IVYA-PINN-based intelligent modeling method for process characteristics of heavy duty gas turbine

Abstract

The invention discloses an IVYA-PINN-based intelligent modeling method for process characteristics of a heavy duty gas turbine, which is used for realizing accurate modeling of the process characteristics of the heavy duty gas turbine. According to the method, on the basis of actual operation data of a certain heavy-duty gas turbine, operation characteristics of the heavy-duty gas turbine are analyzed, a physical differential equation of each component is obtained, and a physical loss item is provided for a physical information neural network. Through Spearman correlation analysis, input and output variables of a neural network are selected, a relationship between the input and output variables is established based on a PRform network, data loss and physical loss are used as a total loss function, and the network is guided to perform parameter updating. The weight coefficient of each loss item is optimized through the ivy algorithm, different loss parts are effectively balanced, it is guaranteed that the model meets the physical law, and meanwhile the interpretability and generalization ability of the model are enhanced. The invention provides a scheme and a thought for solving the problem of precise modeling of the process characteristics of the heavy duty gas turbine.

Description

Technical Field

[0001] The present invention relates to the technical field of heavy-duty gas turbine process characteristic modeling and optimization, and in particular to an IVYA-PINN-based intelligent modeling method for heavy-duty gas turbine process characteristics. Technical Background

[0002] As the national energy development strategy pays more and more attention to clean and low-carbon goals, the transformation of the power structure to clean and efficient is imminent. Combined cycle units use natural gas as an energy source, and generate electricity through the efficient cooperation of gas turbine systems and steam turbine systems. During the power generation process, the emission of polluting gases such as sulfur oxides and nitrogen oxides is low, which can effectively help achieve clean and low-carbon goals. As one of the key components of combined cycle units, the stability and efficiency of the operation of heavy-duty gas turbines are of great significance to the entire power generation process. Therefore, how to achieve safe and stable operation of heavy-duty gas turbine systems is an urgent problem to be solved.

[0003] An accurate controlled object model is the basis for achieving efficient and stable operation of a heavy-duty gas turbine system. However, the system has the characteristics of multivariable coupling and strong nonlinearity. The traditional pure data-driven modeling method lacks direct constraints on physical laws, which may cause the model to produce unrealistic results in practical applications. At the same time, pure data-driven modeling often requires a large amount of historical operating data as support, and the actual performance is heavily dependent on the quality of the training data. In actual engineering, it is difficult to avoid the situation where the data available for collection is limited or the data quality is low, which may cause the network to be unable to fully learn the process characteristics of the object, and the accuracy of the established model is low.

[0004] The traditional process of using neural networks for model identification tends to produce black box models, which have certain limitations in terms of model interpretability and generalization ability. Physical information neural networks embed prior knowledge in physics into neural networks and impose constraints during the training process by appropriately penalizing the loss function of traditional neural networks. In addition, neural networks can also derive the physical equations corresponding to the data by learning some known observation data, thereby improving the accuracy and efficiency of solving differential equations. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent modeling method for heavy-duty gas turbine process characteristics based on IVYA-PINN to achieve accurate modeling of heavy-duty gas turbines. The method takes the historical operating data of a heavy-duty gas turbine as input, combines the physical model with the data, and adds a physical loss term during the training process to ensure that the physical constraints can be met with a small amount of training data, and has good generalization ability while ensuring accuracy.

[0006] The technology adopted in the present invention is an intelligent modeling method for process characteristics of heavy-duty gas turbines based on IVYA-PINN, which is specifically implemented according to the following steps:

[0007] The operating characteristics of heavy-duty gas turbines were analyzed, the Spearman correlation coefficient was used to conduct correlation analysis on the input variables, and the input and output variables required to establish the heavy-duty gas turbine were selected.

[0008] The formula for calculating the Spearman correlation coefficient is:

[0009]

[0010] Among them, R(x i ) and R(y i ) are the values ​​of the two variables, and are the mean values ​​of variables x and y respectively.

[0011] The input variables are determined to be natural gas flow rate, IGV opening, compressor exhaust pressure, speed, load, and exhaust temperature, and the output variables are compressor exhaust pressure, speed, load, and exhaust temperature. The historical operating data of the corresponding variables are obtained and the data is preprocessed.

