A method for predicting the power of a single-shaft combined cycle gas turbine

Abstract

The application discloses a single-shaft combined cycle gas turbine power prediction method, comprising the following steps: obtaining optimal variable data sets, wherein each optimal variable data set comprises a training set, a test set and a verification set; obtaining a QM model based on a quantum particle swarm algorithm, a DM model and an EM model based on a deep neural network and a recurrent neural network algorithm according to the optimal variable data sets, and respectively performing modeling training on the training set; obtaining the absolute error between the prediction result and the actual value of the above-mentioned model, and selecting the model corresponding to the minimum absolute error as the prediction optimal algorithm under the test set to obtain a sample test set; setting a deep neural network classification model, performing optimization based on a quantum particle swarm algorithm, obtaining 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 submodel according to different working conditions to predict the gas turbine power. The application can improve the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of gas turbine power generation technology, specifically a method for predicting the power of a single-shaft combined cycle gas turbine. Background Technology

[0002] Single-shaft gas-steam combined cycle (FSC) generators offer advantages such as short start-up and shutdown times, strong environmental adaptability, rapid load response, and high thermodynamic efficiency, leading to their large-scale development in the domestic market compared to other energy generation methods. The calculated power of the gas turbine directly affects the coordinated optimization of the entire cycle. In existing single-shaft FSC units, the gas turbine and steam turbine share a single generator. The power measurement device monitors the gas turbine power value in two parts: first, it uses the SSS turbine shaft clutch engagement signal (KUPE). When the engagement signal KUPE = 0, the gas turbine power measurement is the actual value measured by the dynamometer. During normal operation of the combined cycle, when the engagement signal KUPE = 1, the gas turbine power measurement is the calculated value. This calculated value is obtained mathematically through logical mechanisms and relevant operating parameters. While the calculation method is accurate, it cannot pre-calculate the gas turbine power value, resulting in a lag. Therefore, building a reliable gas turbine performance prediction model will directly provide advanced, accurate, and fully controllable parameters for combined cycle power generation, ensuring the economic efficiency and safety of the unit.

[0003] The core of constructing accurate prediction models lies in finding suitable modeling methods for the research problem. Existing research has identified three main modeling methods: traditional physical analysis, statistical methods, and data-driven advanced intelligent methods. Among existing technologies, Cao Jun et al. published "Development of a High-Precision Full-Range Simulation System for F-Class Gas-Steam Combined Cycle Units," utilizing APROS software to perform mass and momentum analysis on the gas turbine's thermodynamic processes and conducting real-time dynamic simulation of the full-range processes of F-class gas-steam units following the law of energy conservation. However, this mechanistic method has room for improvement in modeling accuracy. The Supervisory Information System (SIS), as a subsystem of smart power plant control, enables the monitoring, recording, and real-time access to a large amount of parameter data in unit process control, providing a guarantee for advanced intelligent modeling and control strategies. Li Jingxuan et al. published "Research on Hybrid Modeling Method of Gas Turbine Mechanism-Data," designing a hybrid model controller method using intelligent algorithms for mechanistic model error compensation, and conducting verification experiments based on Distributed Control System (DCS) data for different combination designs, thus improving prediction accuracy. Yun Shih-hao et al. published "Research on Sensor Fault Diagnosis of Gas Turbine Control System Based on RS-RBF," which uses an improved equal-frequency discretization method to process sensor fault symptom attributes to construct rough sets, and then builds an RBF network to reduce the false judgment rate of sensor faults. Although previous studies have achieved success in power plant parameter modeling, these studies are all shallow machine learning methods, which cannot capture the deep-seated useful information hidden in the underlying data, resulting in insufficient prediction accuracy. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the issue of low power prediction accuracy in single-shaft combined cycle gas turbines.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] Based on the raw data of the gas turbine power monitoring variables, obtain a sample data vector set;

[0007] The sample data vector set is first subjected to outlier removal and smoothing, and then important monitoring variables are selected by combining the partial mutual information algorithm to obtain the optimal variable dataset. The optimal variable dataset is then divided into the first optimal variable dataset, the second optimal variable dataset, and the third optimal variable dataset, and each optimal variable dataset includes a training set, a test set, and a validation set.

[0008] Based on the optimal variable dataset, the structure of the gas turbine power multivariable system model is identified using the quantum particle swarm optimization algorithm to obtain the QM model. The DM model and EM model are obtained using deep neural network and recurrent neural network algorithms, and modeling training is performed on the training set respectively.

[0009] The QM, DM, and EM models are validated using the test set. The absolute errors between the prediction results and actual values ​​of the QM, DM, and EM models are obtained. The model with the smallest absolute error is selected as the optimal prediction algorithm for the test set. The test set is then labeled and classified. The labeled test set is added back to the test set to obtain a sample test set. A deep neural network classification model structure is set and optimized based on the quantum particle swarm optimization algorithm to obtain the HM model. The HM model is trained using the sample test set and validated on the validation set. Sub-models are adaptively selected for gas turbine power prediction according to different operating conditions.

[0010] Advantages: The raw data acquired from the TCS data acquisition system is processed using mechanistic mathematical analysis and the PMI screening method to obtain the optimal variable dataset. This optimal dataset is then divided into three optimal datasets: a first optimal dataset, a second optimal dataset, and a third optimal dataset. Each optimal dataset includes a training set, a test set, and a validation set. The structure of the multivariate system model for gas turbine power is identified based on the quantum particle swarm optimization algorithm, obtaining the QM model. The DM model and EM model are obtained based on deep neural network and recurrent neural network algorithms. Modeling and prediction are performed on the first optimal variable dataset. An adaptive selection algorithm model, namely the HM model, is constructed using the QPSO-DBN algorithm for adaptive prediction of gas turbine calculated power. The prediction process is not based on a single algorithm prediction result, but rather on the sample operating conditions. The QPSO-DBN algorithm is used to adaptively select the sub-models (QM, DM, and EM models) to obtain more accurate prediction results, resulting in high prediction accuracy.

[0011] In one embodiment of the present invention, obtaining the optimal variable dataset includes the following steps:

[0012] S210, take the sample data vector set as the initial dataset, obtain the mutual information of the i-th element feature and the N+1-th element feature in the initial dataset, and select the element feature corresponding to the maximum mutual information and move it into the initial set.

[0013] S220, in the i-th element feature and the N+1-th element feature of the initial dataset, remove each feature in the i-th element feature that is affected by the initial set to obtain the input residual, and remove each feature in the N+1-th element feature that is affected by the initial set to obtain the output residual;

[0014] S230, based on the input residual and the output residual, obtain the residual mutual information, and find the residual variable when the residual mutual information is maximized;

[0015] S240, put the residual variable into the initial set, obtain the updated initial set, and obtain the variable set according to the feature of the ii-th element and the residual variable. Then, according to the updated initial set and the variable set, return to execute steps S220 to S240 until the variable set is empty, and obtain the Akaike information content criteria corresponding to all variables in the variable set.

[0016] S250, based on the order in which all variables are moved into the initial set at the point of maximum mutual information, and the increase or decrease of the corresponding Akaike information criterion value, the optimal variable dataset is obtained.

