Marine gas turbine compressor performance modeling method based on double agent model

Abstract

The purpose of this invention is to provide a compressor performance modeling method for marine gas turbines based on a dual-surrogate model, belonging to the field of gas turbines. The method includes the following steps: constructing a one-dimensional program-driven compressor performance surrogate model, including sample generation, data preprocessing, and training a BP neural network; constructing a one-dimensional experimental relative error surrogate model driven by experimental data, including experimental data acquisition and preprocessing, and constructing the relative error surrogate model; and combining and validating the dual surrogate models. This invention utilizes limited experimental data to construct a dynamic error correction model, improving the prediction accuracy under all operating conditions; overcomes the bottleneck of model generalization caused by the scarcity of experimental data, expands the model's applicability by relying on the full operating condition coverage capability of one-dimensional data; and ensures a millisecond-level response speed for the model, meeting the engineering requirements of real-time control and rapid iterative design of multiple schemes for gas turbines.

Description

Technical Field

[0001] The present invention relates to a gas turbine simulation method, specifically a compressor simulation method. Background Technology

[0002] As a core component of the aerodynamic system of marine gas turbines, the performance model of the compressor directly impacts the feasibility of the power system design and the effectiveness of operational optimization due to the breadth of operating condition coverage and the accuracy of prediction. Currently, compressor modeling technology faces three major bottlenecks: One-dimensional calculation methods (such as the improved AXIAL algorithm and the classic HARIKA model) rely on the average streamline theory and have the ability to quickly generate data for all operating conditions. However, they have limitations in characterizing complex flow phenomena. Efficiency losses caused by tip clearance leakage, interference of film cooling airflow on the main flow field, and the evolution of secondary flow under off-design conditions are all difficult to model accurately, resulting in deviations in off-design efficiency and flow prediction. In addition, their empirical correction coefficients are mostly calibrated based on test data of traditional models, and their adaptability to new variable geometry compressors such as wide adjustable guide vanes and staged bleed air is significantly reduced, with errors increasing under extreme low speed or high load conditions. Experimental data is the core basis for verifying the accuracy of performance models. However, tests on marine gas turbine compressors rely on large-scale high-altitude simulation test benches or sea trial platforms, resulting in high costs and long cycles for each test, and making it difficult to obtain full-condition sample data. Existing experimental data is mostly concentrated at design points and common cruise conditions, with scarce data on marginal conditions such as low speed with high bleed air and high speed with low load. Furthermore, environmental temperature and humidity fluctuations and sensor measurement deviations introduce random noise during testing, which can easily lead to overfitting when directly used for model training. Existing data fusion techniques have significant drawbacks: some solutions directly use limited experimental data (usually only a few hundred sets) to train surrogate models, leading to a sharp decline in generalization ability under untested conditions due to insufficient sample size; some solutions simply concatenate one-dimensional data with experimental data for training, without considering the differences in error characteristics between the two types of data (one-dimensional data exhibits systematic bias, while experimental data contains random noise), resulting in a fused model with lower accuracy than a model based solely on experimental data; other solutions adjust the one-dimensional model results using static correction coefficients, but these coefficients cannot dynamically change with operating conditions, rendering the correction ineffective across different operating scenarios. In summary, existing technologies cannot simultaneously meet the three major engineering requirements of "full-condition coverage", "high-precision prediction" and "low-cost implementation". There is an urgent need for a modeling method that can efficiently integrate the advantages of one-dimensional data for full-condition coverage with the high-precision characteristics of experimental data. Summary of the Invention

[0003] The purpose of this invention is to provide a compressor performance modeling method for marine gas turbines based on a dual-surrogate model that achieves full operating condition coverage and high-precision prediction coordination.

[0004] The objective of this invention is achieved as follows: The present invention relates to a compressor performance modeling method for marine gas turbines based on a dual-surrogate model, characterized by the following steps: (1) Construct a one-dimensional program-driven compressor performance proxy model, including sample generation, data preprocessing and training of BP neural network; (2) Construct a one-dimensional experimental relative error surrogate model driven by experimental data, including experimental data acquisition and preprocessing, and constructing the relative error surrogate model; (3) Combining and verifying the dual-agent model.