[0012] Data preprocessing includes data noise reduction, missing value filling, and standardization. The specific contents are as follows:

[0013] Data noise reduction: Use filtering algorithms to process random errors, outliers or irrelevant information introduced during data collection, transmission or processing to improve data quality;

[0014] Missing value filling: fill the missing values ​​in the collected data by using the average value of the sample points before and after the missing value sample point;

[0015] Standardization: The Z-score standardization method is used to standardize the data set. The specific formula is:

[0016]

[0017] Among them, x norm is the value after standardization, x is the value before standardization, and x mean is the mean of the original data, x std is the standard deviation of the original data.

[0018] A heavy-duty gas turbine model identification framework based on physical information neural network is established.

[0019] Furthermore, the specific process of establishing the model and training the model is as follows:

[0020] The model of a heavy-duty gas turbine is mainly composed of three modules: a compressor, a combustion chamber and a turbine. The compressor compresses the air sucked in from the outside step by step, stores a certain amount of energy, and transports it to the combustion chamber to mix and burn with the injected fuel, turning it into high-temperature and high-pressure combustion gas. The high-temperature combustion gas flows through the turbine to expand and do work, pushing the turbine to drive the compressor and the load rotor to rotate at high speed to generate electricity.

[0021] When the compressor has a certain structure size, its compression ratio and efficiency mainly depend on the speed, flow rate and intake conditions. azh , equivalent speed N yzh 、Pressure ratio π y and efficiency η y The functional relationship of these four similarity criteria is used to characterize the universal characteristic curve of the compressor. The specific formula is:

[0022]

[0023] A differential equation with pressure and enthalpy as state parameters is established to reflect the dynamic change process of the gas in the combustion chamber. The mass balance equation is:

[0024]

[0025] G a =B1(P r1 -P)

[0026] G g =B2(PP r2 )

[0027] Where, P is the combustion chamber pressure; P r1 , P r2 are the inlet and outlet pressures of the combustion chamber respectively; B1 and B2 are the inlet and outlet air guides of the combustion chamber respectively.

[0028] The energy balance equation is:

[0029]

[0030] Among them, G a is the air flow rate entering the combustion chamber; G f is the gas flow rate entering the combustion chamber; G g is the combustion chamber outlet gas flow rate; ρ rg h is the gas density at the combustion chamber outlet; r2 is the enthalpy of the combustion chamber outlet gas; T a1 is the combustion chamber inlet air temperature; c pa is the average specific heat of the combustion chamber air; H u is the lower calorific value of the fuel; r is the combustion chamber efficiency; V is the combustion chamber volume.

[0031] Heavy-duty gas turbines generally use axial-flow gas turbines with the characteristics of high temperature resistance, large flow, low specific enthalpy, and less technology. According to the mass equation, the exhaust pressure can be known as:

[0032]

[0033] Among them, P t2 is the turbine exhaust pressure; G t1 , G t2 are the turbine intake and exhaust flow rates, respectively; G e G is the flow loss; c is the air flow rate of the press; V t is the flow area of ​​the turbine; g is the compressor extraction cooling coefficient; X gt is the conversion coefficient of cooling air's working capacity; ξ is the flow fouling coefficient.

[0034] The turbine outlet temperature is:

[0035]

[0036] The work produced by the turbine is:

[0037] W t =G t1 (h t1 -h t2 )

[0038] The turbine efficiency is:

[0039] q=nG T0 / (n0G t1 )

[0040] Among them, h t1 ,h t2 are the gas enthalpy at the turbine inlet and outlet respectively; n is the unit speed; n0 is the rated speed of the unit; π t is the pressure ratio of each stage; G T0 is the rated flow rate in the turbine.

[0041] After selecting the physical information differential equation of heavy-duty gas turbine, the physical information neural network is trained. The differential equation guides the parameter update during network training to make it conform to the corresponding physical laws.

[0042] Consider a parameterized nonlinear differential equation of the general form:

[0043]

[0044] where u = u(t, x) represents the hidden solution, f(·) represents the nonlinear operator parameterized by λ, and B(·) and D(·) represent the boundary and initial conditions.