[0017] In one embodiment of the present invention, the mutual information between the i-th element feature and the N+1-th element feature in the initial dataset is determined by the following formula:

[0018]

[0019] In the formula, Represented as the feature of the i-th element in the initial dataset Features of the (N+1)th element The mutual information, f(·), is denoted as the estimated density function for m samples, x hl Let y be the sample in the h-th row and l-th column of the output matrix. l This is represented as the output feature corresponding to the h-th row and l-th column;

[0020] The input residual is obtained using the following formula:

[0021]

[0022] The output residual is obtained using the following formula:

[0023]

[0024] In the formula, V represents the input residual, U represents the output residual, E(·) represents the conditional expectation, and S represents the initial set;

[0025] The set of variables is obtained through the following formula:

[0026]

[0027] The Akaike information content criterion is obtained through the following formula:

[0028]

[0029] In the formula, C1 represents the variable set, AIC represents the Akaike Information Criterion, u2 represents the regression residual of the output residual of the variable selected in the variable set, p1 represents the number of variables selected in the variable set, and m represents the sample.

[0030] In one embodiment of the present invention, the structure of a gas turbine power multivariable system model is identified based on the quantum particle swarm optimization algorithm to obtain a QM model, including the following steps:

[0031] S311, initialize the particle positions in the first search space, and map the particles to a set of identification parameters for identifying the structure of the gas turbine power calculation model to form an initial particle swarm, while initializing the nine-dimensional optimal position and the nine-dimensional global optimal position of the particles.

[0032] S312, based on the initial particle swarm and the optimal position of the initialized particles, obtain the average optimal position of the nine-dimensional particle swarm;

[0033] S313, a portion of the datasets from the first optimal variable dataset, the second optimal variable dataset, and the third optimal variable dataset are input into the identification gas turbine power calculation model structure for identification training to obtain fitness values. The current fitness value is compared with the fitness value of the previous iteration. If the fitness value of the current position is less than the fitness value of the optimal position in the previous iteration, the current particle position is updated to the nine-dimensional optimal position; otherwise, the nine-dimensional optimal position of the previous iteration is maintained.

[0034] S314. Based on the latest nine-dimensional optimal position of the current particle and the fitness of the current particle, obtain the updated nine-dimensional global optimal position of the current particle. Compare 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 less than the nine-dimensional global optimal position of the previous iteration, then update the nine-dimensional global optimal position of the current particle to the nine-dimensional global optimal position; otherwise, keep the nine-dimensional global optimal position of the previous iteration.

[0035] S315: Based on the latest nine-dimensional optimal position and the nine-dimensional global optimal position of the current particle, obtain the nine-dimensional attractor, determine the convergence region of the nine-dimensional particle, and update the position of the nine-dimensional particle.

[0036] S316, If the number of iterations meets the first termination condition, then output the optimal particle; if the number of iterations does not meet the first termination condition, then return to step S312.

[0037] S317, the output optimal particle is brought back to the gas turbine power calculation model structure to obtain the QM model.

[0038] In one embodiment of the present invention, the structure of the gas turbine power calculation model is as follows:

[0039]

[0040] In the formula, Y(s) represents the output variable of the gas turbine calculated power model, U1(s), U2(s), and U3(s) are the input variables in the optimal variable dataset, and K1, K2, K3, T1, T2, T3, τ1, τ2, and τ3 represent the parameters to be identified. -s Let s be the delay factor, and let s be the symbol for the basis projection variable of the real space function in the complex frequency domain.

[0041] In one embodiment of the present invention, wherein,

[0042] The search space is [50, 500], the particle dimension is j (j = 1, 2, ... 9), and the number of particles is i (i = 1, 2, ... 50).

[0043] The average optimal position of the nine-dimensional particle swarm is obtained by the following formula:

[0044]

[0045] 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 number of particles, M = 50, and t represents the number of iterations. This represents the optimal position of the i-th particle in the j-th dimension.

[0046] The formula for updating the optimal position of a particle in nine dimensions is:

[0047]

[0048] In the formula, The current optimal position of the particle in nine dimensions. E represents the nine-dimensional optimal position of the particle in the previous iteration. rr (.) represents the fitness of the current particle;

[0049] The optimal position in nine dimensions is obtained using the following formula:

[0050]

[0051] In the formula, This indicates that the current particle is at the latest nine-dimensional global optimal position;

[0052] The nine-dimensional attractor is obtained using the following formula:

[0053]

[0054] The position of the nine-dimensional particle is updated using the following formula:

[0055]

[0056] In the formula, L i Represented as a nine-dimensional attractor, φ represents a random number in the interval [0,1], and x i,j Let x(t) represent the particle in the t-th iteration. i,j (t+1) represents the particle in the (t+1)th iteration, u represents the random distribution number in the interval [0,1], and α represents the compression expansion factor.

[0057] The compressibility expansion factor is obtained using the following formula:

[0058]

[0059] In one embodiment of the present invention, obtaining the HM model includes the following steps:

[0060] S410, The test set is sent to the trained QM model, DM model and EM model to obtain the prediction results under the test set;

[0061] S420, obtain the absolute value of the error between the prediction result and the actual value, select the model with the smallest absolute value of error as the optimal prediction algorithm under the test set, and label and classify the test set at this time, wherein the category labels of QM model, DM model and EM model are set as the first category, the second category and the third category respectively;

[0062] S430, The test set with the labeled categories is added back to the original test set to form a new sample test set;

[0063] S440, set the deep neural network classification model structure, initialize the position of the particles in the second search space, and map the particles to a set of parameters of the number of nodes in the four hidden layers to form an initial particle swarm, and initialize the four-dimensional optimal position and the four-dimensional global optimal position of the particle swarm.

[0064] S450, based on the four-dimensional optimal position of the initialized particles, obtain the average optimal position of the four-dimensional particle swarm;

[0065] S460, the new sample test set is brought into the deep neural network classification model structure for network training, and the fitness of the average precision of each particle with 10 times the cross-validation is obtained. The fitness of the current average precision is compared with the fitness of the average precision of the previous iteration. If the fitness of the current average precision is less than the fitness of the average precision of the second optimal position in the previous iteration, the current particle position is updated to the optimal position. Otherwise, the optimal position of the previous iteration is maintained.

[0066] S470: Based on the current particle's four-dimensional optimal position and the fitness of the current particle's average precision, obtain the current particle's four-dimensional global optimal position. Compare 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 less than the optimal position of the previous iteration, update the current particle's four-dimensional global optimal position to the four-dimensional global optimal position; otherwise, maintain the four-dimensional global optimal position of the previous iteration.

[0067] S480: Based on the current four-dimensional optimal position and the four-dimensional global optimal position of the particle, obtain the four-dimensional attractor, determine the convergence region of the four-dimensional particle, and update the position of the four-dimensional particle.

[0068] S490, if the number of iterations meets the second termination condition, the optimal four-dimensional particle is output to obtain the HM model; if the number of iterations does not meet the second termination condition, return to step S450; the optimal four-dimensional particle is returned to the model structure to obtain the HM model.

[0069] S4100, the validation set is sent to the HM model trained in step S490 for classification, and then the model with the best validation result among the QM model, the DM model and the EM model is selected for prediction.

