), 7 (c), and 7 (f), the prediction result of the HM model better conforms to the variation trend of the actual value, and the prediction accuracy is improved by adaptively selecting model prediction, so that the purpose of selecting a proper algorithm for modeling prediction under different samples is achieved.

Table 4 is a confusion matrix of classification results of the HM model in the first optimal variable data set

Second optimal variable data set

Third optimal variable dataCollection

The medium AC values are all over 80 percent, which shows that the HM model can accurately select a modeling algorithm adaptive to the problem and the modeling algorithm is applied to the first optimal variable data set

In the above example, the number of correctly identified and classified samples is 93, 83, and 84, respectively, and 15 samples belonging to the EM model are erroneously identified as 8 QM models and 7 DM models. 24 samples belonging to the DM model were misidentified as 7 QM models and 17 EM models. The accuracy of the HM model is affected, while only 1 of the QM models is misidentified. The number of the QM, EM and DM models which are wrongly identified by the HM model in the D2 data set is only 9, 11 and 8, and the accuracy reaches 90.67 percent. The recognition rate of the HM model to the model in the data set D3 is low, and the accuracy rate is only 84%.

TABLE 4 confusion matrix of classification results of HM algorithm

The classification advantages of the proposed HM algorithm model are further verified through table 5, and the classification capability of the DBN model optimized by the QPSO algorithm, that is, the HM model, is significantly better than that of the DBN model without optimization, and the accuracy is higher than about 10%. The C4.5 algorithm is a typical classification algorithm and is used and improved in many application requirements, and as can be seen from table 5, the classification accuracy in the D1, D2, and D3 data sets is 69.56%, 80.67%, and 65.89%, and the average accuracy is 72.04%, which are not higher than the accuracy and average accuracy of the HM algorithm. Through the analysis of the classification results of the test data sets, the usability of the algorithm is verified.

TABLE 5 confusion matrix accuracy for classification results of different algorithms

It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein, and any reference signs in the claims are not intended to be construed as limiting the claim concerned.

The above-mentioned embodiments only represent embodiments of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the concept of the present invention, and these embodiments are all within the protection scope of the present invention.

Claims

1. A method of predicting power for a single shaft combined cycle gas turbine, comprising the steps of:

acquiring a sample data vector set according to original data of a power monitoring variable of the gas turbine;

the method comprises the steps that abnormal points of a sample data vector set are removed and smooth wave processing is conducted, important monitoring variables are selected by combining a partial mutual information algorithm, an optimal variable data set is obtained, the optimal variable data set is divided into a first optimal variable data set, a second optimal variable data set and a third optimal variable data set, and each optimal variable data set comprises a training set, a testing set and a verification set;

2. The method for predicting power of a single-shaft combined cycle gas turbine according to claim 1, wherein said obtaining optimal variable data sets comprises the steps of:

s240, putting the residual variable into the initial set to obtain an updated initial set, obtaining a variable set according to the ii-th element characteristic and the residual variable, returning to execute the steps S220-S240 according to the updated initial set and the variable set until the variable set is empty, and obtaining the Chichi information quantity criterion corresponding to all variables in the variable set;

3. The method for predicting power of a single-shaft combined cycle gas turbine according to claim 2, wherein the mutual information of the ii-th elemental signature and the N + 1-th elemental signature in the initial data set is obtained by the following formula:

in the formula (I), the compound is shown in the specification,

expressed as the ii th element feature in the initial dataset

And the (N + 1) th element feature

F (-) is expressed as an estimated density function of m samples, x _hl Expressed as samples, y, of the h row and l column of the output matrix _l The corresponding output characteristic represented as the h row, l column;

the input residual error is obtained by the following formula:

the output residual is obtained by the following formula:

the set of variables is obtained by the following formula:

in the formula, C ₁ Expressed as a set of variables, AIC as the Chi-pool information content criterion, u ₂ Regression residual, p, expressed as output residual of a variable selected in a set of variables ₁ Expressed as the number of variables selected in the variable set, and m is expressed as a sample.

