A ship gas turbine compressor performance modeling method based on transfer learning

Abstract

The purpose of this invention is to provide a method for modeling the performance of marine gas turbine compressors based on transfer learning, belonging to the field of gas turbines. The method includes the following steps: constructing an initial sample library; performing data preprocessing; selecting a BP neural network for sample training; optimizing the model structure and parameters to complete the verification of a one-dimensional surrogate model of the compressor; conducting CFD simulation data-driven transfer learning optimization: for the required operating points, conducting three-dimensional CFD simulations to obtain compressor efficiency, compressor flow rate data, and flow field distribution; freezing the weights of the first two layers of the surrogate model, fine-tuning the output layer, and using CFD data to locally optimize the model, enabling the model to learn the shared features of the one-dimensional data and CFD data; and integrating the transferred-learned model into the overall performance model of the gas turbine. This invention significantly improves the prediction accuracy of the surrogate model under key operating conditions using a small amount of CFD simulation data, achieving real-time simulation while maintaining engineering interpretability, thus laying the foundation for virtual-real control of 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 aerodynamic component of a gas turbine, the accuracy of the compressor's performance prediction model directly determines the feasibility and optimization potential of the overall engine design. Currently, compressor modeling technology mainly relies on three types of methods, but all of them have significant technical bottlenecks: One-dimensional calculation methods, based on mean streamline theory, rapidly predict compressor characteristics using empirical formulas and simplified flow assumptions. However, they lack effective modeling of three-dimensional effects such as end-region secondary flow and shock wave / boundary layer interference, leading to significant errors in flow-efficiency predictions under off-design conditions. Furthermore, the calibration of empirical coefficients heavily relies on test data from specific engine models, resulting in a sharp decline in applicability in new high-load, wide-chord blade designs. Under extreme conditions, one-dimensional models fail to capture phenomena such as flow separation, leading to large discrepancies between predictions and experimental results.

[0003] While CFD simulations can capture the details of three-dimensional flow, their high computational cost makes them unsuitable for rapid iteration. For example, a single-condition steady-state simulation of a 10-stage high-pressure compressor requires tens of thousands of CPU hours; constructing a complete characteristic curve would be prohibitively expensive in terms of computational resources. Furthermore, the accuracy of CFD models is significantly affected by factors such as the choice of turbulence model and the number of meshes, especially in shock wave / boundary layer interference regions where pressure fluctuation prediction errors can lead to increased efficiency deviations.

[0004] In recent years, surrogate models based on machine learning (such as support vector regression and Gaussian processes) and deep learning (such as deep neural networks) have been used to replace high-cost simulations. However, existing methods generally face two major challenges: First, they are highly dependent on data. Model training requires a large amount of high-precision data, but experimental data is expensive to acquire, and simulation data has inherent biases. Models trained from a single data source show a sharp drop in generalization ability when operating under different conditions or across different aircraft models. Second, it is difficult to fuse heterogeneous data. One-dimensional mechanistic models, CFD simulation data, and experimental data have different accuracy characteristics and distribution patterns, making effective fusion difficult. For example, one-dimensional mechanistic models have low errors under medium load conditions, but the errors increase significantly in high load areas; CFD simulation data has high reliability under steady-state conditions, but the high computational cost makes it difficult to obtain sufficient computational samples. Existing methods (such as simple data concatenation and weighted averaging) cannot effectively utilize the complementarity of heterogeneous data, resulting in the fused model having lower accuracy than the model from a single high-precision data source. Summary of the Invention

[0005] The purpose of this invention is to provide a method for modeling the performance of marine gas turbine compressors based on transfer learning to achieve heterogeneous data fusion.

[0006] The objective of this invention is achieved as follows: This invention discloses a method for modeling the performance of marine gas turbine compressors based on transfer learning, characterized by the following steps: (1) Based on a one-dimensional compressor mechanism model, input-output data are generated within the target operating condition range to construct an initial sample library; wherein, the model input includes compressor speed, rotatable guide vane angle, pressure ratio, and bleed air volume, and the model output includes compressor efficiency and compressor flow rate; (2) Perform data preprocessing, normalize the input and output data, use filtering methods to eliminate noise, and divide the training set and test set in an 8:2 ratio; (3) A BP neural network was selected for sample training. The four nodes of the input layer correspond to the compressor speed, the angle of the rotatable guide vane, the pressure ratio, and the bleed air volume. The two nodes of the output layer correspond to the compressor efficiency and the compressor flow rate. The hidden layer is a two-layer network with ReLU activation function and mean square error (MSE) loss function. L1 regularization and Dropout structure optimization were added to the network to prevent overfitting. A variable learning rate strategy was adopted to achieve higher model accuracy. (4) Optimize the model structure and parameters to complete the verification of the one-dimensional surrogate model of the compressor, compare the prediction errors of the surrogate model and the one-dimensional program at unseen operating points, and ensure that the error of the surrogate model is within an acceptable range; carry out CFD simulation data-driven transfer learning optimization: (5) Conduct three-dimensional CFD simulations for the required operating points to obtain compressor efficiency, compressor flow data and flow field distribution; (6) Freeze the weights of the first two layers of the proxy model, fine-tune the output layer, and use CFD data to perform local optimization of the model so that the model learns the shared features of one-dimensional data and CFD data. (7) Transfer the learned model to the overall performance model of the gas turbine and test and verify it.