[0070] In one embodiment of the present invention, wherein,

[0071] The second search space is [10, 300], the particle dimension is j (j = 1, 2, ... 4), and the number of particles is i (i = 1, 2, ... 50);

[0072] The average optimal position of the four-dimensional particle swarm is obtained by the following formula:

[0073]

[0074] In the formula, C j2 (t) represents the average optimal position of the four-dimensional particle swarm, and M has 50 possible values.

[0075] The fitness of average accuracy is obtained using the following formula:

[0076]

[0077] In the formula, A rr The fitness, expressed as average accuracy, is f. L (x i,j (t) represents the new sample test set The predicted label V obtained by the extreme learning machine under the training samples tran_label This is represented as in the new sample test set. The actual labels of the training samples, F k (X) is a real value, determined by the prediction label being equal to the actual label, where the regularization coefficient of the Extreme Learning Machine is set to 0.001;

[0078] The formula for updating the optimal position of the four particles is:

[0079]

[0080] In the formula, This represents the current optimal four-dimensional position of the particle. Let A be the four-dimensional optimal position of the particle in the previous iteration. rr (.) represents the fitness of the current average particle accuracy;

[0081] The formula for obtaining the latest four-dimensional global optimal position is:

[0082]

[0083] In the formula, This represents the latest four-dimensional global optimal position of the current particle;

[0084] The four-dimensional attractor is obtained using the following formula:

[0085]

[0086] In the formula, L 2i It is represented as a four-dimensional attractor, and φ is represented as a random number in the interval [0, 1].

[0087] The updated four-dimensional particle position is obtained using the following formula:

[0088]

[0089] In the formula, x' i,j Let x'(t) represent the four-dimensional particle in the t-th iteration. i,j (t+1) represents the four-dimensional particle in the (t+1)th iteration.

[0090] In one embodiment of the present invention, when the predicted label equals the actual label, F kThe value of (X) is 1 when the predicted label is not equal to the actual label, F k The value of (X) is 0.

[0091] In one embodiment of the present invention, the first termination condition is t = 1000.

[0092] Compared with the prior art, the beneficial effects of the present invention are: by using a hybrid adaptive selection algorithm model, namely the HM model, to predict the performance of the calculated power of the gas turbine, it can adaptively select sub-models for different operating conditions to predict the calculated power, and the accuracy of the calculated power prediction of the single-shaft gas turbine is significantly higher than that of other algorithms, thus improving the prediction accuracy. Attached Figure Description

[0093] Figure 1 This is a flowchart illustrating a method for predicting the power of a single-shaft combined cycle gas turbine, as described in an embodiment of the present invention.

[0094] Figure 2 This is a flowchart illustrating the process of obtaining the optimal variable dataset according to an embodiment of the present invention.

[0095] Figure 3 This is a flowchart illustrating the acquisition of the QM model according to an embodiment of the present invention.

[0096] Figure 4 This is a flowchart illustrating the acquisition of the HM model according to an embodiment of the present invention.

[0097] Figure 5 This is a schematic diagram of the PMI variable selection process according to an embodiment of the present invention.

[0098] Figure 6 This is a comparison chart of the optimization results of the PSO algorithm and the QPSO algorithm in an embodiment of the present invention.

[0099] Figures 7(a) to (b) are schematic diagrams of the prediction results of different prediction algorithms in the first optimal variable dataset according to the embodiments of the present invention.

[0100] Figures 7(c) to (d) are schematic diagrams of the prediction results of different prediction algorithms in the second optimal variable dataset according to the embodiments of the present invention.

[0101] Figures 7(e) to (f) are schematic diagrams of the prediction results of different prediction algorithms in the third optimal variable dataset according to the embodiments of the present invention. Detailed Implementation

[0102] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0103] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0104] Please see Figure 1 and Figure 2 As shown, this invention provides a method for predicting the power of a single-shaft combined cycle gas turbine, comprising the following steps:

[0105] S100: Obtain a sample data vector set based on the raw data of the gas turbine power monitoring variables;

[0106] S200, the sample data vector set is first subjected to outlier removal and smoothing, and then important monitoring variables are selected by combining the partial mutual information algorithm to obtain the optimal variable dataset. The optimal variable dataset is then divided into the first optimal variable dataset, the second optimal variable dataset, and the third optimal variable dataset, and each optimal variable dataset includes a training set, a test set, and a validation set.

[0107] S300, based on the optimal variable dataset, the structure of the gas turbine power multivariable system model is identified using the quantum particle swarm optimization algorithm, and the QM model is obtained. Based on the deep neural network and recurrent neural network algorithms, the DM model and EM model are obtained, and modeling training is performed on the training set respectively.

[0108] S400: Validate the QM model, DM model, and EM model using the test set; obtain the absolute error between the prediction results and actual values ​​of the QM model, DM model, and EM model; select the model with the smallest absolute error as the optimal prediction algorithm for the test set; label and classify the test set at this point; add the labeled test set back to the test set to obtain a sample test set; set a deep neural network classification model and optimize it based on the quantum particle swarm optimization algorithm to obtain the HM model; train the HM model using the sample test set; validate the trained HM model on the validation set; and adaptively select a sub-model for gas turbine power prediction according to different operating conditions.

[0109] Please see Figure 1As shown, in one embodiment of the present invention, before predicting the power of a single-shaft combined cycle gas turbine in step S100, the relevant parameter values ​​for the calculated power of the gas turbine are first determined based on the mathematical relationship of the internal logic mechanism of the gas turbine. In a single-shaft combined cycle unit, the gas turbine and steam turbine share a generator. The monitoring of the gas turbine power value by its power measurement device can be divided into two parts: the measured value by the dynamometer and the calculated power value of the gas turbine. The calculated power value of the gas turbine can be expressed by the following formula:

[0110] P GT =min(P EL max(10.5984, f4(T1)×PGT) 70 )), (1);

[0111] In the formula, P GT Power, P, is expressed as a measured value of the gas turbine power. EL PGT represents the measured power value of the dynamometer. 70 The calculated gas turbine power value is represented by f4(T1), where T1 represents the compressor inlet temperature, and f4(T1) is the conversion function of the compressor inlet temperature. The calculated gas turbine power value is the power value corresponding to the compressor outlet pressure, obtained through the following formula:

[0112]

[0113] In the formula, p1 represents the compressor inlet pressure, and f1(x), f2(x), and f3(x) represent the gas turbine exhaust temperatures T. otc The conversion function to compressor pressure ratio u1 is obtained through the following formula:

[0114]

[0115]

[0116] f3(x)=f3(T otc -579×1.015+a x )×0.7, (5);

[0117] In the formula, a x It has no actual meaning; it simply takes a value of 0 or -4, and its specific value is determined based on the rotor's axial displacement. Among these, the gas turbine exhaust temperature T... otc Obtain it using the following formula:

[0118] T otc =y1+y2+y3, (6);

[0119] y1 = 0.5B + ​​0.5C, (7);

[0120] y2=-(T1×0.01625+0.0115)×T1, (8);

[0121]

[0122] In the formula, B and C represent the average values ​​of the exhaust temperatures of the 24 gas turbines using two thermocouples, and n represents the gas turbine speed. When n > 47.5 Hz, At that time, y3 = 5. The values ​​of y1 to y3 have no actual meaning; they are only used to obtain the gas turbine exhaust temperature T. otc The intermediate variable is f5(x), which represents the gas turbine exhaust temperature T. otc The conversion function with compressor pressure ratio u1, and its f1(x)~f5(x) respectively correspond to linear conversion functions in different ranges of values.