[0005] The present invention may also include: 1. The sample generation in step (1) includes: using a one-dimensional compressor aerodynamic performance calculation program for marine gas turbines, with input parameters being relative speed N, rotatable guide vane angle IGV, and pressure ratio. , air intake Q; output parameter is isentropic efficiency With the flow rate G, sample data is generated across the entire operating range, covering typical scenarios at the design point.

[0006] 2. The data preprocessing in step (1) includes: using the Z-Score standard normalization method to map the input and output data to the [-1,1] interval; removing abnormal data in one-dimensional calculation; and dividing the training set and test set in a 7:3 ratio.

[0007] 3. The training of the BP neural network in step (1) includes: the input layer has 4 nodes, corresponding to N, IGV, and The hidden layer uses a two-layer structure, with 32 neurons in the first layer and 16 neurons in the second layer; the output layer has 2 nodes, corresponding to... The hidden layer uses the LeakyReLU function; the output layer uses a linear activation function to ensure that the prediction results conform to the range of physical quantities; and the RMSprop optimizer is used to dynamically adjust the learning rate of the corresponding parameters.

[0008] 4. The test data acquisition and preprocessing in step (2) includes: conducting tests on the marine gas turbine compressor test bench, collecting test data, and test parameters including compressor speed N, rotatable guide vane angle IGV, and total pressure ratio. The amount of air drawn up, Q, and the corresponding isentropic efficiency. And flow rate G; and calculate the relative error between the one-dimensional calculation program and the experimental data.

[0009] 5. The construction of the relative error proxy model in step (2) includes: the network input is consistent with the one-dimensional performance proxy model, with 4 nodes, corresponding to N, IGV, and Q; The network output has two nodes, corresponding to the relative efficiency error respectively. Flow rate relative error Network structure: 6 nodes in the input layer, two-layer hidden layer, and 2 nodes in the output layer; the Tanh function is used for the hidden layer activation function; the linear activation function is used for the output layer; the Adam optimizer is used.

[0010] 6. Step (3) is as follows: Combination mechanism: The prediction results of the one-dimensional performance surrogate model are fused with the output of the error-corrected BP surrogate model to obtain the final performance prediction value; Correction efficiency: Correcting flow: ; Model integration and verification: The model reduction toolbox of MWORKS.Sysplorer software was used to build a surrogate model, generating a compressor performance model with dual surrogate models, which was then integrated into the overall performance model of the gas turbine for testing and verification.

[0011] The advantages of this invention are: 1. The compressor performance modeling method based on data fusion and transfer learning proposed in this invention provides an efficient and reliable approach for the digital design, real-time control, and performance optimization of marine gas turbine compressors. The surrogate model constructed by this invention has millisecond-level predictive response capabilities, significantly surpassing traditional CFD simulations (which take hours to days), and meeting the urgent needs of high-time-sensitivity scenarios such as online real-time control of gas turbines and rapid iterative selection of multiple schemes.

[0012] 2. This invention creatively combines the efficiency of one-dimensional programming, the high precision of CFD simulation, and the powerful nonlinear mapping capabilities of data-driven models. Its core innovations lie in its lightweight network architecture and efficient training, a dynamic correction mechanism based on transfer learning, and the fusion of heterogeneous data. While ensuring model expressiveness and engineering interpretability, it achieves a lightweight model design, accurately corrects the systematic prediction bias of one-dimensional programming under complex flow conditions, significantly improves the prediction accuracy of the model under key conditions, and effectively integrates multi-source data with different accuracy characteristics and distribution patterns.

[0013] 3. This method can reduce the reliance on expensive experiments and time-consuming CFD simulations for building high-precision compressor models, shorten the R&D cycle, and reduce testing costs. The high-precision real-time model it constructs lays a solid technical foundation for the digital forward design, intelligent operation and maintenance, online performance optimization, and virtual-real combined control (such as digital twins) of gas turbines, and powerfully promotes the upgrading of ship power systems towards intelligence and efficiency. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram of the BP neural network structure. Figure 3 This is a schematic diagram of the neural network training results. Detailed Implementation