[0045] Construct a PRformer network to approximate u(t,x). The specific process is as follows:

[0046] In PRformer, based on the historical sequence The overall formal process of predicting future sequences is:

[0047]

[0048] in, represents C variables, each with D embedding dimensions.

[0049] The embedding layer obtained by pyramid RNN embedding is processed in the encoding layer of Transformer. In each encoding layer, a multi-head attention mechanism is used to learn the relationship between multiple variables and generate new representations, specifically:

[0050]

[0051] Finally, the output result is obtained by linear projection

[0052] Calculated by automatic differentiation For the derivatives of the time coordinate t and the space coordinate x, the implicit value of f(t,x) under a certain time and space coordinate is calculated.

[0053] In order to obtain the parameters of the deep neural network, the network parameters are updated by minimizing the mean square error to make it conform to the actual operating characteristics of the heavy-duty gas turbine. The specific formula of the mean square error is:

[0054] MSE = MSE u +MSE f

[0055]

[0056]

[0057] in, Represents measurable data or a priori conditions such as boundary conditions and initial conditions, Represents the selected training point of f(t,x).

[0058] Since from MSE f The prior information is absorbed in the model, and the amount of training data required to learn a certain dynamic behavior is significantly reduced. Under the assumption that the differential equation is represented by f(t,x), MSE f The term can be imposed on the loss function to constitute a constraint on the solution of the differential equation.

[0059] In order to solve the problem of limited accuracy of the output results of physical information neural networks when dealing with real-world problems, the gradient of the residual is embedded into the loss function as a new constraint, and the model training accuracy is improved through adaptive selection of training points.

[0060] During training, the residual f of the differential equation is forced to be 0, so the derivative of f will also be 0. Therefore, by enhancing the gradient residual:

[0061]

[0062] This gives us a new loss function:

[0063]

[0064] In order to balance the impact of different loss terms and ensure that the model achieves a good coordination between data fitting and physical constraints, weight coefficients are added to different loss terms to control the relative importance of each part of the loss in training. The loss function after adding the weight coefficient is:

[0065]

[0066] In order to set the optimal weight coefficient, the Ivy algorithm is used to automatically optimize the weight coefficient.

[0067] Population initialization:

[0068] I i =I min +rand(1,D)⊙(I max -I min ),i=1,…,Npop

[0069] Among them, rand(1,D) represents a random number of D dimension [0,1]; I max and I min Represents the upper and lower limits of the search space; ⊙ represents the Hadamard product.

[0070] Ivy is a creeping plant that grows over time. Assuming that the growth rate Gν of the ivy plant is a function of time, member I i The growth rate Gν i (t) is:

[0071] ΔGv i (t+1)=rand 2 ⊙(N(1,D)⊙ΔGv i (t))

[0072] Among them, ΔGv i (t) and ΔGv i(t+1) represents the growth rate of the discrete-time system at time t and t+1.

[0073] By simulating the growth behavior of ivy in nature, members I i Select the nearest and most important neighbor (judged by fitness value) I ii As a leader, represents a sequence of members sorted from best to worst, then:

[0074]

[0075] Member I i By Member I ii The growth behavior towards the light source is expressed as follows:

[0076]

[0077]

[0078] in, Represents Hadamard division.

[0079] In Member I i After searching for the nearest and most important neighbor, member I ii Will follow the best member of the entire population I Best , in member I Best Find a better solution nearby. The specific formula is:

[0080] I i new =I Best ⊙(rand(1,D)+N(1,D)⊙ΔGv i )

[0081] Subsequently, members The new growth rate Calculated by the following formula:

[0082]

[0083] The growth process of ivy plants includes two stages: rapid upward growth, new branch width, and flowering. These two stages are simulated mathematically, and the specific formula is:

[0084]

[0085] Among them, f(I i ) indicates member I i The fitness value is β=(2+rand) / 2.

[0086] During the algorithm operation, after each iteration, the previous population I is compared with the population generated by the current iteration. Merge Then sort the members according to their fitness values ​​from most favorable to least favorable, and we get:

[0087]

[0088] Then, the members with the top Npop ranking are selected as the current final population, i.e.