[0123] Based on the above mathematical quantitative relationship of the internal logic mechanism of the gas turbine TCS, it can be seen that the calculated power of the gas turbine is mainly related to the compressor inlet temperature, compressor inlet pressure, compressor pressure ratio, gas turbine exhaust temperature, rotor axial displacement, and gas turbine speed. Therefore, the above 6 variables are selected as candidate input variables for modeling, and the calculated power of the gas turbine is selected as the output variable.

[0124] The raw data of candidate input and output variables for modeling are obtained from the TCS data acquisition system, and a sample dataset DATA = X∪Y is established, where X represents the features and Y represents the output features. N represents the number of features; in this embodiment, N = 6. The feature contains m values, i.e., m samples. The output feature is Y = {y1, y2, ..., y...}. m}, where each value corresponds to Calculate the power value of the gas turbine for each sample. Construct the input-output matrix. The input-output matrix D is then rewritten as a sample data vector set consisting of N+1 column vectors. in, This represents the first element feature in the sample data vector set. It is actually the first column vector in the input-output matrix D, corresponding to the first feature of the sample dataset.

[0125] Please see Figure 1 and Figure 2 As shown, in one embodiment of the present invention, step S200, obtaining the optimal variable dataset, includes the following steps:

[0126] S210, take the sample data vector set as the initial dataset, obtain the mutual information of the i-th element feature and the N+1-th element feature in the initial dataset, and select the element feature corresponding to the maximum mutual information and move it into the initial set.

[0127] The mutual information between the features of the i-th element and the (N+1)-th element in the initial dataset is obtained through the following formula:

[0128]

[0129] In the formula, Represented as the feature of the i-th element in the initial dataset Features of the (N+1)th element The mutual information, f(·), is denoted as the estimated density function for m samples, x hl Let y be the sample in the h-th row and l-th column of the output matrix. l Let S be the output feature corresponding to the h-th row and l-th column. The initial set is denoted by S, and initially, the initial set is empty.

[0130] S220, in the i-th element feature and the N+1-th element feature of the initial dataset, remove each feature in the i-th element feature that is affected by the initial set to obtain the input residual, and remove each feature in the N+1-th element feature that is affected by the initial set to obtain the output residual.

[0131] The input residual is obtained using the following formula:

[0132]

[0133] The output residual is obtained using the following formula:

[0134]

[0135] In the formula, V represents the input residual, which is the feature of the remaining i-th element in the sample data vector set D'. Each feature in the dataset is the input residual after removing the influence of the initial set information. U represents the output residual, which is the (N+1)th element feature in this data vector set D'. The output residual of each feature after removing the influence of the initial set information. E(·) represents the conditional expectation.

[0136] S230, based on the input residual and the output residual, obtain the residual mutual information, and find the residual variable when the residual mutual information is maximized.

[0137] According to formula (10), the residual mutual information I(V,U) is obtained, where the residual variables are represented by... express.

[0138] S240, put the residual variable into the initial set, obtain the updated initial set, and obtain the variable set according to the feature of the ii-th element and the residual variable. Then, according to the updated initial set and the variable set, return to execute steps S220 to S240 until the variable set is empty, and obtain the Akaike information content criteria corresponding to all variables in the variable set.

[0139] The set of variables is obtained through the following formula:

[0140]

[0141] The Akaike information content criterion is obtained through the following formula:

[0142]

[0143] In the formula, C1 represents the variable set, which is a non-empty set. AIC represents the Akaike Information Criterion, used to represent the residual information of the output residuals corresponding to the residual variables. u2 represents the regression residual of the output residuals of the variables selected from the variable set, and p1 represents the number of variables selected from the variable set. The Akaike Information Criterion is a measure of the regression residuals relative to the number of variables in the input model.

[0144] S250: Based on the order in which all variables are moved into the initial set at the point of maximum mutual information, and the increase or decrease of the corresponding Akaike information criterion value, obtain the optimal variable dataset.

[0145] Wherein, all the variables mentioned are variables in the variable set, and the optimal variable dataset is: Among them, D * Represented as the optimal variable dataset, Let it be the Lth variable. Let L be the (N+1)th variable, and L≤N. Divide the optimal variable dataset into two parts, including the first optimal variable dataset. Second optimal variable dataset and the third optimal variable dataset Each optimal variable dataset includes a training set S1, a test set S2, and a validation set S3.

[0146] In the above steps, the sample data vector set is preprocessed by combining the DBSCAN outlier detection method with Savitzky-Golay smoothing filtering, and the processed data is mapped to the [0,1] interval. Then, the partial mutual information (PMI) algorithm is used to select important modeling variables to obtain the optimal variable dataset after processing.

[0147] Please see Figure 1 andFigure 3 As shown, in one embodiment of the present invention, in step S300, the first optimal variable dataset... Second optimal variable dataset and the third optimal variable dataset A multivariate system model for gas turbine power is identified using the Quantum Particle Swarm Optimization (QPSO) algorithm, and predictions are made using deep neural network (DBN1) and recurrent neural network (Elman) algorithms. For ease of reference to the accompanying drawings, in this embodiment, the QPSO identification model is represented by QM, the DBN1 model by DM, and the Elman model by EM.

[0148] Obtaining the QM model, DM model, and EM model, and training them, mainly includes the following steps:

[0149] S310, obtain the structure of the gas turbine power multivariable system model, identify the structure of the gas turbine power multivariable system model, and obtain the QM model.

[0150] The structure of the QM model for a gas turbine power multivariable system was identified based on the quantum particle swarm optimization (QPSO) algorithm. Due to the uncertainty of the transfer function parameters and the structural characteristics, parameter estimation is challenging. To ensure model simplicity and robustness, the chosen model structure is the most common transfer function form in control systems, exhibiting second-order object characteristics. Its structure is as follows:

[0151]

[0152] In the formula, G(s) represents the transfer function, and K represents the amplification factor. τ Let T represent the delay time, and e represent the time constant. -s Let s be the delay factor, and let s be the symbol for the basis projection variable of the real space function in the complex frequency domain.

[0153] Combination The variables in the data are quantitatively identified using the QPSO algorithm. With input independent variables The mathematical relationship between them, and the structure of its gas turbine power calculation model are as follows:

[0154]

[0155] In the formula, Y(s) represents the output variable of the gas turbine calculated power model, and U1(s), U2(s), and U3(s) are the input variables in the optimal variable dataset, respectively. The QPSO algorithm needs to identify nine parameters, namely K1, K2, K3, T1, T2, T3, τ1, τ2, and τ3. By searching the nine-dimensional space, an optimal solution is obtained to minimize the fitness value. The fitness function is the mean squared error function, which is obtained through the following formula:

[0156]

[0157] In the formula, E rr Let H be the fitness function, where H represents the total number of samples in the optimal variable dataset, and H≦m. k has no meaning and its range is k=1…H. y(k) represents the actual power of the gas turbine, and y'(k) represents the predicted power of the gas turbine calculated by the power model.