[0015] The invention will now be described in more detail with reference to the accompanying drawings: Combination Figure 1-3 The specific process of this invention is as follows: (I) Construction of a one-dimensional program-driven compressor performance proxy model (1) Sample generation: A one-dimensional compressor aerodynamic performance calculation program for marine gas turbines was used. The input parameters were relative speed (N, speed range 0.4~1.1), rotatable guide vane angle (IGV, range -40°~3°, the angle is linear with speed between 0.8~0.9 speeds, while the upper and lower limits of the angle remain unchanged for speeds less than 0.8 and 0.9), and pressure ratio ( ), and induced draft (Q, ranging from 0 to 3.5% of total flow); the output parameter is isentropic efficiency ( ( ) and flow rate (G). 42,878 sets of sample data were generated across the entire operating range, covering typical scenarios such as design points.

[0016] (2) Data preprocessing: The Z-Score standard normalization method is used to map the input and output data to the interval [-1,1]; abnormal data in one-dimensional calculation (such as invalid samples with negative flow and efficiency exceeding 100%) are removed; the training set and test set are divided in a 7:3 ratio to ensure the consistency of sample distribution.

[0017] (3) BP neural network training: The input layer has 4 nodes, corresponding to N, IGV, and The hidden layer uses a two-layer structure, with 32 neurons in the first layer and 16 neurons in the second layer; the output layer has 2 nodes, corresponding to... The hidden layer uses the LeakyReLU function (slope coefficient set to 0.01) to solve the neuron "death" problem of the traditional ReLU function; the output layer uses a linear activation function to ensure that the prediction results conform to the range of physical quantities; the RMSprop optimizer is used to dynamically adjust the learning rate of the corresponding parameters to improve training stability.

[0018] (II) Construction of a one-dimensional experimental relative error surrogate model driven by experimental data (1) Data Acquisition and Preprocessing: Tests were conducted on a marine gas turbine compressor test bench to collect key operating condition data. Test parameters included compressor speed (N, speed range 0.5~1.05), rotatable guide vane angle (IGV, range -30°~0°), and total pressure ratio (N). ), induced draft (Q, ranging from 0 to 3.5% of total flow), and the corresponding isentropic efficiency ( ) and flow rate (G); and calculate the relative error between the one-dimensional calculation program and the experimental data; (2) Construction of the relative error surrogate model: The network input is consistent with the one-dimensional performance surrogate model, with 4 nodes, corresponding to N, IGV, and Q; The network output has two nodes, corresponding to the relative efficiency error respectively. Flow rate relative error Network structure: 6 nodes in the input layer, a two-layer hidden layer (24 neurons in the first layer and 12 neurons in the second layer), and 2 nodes in the output layer; the Tanh function is used for the hidden layer activation function to improve the nonlinear mapping capability; the linear activation function is used for the output layer to ensure the continuity of the error prediction value; the Adam optimizer is used to balance convergence speed and update stability.

[0019] (III) Combination and Validation of the Dual-Agent Model (1) Combination mechanism: The prediction results of the one-dimensional performance surrogate model are fused with the output of the error-corrected BP surrogate model to obtain the final performance prediction value; correction efficiency: Correcting flow: ; (2) Model integration verification: Optionally, the model reduction toolbox of MWORKS.Sysplorer software is used to construct a proxy model, generate a compressor performance modelica with dual proxy models, and integrate it into the overall performance model of the gas turbine for testing and verification.

[0020] The compressor characteristics are calculated using the one-dimensional HARIKA algorithm, which is based on the stage superposition method. The algorithm performs empirical corrections along the blade height and also performs empirical corrections on the characteristic parameters of each blade profile, thus enabling accurate calculation of the parameters of the outlet blade throat.

[0021] The one-dimensional calculation method for compressor performance is as follows: Based on the known compressor geometry, inlet total temperature, total pressure, and circumferential velocity, the velocity triangle on the average radius of the rotor inlet cross-section is solved. The total pressure ratio and adiabatic efficiency of this stage are calculated empirically. Using the same method, the velocity triangle at the stage outlet can be further obtained. This outlet parameter is used as the inlet parameter for the next stage, and the outlet parameter at that flow point in the next stage can be calculated. This process is repeated until the outlet parameter of the last stage of the compressor is obtained. Then, the flow rate and rotational speed are changed, and the above process is repeated until the performance parameters of the entire compressor are obtained.