[0089] At this point, the initial population has continuously searched the space to determine the optimal member in the population, and output the optimal solution of the Ivy Algorithm when the maximum number of iterations is reached. If the maximum number of iterations is not reached, the number of iterations t is updated and the search is continued.

[0090] From the above description, it can be seen that this scheme proposes a heavy-duty gas turbine modeling method based on LVYA-PINN, which not only ensures that the model meets the laws of physics, but also enhances the interpretability and generalization ability of the model. The Ivy algorithm is used to optimize the weight coefficient of the loss term, effectively balancing different loss parts, so that the model can better fit the actual data while meeting the laws of physics. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 The present invention is a flow chart of an IVYA-PINN-based intelligent modeling method for heavy-duty gas turbine process characteristics.

[0092] Figure 2 This is the structure diagram of the physical information neural network.

[0093] Figure 3 This is a schematic diagram of the heavy-duty gas turbine structure. Specific implementation methods

[0094] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0095] In this example, refer to Figure 1 As shown, the present invention proposes a flow chart of a heavy-duty gas turbine process characteristic intelligent modeling method based on IVYA-PINN, comprising the steps of:

[0096] The operating characteristics of heavy-duty gas turbines were analyzed, the Spearman correlation coefficient was used to conduct correlation analysis on the input variables, and the input and output variables required to establish the heavy-duty gas turbine were selected.

[0097] The formula for calculating the Spearman correlation coefficient is:

[0098]

[0099] Among them, R(x i ) and R(y i ) are the values ​​of the two variables, and are the mean values ​​of variables x and y respectively.

[0100] The input variables are determined to be natural gas flow rate, IGV opening, compressor exhaust pressure, speed, load, and exhaust temperature, and the output variables are compressor exhaust pressure, speed, load, and exhaust temperature. The historical operating data of the corresponding variables are obtained and the data is preprocessed.

[0101] Data preprocessing includes data noise reduction, missing value filling, and standardization. The specific contents are as follows:

[0102] Data noise reduction: Use filtering algorithms to process random errors, outliers or irrelevant information introduced during data collection, transmission or processing to improve data quality;

[0103] Missing value filling: fill the missing values ​​in the collected data by using the average value of the sample points before and after the missing value sample point;

[0104] Standardization: The Z-score standardization method is used to standardize the data set. The specific formula is:

[0105]

[0106] Among them, x norm is the value after standardization, x is the value before standardization, and x mean is the mean of the original data, x std is the standard deviation of the original data.

[0107] A heavy-duty gas turbine model identification framework based on physical information neural network is established.

[0108] Furthermore, the specific process of establishing the model and training the model is as follows:

[0109] The model of a heavy-duty gas turbine is mainly composed of three modules: a compressor, a combustion chamber and a turbine. The compressor compresses the air sucked in from the outside step by step, stores a certain amount of energy, and transports it to the combustion chamber to mix and burn with the injected fuel, turning it into high-temperature and high-pressure combustion gas. The high-temperature combustion gas flows through the turbine to expand and do work, pushing the turbine to drive the compressor and the load rotor to rotate at high speed to generate electricity.

[0110] When the compressor has a certain structure size, its compression ratio and efficiency mainly depend on the speed, flow rate and intake conditions. azh , equivalent speed N yzh 、Pressure ratio π y and efficiency η y The functional relationship of these four similarity criteria is used to characterize the universal characteristic curve of the compressor. The specific formula is:

[0111]

[0112] A differential equation with pressure and enthalpy as state parameters is established to reflect the dynamic change process of the gas in the combustion chamber. The mass balance equation is:

[0113]

[0114] G a =B1(P r1 -P)

[0115] G g =B2(PP r2 )

[0116] Where, P is the combustion chamber pressure; P r1 , P r2 are the inlet and outlet pressures of the combustion chamber respectively; B1 and B2 are the inlet and outlet air guides of the combustion chamber respectively.