[0158] After determining the structure of the multivariable system model for gas turbine power, its K1, K2, K3, T1, T2, T3, and τ are... 1 τ 2 τ 3 Identification is performed. The identification process includes the following steps:

[0159] S311, initialize the particle positions in the first search space, and map the particles to a set of identification parameters for identifying the structure of the gas turbine power calculation model to form an initial particle swarm, while initializing the nine-dimensional optimal position and the nine-dimensional global optimal position of the particles.

[0160] The search space is [50, 500], and the initial particle is x. i,j (t), the particle dimension is j, (j=1,2,…9, the number of particles is i, (i=1,2,…50), and the optimal nine-dimensional position of the initialized particle is... The initial nine-dimensional global optimal position of the particle is g. best =[0…0] 1×j .

[0161] S312, based on the initial particle swarm and the optimal position of the initialized particles, obtain the average optimal position of the nine-dimensional particle swarm.

[0162] The average optimal position of the nine-dimensional particle swarm is obtained by the following formula:

[0163]

[0164] In the formula, C j (t) represents the average optimal position of the particle swarm, M is a constant, and its value is the number of particles. In this embodiment, M = 50, and t represents the number of iterations. It represents the optimal position of the i-th particle in the j-th dimension.

[0165] S313, a portion of the datasets from the first optimal variable dataset, the second optimal variable dataset, and the third optimal variable dataset are input into the identification gas turbine power calculation model structure for identification training to obtain fitness values. The current fitness value is compared with the fitness value of the previous iteration. If the fitness value of the current position is less than the fitness value of the optimal position in the previous iteration, the current particle position is updated to the nine-dimensional optimal position; otherwise, the nine-dimensional optimal position of the previous iteration is maintained.

[0166] The formula for updating the optimal position of a particle in nine dimensions is as follows:

[0167]

[0168] In the formula, The current optimal position of the particle in nine dimensions. E represents the nine-dimensional optimal position of the particle in the previous iteration. rr (.) represents the fitness of the current particle.

[0169] The fitness value is calculated by taking the sum of the mean squared errors of 55% of the data from the first, second, and third optimal variable datasets.

[0170] S314. Based on the latest nine-dimensional optimal position of the current particle and the fitness of the current particle, obtain the updated nine-dimensional global optimal position of the current particle. Compare 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 less than the nine-dimensional global optimal position of the previous iteration, then update the nine-dimensional global optimal position of the current particle to the nine-dimensional global optimal position; otherwise, keep the nine-dimensional global optimal position of the previous iteration.

[0171] The optimal position in the nine dimensions is obtained using the following formula:

[0172]

[0173] In the formula, This indicates that the current particle is at the latest nine-dimensional global optimal position.

[0174] S315: Based on the latest nine-dimensional optimal position and the nine-dimensional global optimal position of the current particle, obtain the nine-dimensional attractor, determine the convergence region of the nine-dimensional particle, and update the position of the nine-dimensional particle.

[0175] The nine-dimensional attractor is obtained through the following formula:

[0176]

[0177] The position of the nine-dimensional particle is updated using the following formula:

[0178]

[0179] In the formula, L i Represented as a nine-dimensional attractor, φ represents a random number in the interval [0,1], and x i,j Let x(t) represent the particle in the t-th iteration. i,j (t+1) represents the particle in the (t+1)th iteration, u represents the random distribution number in the interval [0,1], and α represents the compression expansion factor.

[0180] The compressibility expansion factor is obtained using the following formula:

[0181]

[0182] Where M = 50, the compression expansion factor adopts a linear reduction strategy.

[0183] S316. If the number of iterations meets the first termination condition, output the optimal particle; if the number of iterations does not meet the first termination condition, return to step S312.

[0184] S317, the output optimal particle is brought back to the gas turbine power calculation model structure to obtain the QM model.

[0185] In this embodiment, the first termination condition for the number of iterations is t = 1000. The identified parameters are then back-substituted into the gas turbine power calculation model structure to obtain the QM model.

[0186] S320, Set the number of input neurons of the deep neural network to 3 and the number of output neurons to 1 to construct the DM model. The DM model is a four-layer hidden layer structure with each hidden layer node set to [64,71,11,51]. Its top layer is a BP neural network with a learning rate θ set to 0.09.

[0187] S330, the number of input neurons of the recurrent neural network is set to 3, the number of output neurons is set to 3, and the number of output neurons is set to 1 to construct the EM model, and the number of hidden layer nodes of the recurrent neural network is set to 4.

[0188] S340, the training set is sent to the constructed QM model, DM model and EM model for model training.

[0189] Among them, obtaining the DM model and EM model is an existing technology.

[0190] Please see the sameFigure 1 and Figure 4 As shown, in one embodiment of the present invention, obtaining the HM model in step S400 includes the following steps:

[0191] S410, the test set is sent to the trained QM model, DM model and EM model to obtain the prediction results under the test set.

[0192] S420, obtain the absolute value of the error between the predicted result and the actual value, select the model corresponding to the smallest absolute value of the error as the optimal prediction algorithm under the test set, and label and classify the test set at this time, wherein the category labels of QM model, DM model and EM model are set to the first category, the second category and the third category respectively.

[0193] S430, the test set with the labeled categories is added back to the original test set to form a new sample test set.

[0194] In this embodiment, 700 sets of samples are randomly selected. The DBN classification model network structure, i.e., the HM model, is optimized using the QPSO algorithm.

[0195] S440 sets up the deep neural network classification model structure, initializes the position of particles in the second search space, and maps the particles to a set of parameters of the number of nodes in the four hidden layers to form an initial particle swarm. It also initializes the four-dimensional optimal position and the four-dimensional global optimal position of the particle swarm.

[0196] The deep neural network classification model, also known as the DBN classification model, is a typical structure with four hidden layers. The upper layer is an Extreme Learning Machine (ELM), with a second search space of [10, 300]. The particle dimension is j (j = 1, 2, ..., 4), and the number of particles is i (i = 1, 2, ..., 50). The optimal four-dimensional position of the initialized particles is... The initial four-dimensional global optimal position of the particle is g. best2 =[0…0] 1×j .

[0197] S450, based on the four-dimensional optimal position of the initialized particles, obtain the average optimal position of the four-dimensional particle swarm.

[0198] The average optimal position of the four-dimensional particle swarm is obtained by the following formula:

[0199]

[0200] Among them, C j2 (t) represents the average optimal position of the four-dimensional particle swarm, and M has 50 possible values.

[0201] S460, the new sample test set is brought into the deep neural network classification model structure model for network training, and the fitness of the average precision of each particle with 10 times the cross-validation is obtained. The fitness of the current average precision is compared with the fitness of the average precision of the previous iteration. If the fitness of the current average precision is less than the fitness of the average precision of the second optimal position in the previous iteration, the current particle position is updated to the optimal position. Otherwise, the optimal position of the previous iteration is maintained.