[0022] The main theoretical formulas used in the HARIKA algorithm are as follows, for flow calculation: in The traffic reserve coefficient is an adjustable data input by the user.

[0023] Efficiency calculation: Stage pressure ratio calculation: Using these calculation formulas, combined with other aerodynamic formulas and empirical corrections, the characteristic curve of the compressor was calculated, and the performance prediction of the compressor was completed.

[0024] In this invention, both the dual-surrogate model (one-dimensional performance surrogate model and one-dimensional-experimental relative error surrogate model) are constructed based on an improved backpropagation (BP) neural network. The core principle is to achieve accurate fitting from the compressor input operating parameters to the output performance parameters (or relative error) through nonlinear mapping of multiple layers of neurons. Its key characteristics and engineering adaptability are as follows: Taking a one-dimensional performance proxy model as an example, the forward propagation formula under a certain operating condition is: Where W1 is the input layer - first hidden layer weight, f is the Leaky ReLU activation function, W2 is the first layer - second hidden layer weight, and W3 is the second hidden layer - output layer weight.

[0025] During forward propagation, the input layer receives compressor operating parameters, which are then processed by the activation function of the hidden layer neurons and transmitted to the output layer to output the performance prediction value. During backpropagation, the deviation between the prediction value and the true value is calculated through the loss function, and the deviation is backpropagated using the gradient descent method to iteratively update the weights and biases of each layer and minimize the loss function.

[0026] The Leaky ReLU activation function used in constructing neural networks is an improvement on the traditional ReLU function, solving the problem of neurons "dying" (weights never being updated) in the negative gradient region. Its expression is: This structure can adapt to the gradient characteristics of low-speed operation. At low speed, the gradients of compressor efficiency and flow rate with respect to operating parameters are prone to being negative (such as the flow rate decreasing due to an increase in bleed air volume). Traditional ReLU will set the negative gradient to zero, resulting in the corresponding weights not being able to be updated. The negative slope of Leaky ReLU can preserve small gradients, ensuring that the model can still learn effectively under low-speed operation. The Tanh function used in constructing the experimental error surrogate model is a commonly used nonlinear activation function, and its expression is: This structure has strong nonlinear mapping characteristics. Compared with Leaky ReLU, Tanh has a steeper gradient in the middle region (x∈[-2,2]), making it more sensitive to changes in the input signal and suitable for fitting nonlinear relationships with small errors. Tanh's zero-mean output can reduce the impact of weight initialization bias on training, avoid model shift due to small sample size, and ensure that the test set error is stable and meets the standard.