[0117] The energy balance equation is:

[0118]

[0119] Among them, G a is the air flow rate entering the combustion chamber; G f is the gas flow rate entering the combustion chamber; G g is the combustion chamber outlet gas flow rate; ρ rg h is the gas density at the combustion chamber outlet; r2 is the enthalpy of the combustion chamber outlet gas; T a1 is the combustion chamber inlet air temperature; c pa is the average specific heat of the combustion chamber air; H u is the lower calorific value of the fuel; r is the combustion chamber efficiency; V is the combustion chamber volume.

[0120] Heavy-duty gas turbines generally use axial-flow gas turbines with the characteristics of high temperature resistance, large flow, low specific enthalpy, and less technology. According to the mass equation, the exhaust pressure can be known as:

[0121]

[0122] Among them, P t2 is the turbine exhaust pressure; G t1 , G t2 are the turbine intake and exhaust flow rates, respectively; G e G is the flow loss; c is the air flow rate of the press; V t is the flow area of ​​the turbine; g is the compressor extraction cooling coefficient; X gt is the conversion coefficient of cooling air's working capacity; ξ is the flow fouling coefficient.

[0123] The turbine outlet temperature is:

[0124]

[0125] The work produced by the turbine is:

[0126] W t =G t1 (h t1 -h t2 )

[0127] The turbine efficiency is:

[0128]

[0129] Among them, h t1 ,h t2 are the gas enthalpy at the turbine inlet and outlet respectively; n is the unit speed; n0 is the rated speed of the unit; π t is the pressure ratio of each stage; G T0 is the rated flow rate in the turbine.

[0130] After selecting the physical information differential equation of heavy-duty gas turbine, the physical information neural network is trained. The differential equation guides the parameter update during network training to make it conform to the corresponding physical laws.

[0131] Consider a parameterized nonlinear differential equation of the general form:

[0132]

[0133] where u = u(t, x) represents the hidden solution, f(·) represents the nonlinear operator parameterized by λ, and B(·) and D(·) represent the boundary and initial conditions.

[0134] Construct a PRformer network to approximate u(t,x). The specific process is as follows:

[0135] In PRformer, based on the historical sequence The overall formal process of predicting future sequences is:

[0136]

[0137] in, represents C variables, each with D embedding dimensions.

[0138] The embedding layer obtained by pyramid RNN embedding is processed in the encoding layer of Transformer. In each encoding layer, a multi-head attention mechanism is used to learn the relationship between multiple variables and generate new representations, specifically:

[0139]

[0140] Finally, the output result is obtained by linear projection

[0141] Calculated by automatic differentiation For the derivatives of the time coordinate t and the space coordinate x, the implicit value of f(t,x) under a certain time and space coordinate is calculated.

[0142] In order to obtain the parameters of the deep neural network, the network parameters are updated by minimizing the mean square error to make it conform to the actual operating characteristics of the heavy-duty gas turbine. The specific formula of the mean square error is:

[0143] MSE = MSE u +MSE f

[0144]

[0145]

[0146] in, Represents measurable data or a priori conditions such as boundary conditions and initial conditions, Represents the selected training point of f(t,x).

[0147] Since from MSE f The prior information is absorbed in the model, and the amount of training data required to learn a certain dynamic behavior is significantly reduced. Under the assumption that the differential equation is represented by f(t,x), MSE f The term can be imposed on the loss function to constitute a constraint on the solution of the differential equation.

[0148] In order to solve the problem of limited accuracy of the output results of physical information neural networks when dealing with real-world problems, the gradient of the residual is embedded into the loss function as a new constraint, and the model training accuracy is improved through adaptive selection of training points.

[0149] During training, the residual f of the differential equation is forced to be 0, so the derivative of f will also be 0. Therefore, by enhancing the gradient residual:

[0150]

[0151] This gives us a new loss function:

[0152]

[0153] In order to balance the impact of different loss terms and ensure that the model achieves a good coordination between data fitting and physical constraints, weight coefficients are added to different loss terms to control the relative importance of each part of the loss in training. The loss function after adding the weight coefficient is:

[0154]

[0155] In order to set the optimal weight coefficient, the Ivy algorithm is used to automatically optimize the weight coefficient.