[0202] The fitness of average accuracy is obtained using the following formula:

[0203]

[0204] In the formula, A rr The fitness, expressed as average accuracy, is f. L (x i,j (t) represents the new sample test set The predicted label V obtained by the extreme learning machine under the training samples tran_label This is represented as in the new sample test set. The actual labels of the training samples. The regularization coefficient for the Extreme Learning Machine is set to 0.001.

[0205] When the predicted label equals the actual label, F k The value of (X) is 1 when the predicted label is not equal to the actual label, F k The value of (X) is 0.

[0206] The formula for updating the optimal position of the four particles is:

[0207]

[0208] In the formula, This represents the current optimal four-dimensional position of the particle. Let A be the four-dimensional optimal position of the particle in the previous iteration. rr (.) represents the fitness of the current average particle accuracy.

[0209] S470: Based on the current particle's four-dimensional optimal position and the fitness of the current particle's average precision, obtain the current particle's four-dimensional global optimal position. Compare the current 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 less than the optimal position of the previous iteration, update the current particle's four-dimensional global optimal position to the current four-dimensional global optimal position; otherwise, maintain the four-dimensional global optimal position of the previous iteration.

[0210] The latest four-dimensional global optimal position will be obtained using the following formula:

[0211]

[0212] In the formula, This represents the latest four-dimensional global optimal position of the current particle.

[0213] S480: Based on the current four-dimensional optimal position and the four-dimensional global optimal position of the particle, obtain the four-dimensional attractor, determine the convergence region of the four-dimensional particle, and update the position of the four-dimensional particle.

[0214] The four-dimensional attractor is obtained using the following formula:

[0215]

[0216] In the formula, L 2i It is represented as a four-dimensional attractor, and φ is represented as a random number in the interval [0, 1].

[0217] The updated four-dimensional particle position is obtained using the following formula:

[0218]

[0219] In the formula, x' i,j Let x'(t) represent the four-dimensional particle in the t-th iteration. i,j (t+1) represents the four-dimensional particle in the (t+1)th iteration, u represents the random distribution number in the interval [0.1], and α represents the compression expansion factor.

[0220] The compressibility expansion factor is obtained using the following formula:

[0221]

[0222] Where M = 50, the compression expansion factor adopts a linear reduction strategy.

[0223] S490: If the number of iterations meets the second termination condition, the optimal four-dimensional particle is output to obtain the HM model. If the number of iterations does not meet the second termination condition, return to step S450; the optimal four-dimensional particle is returned to the model structure to obtain the HM model. S4100: The validation set is sent to the HM model trained in step S490 for classification. Then, the model with the best validation result among the QM model, the DM model, and the EM model is selected for prediction.

[0224] In this embodiment, the second termination condition for the number of iterations is t = 100. The optimal four-dimensional particle output at this time is the optimal number of neurons in each of the four hidden layers. The optimal four-dimensional particle is then fed back into the HM model structure to obtain the HM model.

[0225] The validation set is sent to the HM model for classification, and the prediction model that can obtain the best gas turbine power prediction results is selected. Then, the corresponding prediction model is used for prediction.

[0226] The prediction models mentioned above are the QM model, DM model, and EM model. The same method was also validated on the second and third optimal variable datasets.

[0227] Please see Figure 5 As shown, in one embodiment of the present invention, the data in this study was collected from the TCS platform of Unit 1 of a 420MW single-shaft gas turbine combined cycle unit of a domestic power generation company. The sampling frequency was set to 1 minute, and three sets of data from continuous operation over 50 hours were collected as the first optimal variable dataset. Second optimal variable dataset Third optimal variable dataset Each group contains 3000 samples and is divided into training set S1, test set S2 and validation set S3. The data information is shown in Table 1.

[0228] Table 1 Experimental Data Information

[0229]

[0230] The PMI algorithm was used to screen candidate input variables for modeling. The screening results are as follows: Figure 5 As shown. Index represents the variable label, with labels 1-6 representing rotor axial displacement, compressor inlet pressure, compressor inlet temperature, compressor pressure ratio, gas turbine exhaust temperature, and gas turbine speed, respectively. The Akaike Information Criterion (AIC) is used to calculate the value; when the AIC value reaches its minimum, it indicates that the set of independent variables is optimal for modeling. From... Figure 5 It can be seen that the Akaike Information Criterion (AIC) value reaches its minimum when the variable is selected as the gas turbine exhaust temperature, with a minimum value of -25869.9363. The variables with the highest information entropy in the first three selection processes—compressor pressure ratio, compressor inlet temperature, and gas turbine exhaust temperature—are selected as the final modeling variables.

[0231] Through mean relative error (MAPE), mean squared error (MSE), and correlation coefficient (R²), 2 The Accuracy (AC) performance evaluation criterion, Akaike Information Criterion (AIC), is based on the following formula:

[0232]

[0233]

[0234]

[0235]

[0236] In the formula, m is the sample size; x p Calculate the actual power output of the gas turbine; x' p Calculate the predicted power output for the gas turbine; To calculate the average of the actual power values; A represents the number of correctly classified cases; D represents the number of misclassified cases. MAPE, MSE, and R² represent the reliability, accuracy, and fit of the model prediction results, respectively. The smaller the values ​​of MAPE and MSE, the better, and the closer R² is to 1, the better.

[0237] Please see Figure 6 As shown, in one embodiment of the present invention, to verify the performance of the prediction method for the power of a single-shaft combined cycle gas turbine, the QM model, DM model, and EM model are used. Label 100 represents the fitness error curve of the Particle Swarm Optimization (PSO) algorithm, and label 200 represents the fitness error curve of the QPSO algorithm. Table 2 shows the optimal parameters for the calculated power of the gas turbine identified by QPSO with respect to the u1, T1, and Totc models. The QM model estimation process will... The first 55% of the data was used as the training set for parallel identification, and validated using 300 test sets for each set. The fitness function was the sum of the mean squared errors of the three sets of data. The number of QPSO iterations was set to 1000, the population particle count was 50, K1, K2, K3 ∈ [-50, 50], T1, T2, T3 ∈ [0, 1000], τ1, τ2, τ3 ∈ [0, 100]. Based on the test set, the MSE results of the three identification performances were 0.9754, 1.6911, and 2.0183, respectively.

[0238] Table 2 QM Model Identification Parameters

[0239]

[0240] Figure 6 To identify the fitness error evolution curves of the QM model for the QPSO algorithm and the Particle Swarm Optimization (PSO) algorithm, from... Figure 6 It can be seen that the QPSO algorithm is superior to the PSO algorithm in the optimization process of the nine parameters. In the process of 1500 iterations, the iteration speed and optimization accuracy of the QPSO algorithm are significantly higher than those of the PSO algorithm.