4. The method of predicting power of a single-shaft combined cycle gas turbine as claimed in claim 1, wherein identifying a gas turbine power multivariate system model structure based on quantum-swarm algorithm, obtaining a QM model, comprises the steps of:

s311, initializing particle positions in a first search space, mapping the particles into a group of identification parameters for identifying a gas turbine calculation power model structure to form an initial particle swarm, and initializing a nine-dimensional optimal position and a nine-dimensional global optimal position of the particles;

5. The method of predicting power in a single-shaft combined cycle gas turbine as set forth in claim 4, wherein said gas turbine computational power model structure is:

wherein Y(s) is the output variable of the gas turbine power model gas turbine power, U ₁ (s),U ₂ (s),U ₃ (s) input variables, K, respectively, in the optimal variable data set ₁ 、K ₂ 、K ₃ 、T ₁ 、T ₂ 、T ₃ 、τ ¹ 、τ ² 、τ ³ Expressed as the parameter to be identified, e ^-s Expressed as a delay factor, s is expressed as realThe spatial function projects the variable symbols at the base of the complex frequency domain space.

6. The single shaft combined cycle gas turbine power prediction method of claim 4, wherein,

an optimal position in the j dimension represented as the ith particle;

the nine-dimensional optimal particle position updating formula is as follows:

in the formula (I), the compound is shown in the specification,

the nine-dimensional optimal position of the current particle,

in the formula (I), the compound is shown in the specification,

the nine-dimensional attractor is obtained by the following formula:

updating the nine-dimensional particle position is by the following equation:

Wherein the compression expansion factor is obtained by the following formula:

7. the method of predicting single-shaft combined cycle gas turbine power of claim 1, wherein obtaining the HM model comprises the steps of:

s420, obtaining an absolute value of an error between the prediction result and an actual value, selecting a model corresponding to the minimum absolute value of the error as a prediction optimal algorithm under a test set, and performing label classification on the test set at the moment, wherein the QM model, the DM model and the EM model are respectively set as a first class, a second class and a third class;

s430, adding the test set with the marked category into the original test set again to form a new sample test set;

s450, obtaining the average optimal position of the four-dimensional particle swarm according to the four-dimensional optimal position of the initialization particle;

s460, bringing a new sample test set into the deep neural network classification model structure for network training, obtaining the fitness of 10 times of cross validation average precision of each particle, comparing the fitness of the current average precision with the fitness of the average precision of the previous iteration, if the fitness of the current average precision is smaller than the fitness of the average precision of the second optimal position of the previous iteration, updating the current particle position to the optimal position, otherwise, keeping the optimal position of the previous iteration;

s470, acquiring a four-dimensional global optimal position of the current particle according to the four-dimensional optimal position of the current particle and the fitness of the average precision of the current particle, comparing the four-dimensional global optimal position with the four-dimensional global optimal position of the previous iteration, if the current four-dimensional global optimal position is smaller than the optimal position of the previous iteration, updating the four-dimensional global optimal position of the current particle to the four-dimensional global optimal position, otherwise, keeping the four-dimensional global optimal position of the previous iteration;

s480, acquiring a four-dimensional attractor according to the four-dimensional optimal position and the four-dimensional global optimal position of the current particle, determining a four-dimensional particle convergence region, and updating the four-dimensional particle position;

s490, if the iteration number satisfies a second termination condition, outputting the optimal four-dimensional particle to obtain an HM model, and if the iteration number does not satisfy the second termination condition, returning to the step S450; carrying back the output optimal four-dimensional particles to the model structure to obtain the HM model;

8. The single shaft combined cycle gas turbine power prediction method of claim 7, wherein,

the fitness of the average precision is obtained by the following formula:

in the formula, A _rr Fitness expressed as mean accuracy, f _L (x _i,j (t)) is expressed in the new sample test set

Prediction label, V, obtained by prediction of extreme learning machine under training sample _{tran_label} Expressed as in a new sample test set

the particle four-bit optimal position updating formula is as follows:

in the formula (I), the compound is shown in the specification,

expressed as the four-dimensional optimal position of the current particle,

expressed as the four-dimensional optimal position of the particle of the previous iteration, A _rr A fitness expressed as the average precision of the current particles;

in the formula (I), the compound is shown in the specification,

the four-dimensional attractor is obtained by the following formula:

in formula (II), x' _i,j (t) is represented as the four-dimensional particle, x 'of the t iteration' _i,j (t + 1) represents the four-dimensional particle for the t +1 th iteration.

9. The method of predicting single shaft combined cycle gas turbine power of claim 8, wherein F is the actual label when the predicted label is equal to the actual label _k (X) has a value of 1, and F is the value of 1 when the predicted tag is not equal to the actual tag _k The value of (X) is 0.

10. The single shaft combined cycle gas turbine power prediction method of claim 4, where the first termination condition is t =1000.