[0156] Population initialization:

[0157] I i =I min +rand(1,D)⊙(I max -I min ),i=1,…,Npop

[0158] Among them, rand(1,D) represents a random number of D dimension [0,1]; I max and I min Represents the upper and lower limits of the search space; ⊙ represents the Hadamard product.

[0159] Ivy is a creeping plant that grows over time. Assuming that the growth rate Gν of the ivy plant is a function of time, member I i The growth rate Gν i (t) is:

[0160] ΔGv i (t+1)=rand 2 ⊙(N(1,D)⊙ΔGv i (t))

[0161] Among them, ΔGv i (t) and ΔGv i(t+1) represents the growth rate of the discrete-time system at time t and t+1.

[0162] By simulating the growth behavior of ivy in nature, members I i Select the nearest and most important neighbor (judged by fitness value) I ii As a leader, represents a sequence of members sorted from best to worst, then:

[0163]

[0164] Member I i By Member I ii The growth behavior towards the light source is expressed as follows:

[0165]

[0166]

[0167] in, Represents Hadamard division.

[0168] In Member I i After searching for the nearest and most important neighbor, member I ii Will follow the best member of the entire population I Best , in member I Best Find a better solution nearby. The specific formula is:

[0169]

[0170] Subsequently, member I i new The new growth rate ΔGν i new Calculated by the following formula:

[0171]

[0172] The growth process of ivy plants includes two stages: rapid upward growth, new branch width, and flowering. These two stages are simulated mathematically, and the specific formula is:

[0173]

[0174] Among them, f(I i ) indicates member I i The fitness value is β=(2+rand) / 2.

[0175] During the algorithm operation, after each iteration, the previous population The group generated by the current iteration Merge Then sort the members according to their fitness values ​​from most favorable to least favorable, and we get:

[0176]

[0177] Then, the members with the top Npop ranking are selected as the current final population, i.e.

[0178] At this point, the initial population has continuously searched the space to determine the optimal member in the population, and output the optimal solution of the Ivy Algorithm when the maximum number of iterations is reached. If the maximum number of iterations is not reached, the number of iterations t is updated and the search is continued.

[0179] From the above description, it can be seen that this scheme proposes a heavy-duty gas turbine modeling method based on LVYA-PINN, which not only ensures that the model meets the laws of physics, but also enhances the interpretability and generalization ability of the model. The Ivy algorithm is used to optimize the weight coefficient of the loss term, effectively balancing different loss parts, so that the model can better fit the actual data while meeting the laws of physics.

[0180] The applicant of the present invention has described the specific implementation of the present invention in conjunction with the drawings of the specification. The technicians in the industry should understand that the above specific implementation and description are only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A heavy-duty gas turbine process characteristics intelligent modeling method based on IVYA-PINN, characterized in that: Follow the steps below to implement it: Step 1: Conduct mechanism analysis on heavy-duty gas turbines, analyze differential equations in various components, and provide physical constraints for the physical information neural network; Step 2: Use the Spearman correlation coefficient to perform correlation analysis on the input variables, select the input and output variables required for modeling, and collect the corresponding historical operation data; Step 3: Perform data noise reduction, missing value filling and standardization on the collected data; Step 4: Train the physical information neural network model, combine the data-driven error and the physical-driven error as the network's loss function, guide the network parameter update, and establish a heavy-duty gas turbine model; Step 5: Use the Ivy algorithm to automatically optimize the weight coefficients of each loss term in the loss function to improve the model's ability to integrate physical knowledge and actual data.

2. According to claim 1, a heavy-duty gas turbine process characteristic intelligent modeling method based on IVYA-PINN is characterized in that: The step 1 specifically involves performing a mechanism analysis on the heavy-duty gas turbine, analyzing the differential equations in each component, and providing physical constraints for the physical information neural network. The specific steps are as follows: Step 2.