[0241] Table 3 shows the performance parameter evaluation results for the QM model, DM model, EM model, and HM model. As can be seen from Table 3, the HM model is the best in the first optimal variable dataset. Second optimal variable dataset and the third optimal variable dataset Both datasets achieved relatively good prediction results, particularly in the first optimal variable dataset. The HM model outperforms the EM model in terms of MAPE, MSE, and fit coefficient R. 2 The values ​​increased by 36.01%, 55.54%, and 0.17% respectively, indicating a significant improvement in prediction accuracy. This was achieved on the second optimal variable dataset. In the QM, DM, and EM models, due to the expanded range and increased frequency of data changes, the MSE and MAPE values ​​in the model prediction results are significantly larger. The HM model, compared to the EM model, shows higher values ​​in MAPE, MSE, and R². 2 The values ​​increased by 40.18%, 51.04%, and 2.15%, respectively. This is in the third optimal variable dataset. In the figure, the fitting coefficients R of the QM model, DM model, and EM model are... 2 All values ​​are less than 0.8, indicating a poor fit. The HM model performs better than the DM model in terms of MAPE, MSE, and R. 2 The values ​​increased by 20.44%, 31.26%, and 8.68% respectively, compared to the first optimal variable dataset. With the second optimal variable dataset R 2 The improvement is significant, and the above datasets verify the adaptability of the HM model in adaptively selecting predictions.

[0242] Table 3 Evaluation metrics results for different algorithms

[0243]

[0244] Figures 7(a) to 7(f) The absolute advantages of the HM model are further illustrated using line charts and 3D bar charts. Figure 7(b) , 7(d) As can be seen from 7(f), the HM model has more absolute error values ​​less than or equal to 1 than other models, with 222, 134, and 111 respectively. The number of values ​​greater than 5 is also smaller for the HM model, with 4, 23, and 24 respectively. From... Figure 7(a) , 7(c) As can be seen from 7(f), the prediction results of the HM model are more consistent with the changing trend of the actual values. By adaptively selecting the model for prediction, the prediction accuracy is improved, and the goal of selecting the appropriate algorithm for modeling and prediction under different samples is achieved.

[0245] Table 4 shows the confusion matrix of the classification results of the HM model, in the first optimal variable dataset. Second optimal variable dataset Third optimal variable dataset The AC values ​​are all above 80%, indicating that the HM model can accurately select a modeling algorithm suitable for the problem, in the first optimal variable dataset. In the D2 dataset, the number of correctly identified and classified samples were 93, 83, and 84 respectively. 15 samples belonging to the EM model were incorrectly identified as 8 samples as the QM model and 7 as the DM model. 24 samples belonging to the DM model were incorrectly identified as 7 samples as the QM model and 17 as the EM model. This affected the accuracy of the HM model, while only 1 sample of the QM model was incorrectly identified. In the D3 dataset, the number of QM, EM, and DM models incorrectly identified by the HM model were only 9, 11, and 8 respectively, achieving an accuracy of 90.67%. The HM model had a lower recognition rate in the D3 dataset, with an accuracy of only 84%.

[0246] Table 4. Confusion matrix of HM algorithm classification results

[0247]

[0248] Table 5 further verifies the classification advantage of the proposed HM algorithm model. The DBN model optimized by the QPSO algorithm, i.e., the HM model, has a significantly better classification ability than the unoptimized DBN model, with an accuracy rate consistently higher than 10%. The C4.5 algorithm is a typical classification algorithm that has been used and improved in many applications. As shown in Table 5, the classification accuracy rates in datasets D1, D2, and D3 are 69.56%, 80.67%, and 65.89%, respectively, with an average accuracy rate of 72.04%, all of which are not higher than the accuracy rate and average accuracy rate of the HM algorithm. The analysis of the classification results on the above experimental datasets verifies the usability of the proposed algorithm.

[0249] Table 5. Confusion Matrix and Accuracy of Different Algorithms in Classification Results

[0250]

[0251] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0252] The above-described embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above-described embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for predicting the power of a single-shaft combined cycle gas turbine, characterized in that, Includes the following steps: Based on the raw data of the gas turbine power monitoring variables, obtain a sample data vector set; The sample data vector set is first subjected to outlier removal and smoothing, and then important monitoring variables are selected by combining the partial mutual information algorithm to obtain the optimal variable dataset. The optimal variable dataset is then divided into the first optimal variable dataset, the second optimal variable dataset, and the third optimal variable dataset, and each optimal variable dataset includes a training set, a test set, and a validation set. Based on the optimal variable dataset, the structure of the gas turbine power multivariable system model is identified using the quantum particle swarm optimization algorithm to obtain the QM model. The DM model and EM model are obtained using deep neural network and recurrent neural network algorithms, and modeling training is performed on the training set respectively. The QM, DM, and EM models are validated using the test set. The absolute errors between the prediction results and actual values ​​of the QM, DM, and EM models are obtained. The model with the smallest absolute error is selected as the optimal prediction algorithm for the test set. The test set is then labeled and classified. The labeled test set is added back to the test set to obtain a sample test set. A deep neural network classification model structure is set and optimized based on the quantum particle swarm optimization algorithm to obtain the HM model. The HM model is trained using the sample test set and validated on the validation set. Sub-models are adaptively selected for gas turbine power prediction according to different operating conditions.

2. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 1, characterized in that, Obtaining the optimal variable dataset includes the following steps: S210, take the sample data vector set as the initial dataset, obtain the mutual information of the i-th element feature and the N+1-th element feature in the initial dataset, and select the element feature corresponding to the maximum mutual information and move it into the initial set. S220, in the i-th element feature and the N+1-th element feature of the initial dataset, remove each feature in the i-th element feature that is affected by the initial set to obtain the input residual, and remove each feature in the N+1-th element feature that is affected by the initial set to obtain the output residual; S230, based on the input residual and the output residual, obtain the residual mutual information, and find the residual variable when the residual mutual information is maximized; S240, put the residual variable into the initial set, obtain the updated initial set, and obtain the variable set according to the feature of the ii-th element and the residual variable. Then, according to the updated initial set and the variable set, return to execute steps S220 to S240 until the variable set is empty, and obtain the Akaike information content criteria corresponding to all variables in the variable set. S250, based on the order in which all variables are moved into the initial set at the point of maximum mutual information, and the increase or decrease of the corresponding Akaike information criterion value, the optimal variable dataset is obtained.

3. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 2, characterized in that, in, The mutual information between the i-th element feature and the (N+1)-th element feature in the initial dataset is obtained through the following formula: In the formula, Represented as the feature of the i-th element in the initial dataset Features of the (N+1)th element The mutual information, f(·), is denoted as the estimated density function for m samples, x hl Let y be the sample in the h-th row and l-th column of the output matrix. l This is represented as the output feature corresponding to the h-th row and l-th column; The input residual is obtained using the following formula: The output residual is obtained using the following formula: In the formula, V represents the input residual, U represents the output residual, E(·) represents the conditional expectation, and S represents the initial set; The set of variables is obtained through the following formula: The Akaike information content criterion is obtained through the following formula: In the formula, C1 represents the variable set, AIC represents the Akaike Information Criterion, u2 represents the regression residual of the output residual of the variable selected in the variable set, p1 represents the number of variables selected in the variable set, and m represents the sample.

4. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 1, characterized in that, The structure of a multivariable gas turbine power system model is identified based on the quantum particle swarm optimization algorithm, and the QM model is obtained. The steps include: S311, initialize the particle positions in the first search space, and map the particles to a set of identification parameters for identifying the structure of the gas turbine power calculation model to form an initial particle swarm, while initializing the nine-dimensional optimal position and the nine-dimensional global optimal position of the particles. S312, based on the initial particle swarm and the optimal position of the initialized particles, obtain the average optimal position of the nine-dimensional particle swarm; S313, a portion of the datasets from the first optimal variable dataset, the second optimal variable dataset, and the third optimal variable dataset are input into the identification gas turbine power calculation model structure for identification training to obtain fitness values. The current fitness value is compared with the fitness value of the previous iteration. If the fitness value of the current position is less than the fitness value of the optimal position in the previous iteration, the current particle position is updated to the nine-dimensional optimal position; otherwise, the nine-dimensional optimal position of the previous iteration is maintained. S314. Based on the latest nine-dimensional optimal position of the current particle and the fitness of the current particle, obtain the updated nine-dimensional global optimal position of the current particle. Compare 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 less than the nine-dimensional global optimal position of the previous iteration, then update the nine-dimensional global optimal position of the current particle to the nine-dimensional global optimal position; otherwise, keep the nine-dimensional global optimal position of the previous iteration. S315: Based on the latest nine-dimensional optimal position and the nine-dimensional global optimal position of the current particle, obtain the nine-dimensional attractor, determine the convergence region of the nine-dimensional particle, and update the position of the nine-dimensional particle. S316, If the number of iterations meets the first termination condition, then output the optimal particle; if the number of iterations does not meet the first termination condition, then return to step S312. S317, the output optimal particle is brought back to the gas turbine power calculation model structure to obtain the QM model.

5. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 4, characterized in that, in, The structure of the gas turbine power calculation model is as follows: In the formula, Y(s) represents the output variable of the gas turbine calculated power model, U1(s), U2(s), and U3(s) are the input variables in the optimal variable dataset, and K1, K2, K3, T1, T2, T3, τ1, τ2, and τ3 represent the parameters to be identified. -s Let s be the delay factor, and let s be the symbol for the basis projection variable of the real space function in the complex frequency domain.

6. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 4, characterized in that, in, The search space is [50, 500], the particle dimension is j (j = 1, 2, ... 9), and the number of particles is i (i = 1, 2, ... 50). The average optimal position of the nine-dimensional particle swarm is obtained by 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 number of particles, M = 50, and t represents the number of iterations. This represents the optimal position of the i-th particle in the j-th dimension. The formula for updating the optimal position of a particle in nine dimensions is: In the formula, The current optimal position of the particle in nine dimensions. E represents the nine-dimensional optimal position of the particle in the previous iteration. rr (.) represents the fitness of the current particle; The optimal position in nine dimensions is obtained using the following formula: In the formula, This indicates that the current particle is at the latest nine-dimensional global optimal position; The nine-dimensional attractor is obtained using the following formula: The position of the nine-dimensional particle is updated using the following formula: In the formula, L i Represented as a nine-dimensional attractor, φ represents a random number in the interval [0,1], and x i,j Let x(t) represent the particle in the t-th iteration. i,j (t+1) represents the particle in the (t+1)th iteration, u represents the random distribution number in the interval [0,1], and α represents the compression expansion factor; The compressibility expansion factor is obtained using the following formula:

7. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 1, characterized in that, Obtaining the HM model includes the following steps: S410, The test set is sent to the trained QM model, DM model and EM model to obtain the prediction results under the test set; S420, obtain the absolute value of the error between the prediction result and the actual value, select the model with the smallest absolute value of error as the optimal prediction algorithm under the test set, and label and classify the test set at this time, wherein the category labels of QM model, DM model and EM model are set as the first category, the second category and the third category respectively; S430, The test set with the labeled categories is added back to the original test set to form a new sample test set; S440, set the deep neural network classification model structure, initialize the position of the particles in the second search space, and map the particles to a set of parameters of the number of nodes in the four hidden layers to form an initial particle swarm, and initialize the four-dimensional optimal position and the four-dimensional global optimal position of the particle swarm. S450, based on the four-dimensional optimal position of the initialized particles, obtain the average optimal position of the four-dimensional particle swarm; S460, the new sample test set is brought into the deep neural network classification model structure for network training, and the fitness of the average precision of each particle with 10 times the cross-validation is obtained. The fitness of the current average precision is compared with the fitness of the average precision of the previous iteration. If the fitness of the current average precision is less than the fitness of the average precision of the second optimal position in the previous iteration, the current particle position is updated to the optimal position. Otherwise, the optimal position of the previous iteration is maintained. S470: Based on the current particle's four-dimensional optimal position and the fitness of the current particle's average precision, obtain the current particle's four-dimensional global optimal position. Compare 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 less than the optimal position of the previous iteration, update the current particle's four-dimensional global optimal position to the four-dimensional global optimal position; otherwise, maintain the four-dimensional global optimal position of the previous iteration. S480: Based on the current four-dimensional optimal position and the four-dimensional global optimal position of the particle, obtain the four-dimensional attractor, determine the convergence region of the four-dimensional particle, and update the position of the four-dimensional particle. S490, if the number of iterations meets the second termination condition, the optimal four-dimensional particle is output to obtain the HM model; if the number of iterations does not meet the second termination condition, return to step S450; the optimal four-dimensional particle is returned to the model structure to obtain the HM model. S4100, the validation set is sent to the HM model trained in step S490 for classification, and then the model with the best validation result among the QM model, the DM model and the EM model is selected for prediction.

8. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 7, characterized in that, in, The second search space is [10, 300], the particle dimension is j (j = 1, 2, ... 4), and the number of particles is i (i = 1, 2, ... 50); The average optimal position of the four-dimensional particle swarm is obtained by the following formula: In the formula, C j2 (t) represents the average optimal position of the four-dimensional particle swarm, and M has 50 possible values. The fitness of average accuracy is obtained using the following formula: In the formula, A rr The fitness, expressed as average accuracy, is f. L (x i,j (t) represents the new sample test set The predicted label V obtained by the extreme learning machine under the training samples tran_label This is represented as in the new sample test set. The actual labels of the training samples, F k (X) is a real value, determined by the prediction label being equal to the actual label, where the regularization coefficient of the Extreme Learning Machine is set to 0.001; The formula for updating the optimal position of the four particles is: In the formula, This represents the current optimal four-dimensional position of the particle. Let A be the four-dimensional optimal position of the particle in the previous iteration. rr (.) represents the fitness of the current average particle accuracy; The formula for obtaining the latest four-dimensional global optimal position is: In the formula, This represents the latest four-dimensional global optimal position of the current particle; The four-dimensional attractor is obtained using the following formula: In the formula, L 2i It is represented as a four-dimensional attractor, and φ is represented as a random number in the interval [0, 1]. The updated four-dimensional particle position is obtained using the following formula: In the formula, x' i,j Let x'(t) represent the four-dimensional particle in the t-th iteration. i,j (t+1) represents the four-dimensional particle in the (t+1)th iteration.

9. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 8, characterized in that, When the predicted label equals the actual label, F k The value of (X) is 1 when the predicted label is not equal to the actual label, F k The value of (X) is 0.

10. The method for predicting the power of a single-shaft combined cycle gas turbine according to claim 4, characterized in that, The first termination condition is t = 1000.

Images