[0027] The RMSprop optimizer for training one-dimensional performance surrogate models is an adaptive learning rate optimizer based on gradient squared accumulation. Its core principle is to maintain a "gravity squared moving average" for each parameter, dynamically adjusting the learning rate. Example: Constructing a one-dimensional program-driven proxy model: Step 1: Sample generation and data preprocessing. A program for calculating the aerodynamic performance of a one-dimensional compressor in a marine gas turbine, based on the HARIKA algorithm (the core algorithm is a stacking method, including empirical corrections for blade height and blade profile characteristic parameters, which can quickly output compressor performance parameters), is used. Input parameters include relative speed (N, range 0.4~1.1), steerable guide vane angle (IGV, range -40°~3°, linearly varying with speed from 0.8 to 0.9, fixed at -40° for speeds less than 0.8, and fixed at 3° for speeds greater than 0.9), pressure ratio (πc), and bleed air volume (Q, range 0~3.5% of total flow). Output parameters are isentropic efficiency (η) and flow rate (G). 42,878 sets of samples were generated within the above operating conditions, covering typical scenarios such as design points. 80% (34,302 sets) were used as the training set, and 20% (8,576 sets) as the test set. Data preprocessing was performed on the generated samples: Z-Score normalization was used to map the input and output parameters to the [-1,1] interval (formula: (μ is the mean of the parameters, σ is the standard deviation); outlier samples with flow rate less than 0, efficiency greater than 1.0 or less than 0.6 were removed. Step Two: Neural Network Construction and Training. Since the output of the one-dimensional performance model is steady-state data, an improved BP neural network is selected for training. The network structure is determined as follows: The input layer has 4 nodes, corresponding to N, IGV, πc, and Q respectively; the hidden layer has a two-layer structure to balance model fitting ability and training efficiency, and both layers use the Leaky ReLU activation function (…). To address the "death" problem of neurons; the output layer has two nodes, corresponding to η and G respectively, and uses a linear activation function to ensure the rationality of physical quantities. The optimizer is RMSprop, and the loss function is the root mean square error (RMSE). (where m is the batch sample size). Step 3: Model Parameter Optimization. By analyzing and comparing the model's predicted values ​​with the one-dimensional program output values ​​under different sample inputs, the training effect is measured and parameter optimization is guided. Parameters to be optimized include: initial learning rate, number of hidden layer neurons, number of training iterations, and training batch size. When the average prediction error of the test set is ≤3% and the error trends of the training set and the test set are consistent, the one-dimensional program-driven surrogate model is considered to have met the standard, and the single-condition prediction response time is ≤0.1ms. Construction of a one-dimensional experimental relative error surrogate model driven by experimental data: Step 1: Data Acquisition and Error Calculation. Tests were conducted on a marine gas turbine compressor test bench, collecting 826 sets of key operating condition data. Input parameters included relative speed (N, range 0.5~1.05), rotatable guide vane angle (IGV, range -30°~0°), pressure ratio (πc), and bleed air volume (Q, range 0~3.5% of total flow). Output parameter was the measured isentropic efficiency (…). ) and flow ( ). Call the one-dimensional program, input the test condition parameters, and obtain the one-dimensional predicted value. , ), calculate relative error: efficiency error Flow error .

[0028] Step Two: Error Surrogate Model Construction and Training. A backpropagation (BP) neural network is selected to construct the error surrogate model. The network structure is consistent with the input of the one-dimensional performance model: the input layer has 4 nodes (N, IGV, πc, Q) to ensure matching of operating parameters; the hidden layer has a two-layer structure and uses the Tanh activation function (…). (to improve the nonlinear mapping capability for small errors); the output layer has 2 nodes, corresponding to eη, respectively. A linear activation function is used to ensure the continuity of the error.

[0029] The optimizer chosen was Adam, the loss function was root mean square error, the training batch size was 2, the number of iterations was 2000, and the hyperparameters were continuously optimized. Dual-proxy model combination and testing verification: Step 1: Implementation of the dual-surrogate model combination mechanism. The one-dimensional performance surrogate model and the error surrogate model are merged, and the combination formula is: Corrected efficiency. Corrected flow rate ( , Output for a one-dimensional proxy model. , (Output for the error surrogate model). Step 2: Model Integration and Simulation Setup. The surrogate model is constructed using the "Model Reduction Toolbox" in MWORKS.Sysplorer software. The simulation model generation and export functions are then used to generate a dual surrogate model module in Modelica format: containing a one-dimensional performance calculation submodule, an error calculation submodule, and a combined calculation submodule. A dual-agent model for the low-pressure compressor and a dual-agent model for the high-pressure compressor were constructed respectively. The simulation conditions were set as follows: the transient process from idle speed (N=0.45) to design speed (N=1.0) (duration 100s, step size 0.01s). The IGV was adjusted according to the law in Example 1, and the bleed air volume was dynamically adjusted with the load. Step 3: Testing, Verification, and Performance Comparison. The experimental data was compared with the simulation results of the integrated model: the average error in efficiency under all operating conditions was 0.48%, and the average error in flow rate was 0.79%, meeting engineering accuracy requirements; the total dynamic simulation time was 8 seconds, much faster than traditional CFD; compared to traditional modeling methods, this invention requires less experimental data and shortens the modeling cycle. This invention aims to address the key bottlenecks of traditional technologies, such as difficulty in achieving full-condition coverage, high testing costs, and slow response. By innovatively constructing a "one-dimensional performance proxy model + error correction proxy model" architecture, it achieves full-condition coverage through a one-dimensional program while dynamically correcting errors using limited experimental data. Furthermore, it employs a lightweight BP neural network and an efficient optimizer to ensure the model's real-time performance and generalization capabilities. Compared with existing technologies, the innovations of this invention are: 1) The input parameters of the error correction model and the one-dimensional performance model are unified, avoiding the deviation of the working condition matching; 2) The dynamic weight adjustment mechanism improves the reliability of prediction of untested working conditions; 3) It does not rely on time-consuming CFD, and the engineering implementation cost is low. This method provides core support for the digital design, intelligent operation and maintenance, and online control of marine gas turbines, and has significant strategic value in promoting the intelligent upgrading of marine power systems, shortening the R&D cycle, and reducing testing costs. The widespread application of this invention will effectively promote the digital transformation of the shipbuilding industry and provide technical support for the development of green and low-carbon ships.