1. Before starting data collection, analyze the physical differential equations satisfied in the compressor and establish the equivalent flow rate G azh , equivalent speed N yzh , pressure ratio π y and efficiency η y The functional relationship of these four similarity criteria is used to characterize the universal characteristic curve of the compressor. The specific formula is: Step 2.2: Establish a differential equation with pressure and enthalpy as state parameters to reflect the dynamic change process of the gas in the combustion chamber. The mass balance equation is: Step 2.3: The energy balance equation is: Step 2.4: Analyze the physical differential equations satisfied in the turbine. The exhaust pressure can be obtained based on the mass equation: Step 2.5, the outlet temperature of the turbine is: Step 2.6: The work produced by the turbine is: W t =G t1 (h t1 -h t2 ) Step 2.7, turbine efficiency is: q=nG T0 / (n0G t1 ) 3. The heavy-duty gas turbine process characteristic intelligent modeling method based on IVYA-PINN according to claim 1 is characterized in that: In step 2, the Spearman correlation coefficient is used to perform correlation analysis on the input variables, select the input and output variables required for modeling, and collect the corresponding historical operation data. The specific steps are as follows: Step 3.1: Use the Spearman correlation coefficient to perform correlation analysis on the input variables and select the input and output variables required to establish the heavy-duty gas turbine. The Spearman correlation coefficient calculation formula is: Step 3.2, select the input variables as natural gas flow rate, IGV opening, compressor exhaust pressure, speed, load, and exhaust temperature, and the output variables as compressor exhaust pressure, speed, load, and exhaust temperature; Step 3.3: Collect the historical operation data of the measuring point data determined in step 3.

2.

4. The heavy-duty gas turbine process characteristic intelligent modeling method based on IVYA-PINN according to claim 1 is characterized in that: The step 3 specifically involves performing data noise reduction, missing value filling, and standardization on the collected data. The specific steps are as follows: Step 4.1: Use continuous wavelet transform to reduce noise on the collected data. Continuous wavelet transform decomposes the signal by convolving it with the scale-shifted wavelet function. The specific formula is: Step 4.2: Fill the missing values ​​in the collected data by taking the average value of the sample points before and after the missing value sample point; Step 4.3: Use the Z-score standardization method to standardize the collected historical operation data. The Z-score standardization calculation formula is:

5. The heavy-duty gas turbine process characteristic intelligent modeling method based on IVYA-PINN according to claim 1, characterized in that: The step 4 specifically trains the physical information neural network model, combines the data-driven error and the physical-driven error as the network's loss function, guides the network parameter update, and establishes a heavy-duty gas turbine model. The specific steps are as follows: Step 5.1: Construct a PRformer network to learn the relationship between input and output variables and approximate the solution of the physical differential equation; Step 5.2, design the loss function, including the error between the neural network prediction value and the actual collected data and the residual obtained by automatically differentiating the prediction value and inserting it into the physical equation. The specific formula is: MSE=MSE u +MSE f Step 5.3: Embed the gradient of the residual as a new constraint into the loss function to obtain a new loss function:

6. The heavy-duty gas turbine process characteristic intelligent modeling method based on IVYA-PINN according to claim 1, characterized in that: In step 5, the Ivy algorithm is specifically used to automatically optimize the weight coefficients of each loss term in the loss function to improve the model's ability to integrate physical knowledge and actual data. The specific steps are as follows: Step 6.1, population initialization, I i =I min +rand(1,D)⊙(I max -I min ),i=1,…,Npop Step 6.2: By simulating the growth behavior of ivy in nature, member I i Select the nearest and most important neighbor (judged by fitness value) I ii As a leader, member I i By Member I ii The growth behavior towards the light source is expressed as follows: Step 6.3: In member I i After searching for the nearest and most important neighbor, member I ii Will follow the best member of the entire population I Best , in member I Best Find a better solution nearby. The specific formula is: Step 6.4, the growth process of ivy plants includes two stages, the rapid upward growth stage, the new branch width and the flowering stage. These two stages are simulated mathematically, and the specific formula is: Step 6.5: During the algorithm operation, after each iteration, the previous population The group generated by the current iteration Merge Then sort the members according to their fitness values ​​from most favorable to least favorable, and we get: Step 6.6: Select the top Npop members as the current final population, i.e. Step 6.7: Determine whether the maximum number of iterations has been reached or the optimal solution has been found. If the optimal solution has been found, output the optimal member position. If the above conditions are not met, return to step 6.2 until the optimal solution is found.

Images