Claims

1. A method for modeling the performance of a compressor of a marine gas turbine based on a dual agent model, characterized in that: Includes the following steps: (1) Construct a one-dimensional program-driven compressor performance proxy model, including sample generation, data preprocessing and training of BP neural network; (2) Construct a one-dimensional experimental relative error surrogate model driven by experimental data, including experimental data acquisition and preprocessing, and constructing the relative error surrogate model; (3) Combining and verifying the dual-agent model.

2. The compressor performance modeling method for marine gas turbines based on a dual-surrogate model according to claim 1, characterized in that: The sample generation in step (1) comprises: using a one-dimensional compressor aerodynamic performance calculation program of a marine gas turbine, input parameters are relative rotating speed N, rotatable guide vane angle IGV, pressure ratio , and air bleed amount Q; output parameters are isentropic efficiency and flow G, sample data is generated in a full working condition interval, covering typical scenes of design points.

3. The compressor performance modeling method for marine gas turbines based on a dual-surrogate model according to claim 1, characterized in that: The data preprocessing in step (1) includes: using the Z-Score standard normalization method to map the input and output data to the [-1,1] interval; removing outlier data in one-dimensional calculations; and dividing the training set and test set into a 7:3 ratio.

4. The compressor performance modeling method for marine gas turbines based on a dual-surrogate model according to claim 1, characterized in that: Step (1) in training the BP neural network includes: the input layer has 4 nodes, corresponding to N, IGV, and The hidden layer uses a two-layer structure, with 32 neurons in the first layer and 16 neurons in the second layer; the output layer has 2 nodes, corresponding to... The hidden layer uses the LeakyReLU function; the output layer uses a linear activation function to ensure that the prediction results conform to the range of physical quantities; and the RMSprop optimizer is used to dynamically adjust the learning rate of the corresponding parameters.

5. The compressor performance modeling method for marine gas turbines based on a dual-surrogate model according to claim 1, characterized in that: Step (2) includes data acquisition and preprocessing: conducting tests on a marine gas turbine compressor test bench, collecting test data, and including test parameters such as compressor speed N, rotatable guide vane angle IGV, and total pressure ratio. The amount of air drawn up, Q, and the corresponding isentropic efficiency. And flow rate G; and calculate the relative error between the one-dimensional calculation program and the experimental data.

6. The compressor performance modeling method for marine gas turbines based on a dual-surrogate model according to claim 1, characterized in that: Step (2) involves constructing a relative error proxy model, which includes: the network input is consistent with the one-dimensional performance proxy model, with 4 nodes corresponding to N, IGV, and 1 / 2 GHz respectively. Q; The network output has two nodes, corresponding to the relative efficiency error respectively. Flow rate relative error Network structure: 6 nodes in the input layer, two-layer hidden layer, and 2 nodes in the output layer; the Tanh function is used for the hidden layer activation function; the linear activation function is used for the output layer; the Adam optimizer is used.

7. The compressor performance modeling method for marine gas turbines based on a dual-surrogate model according to claim 1, characterized in that: Step (3) is as follows: Combination mechanism: The prediction results of the one-dimensional performance surrogate model are fused with the output of the error-corrected BP surrogate model to obtain the final performance prediction value; Correction efficiency: Correcting flow: ; Model integration and verification: The model reduction toolbox of MWORKS.Sysplorer software was used to build a surrogate model, generating a compressor performance model with dual surrogate models, which was then integrated into the overall performance model of the gas turbine for testing and verification.

Images