[0007] The present invention may also include: 1. In step (1), the compressor characteristics calculation of the compressor mechanism model adopts the one-dimensional HARIKA algorithm. The HARIKA algorithm is based on the stage superposition method, and performs empirical corrections along the blade height and also performs empirical corrections on the characteristic parameters of the blade profile, thereby calculating the parameters of the outlet blade throat: The calculation method for the one-dimensional mechanistic model of compressor performance is as follows: Based on the known compressor geometry, inlet total temperature, total pressure, and circumferential velocity parameters, solve for the velocity triangle on the average radius of the rotor inlet section; use empirical methods to solve for the total pressure ratio and adiabatic efficiency of this stage; use the same method to further obtain the velocity triangle at the stage outlet, and use this outlet parameter as the inlet parameter of the next stage to calculate the outlet parameter of the next stage at this flow rate point; repeat this process until the outlet parameter of the last stage of the compressor is obtained, then change the flow rate and speed, and repeat the above process until the performance parameters of the entire compressor are obtained; The theoretical formula used in the HARIKA algorithm for flow calculation is as follows: Where G is the compressor mass flow rate. This is the total pressure at the compressor inlet. This refers to the total temperature at the compressor inlet. The traffic reserve coefficient is an adjustable data input by the user. Efficiency calculation: Stage pressure ratio calculation: By combining aerodynamic formulas and empirical corrections, the characteristic curve of the compressor is calculated, and the compressor performance is predicted.

[0008] 2. The neural network adopts an improved BP neural network architecture, which realizes the nonlinear mapping from input parameters to output parameters through the error backpropagation algorithm. The network structure is defined as follows: Input layer input vector Where n is the rotational speed, For the guide vane angle, For pressure ratio, To determine the air intake, the number of input layer nodes is 4, and the data is normalized to the range [-1,1]. The hidden layers employ a two-layer fully connected structure, with the number of neurons in each layer dynamically optimized through sensitivity analysis. The activation function for the hidden layers is the modified linear unit ReLU. in This is the weight matrix. For bias vectors, For output from the previous layer; Output layer For efficiency and flow rate, linear activation functions are used respectively: The regularization strategy chosen is L1 regularization, which adds the sum of the absolute values ​​of the weights to the loss function to suppress overfitting during network training. In the formula, λ is the penalty coefficient, which is optimized through cross-validation; During training, the Dropout mechanism randomly discards hidden layer neurons with a probability of p=0.5. During testing, all neurons are enabled and their weights are scaled.

[0009] 3. In step (6), the transfer learning freezes the weights of the first two layers, retains the general flow characteristics of one-dimensional data training, fine-tunes the output layer, uses CFD data to adjust the weights of the output layer, and corrects the systematic deviation of the one-dimensional mechanism model under complex flow conditions. Let the parameters of the neural network be... W1 is the weight matrix from the input layer to the first hidden layer, W2 is the weight matrix from the first hidden layer to the second hidden layer, and W3 is the weight matrix from the second hidden layer to the output layer. During transfer learning, W1 and W2 are frozen, and only W3 is optimized. The loss function is then modified as follows: The hyperparameters are determined through cross-validation, and the dynamic fusion weights are defined as follows: The final prediction output is a weighted fusion: .

[0010] The advantages of this invention are: 1. 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 using this invention possesses millisecond-level predictive response capabilities, significantly surpassing traditional CFD simulations (which take hours to days), thus meeting the urgent needs of time-sensitive scenarios such as online real-time control of gas turbines and rapid iterative selection of multiple solutions.

[0011] 2. This invention creatively combines the efficiency of one-dimensional mechanistic models, the high precision of CFD simulations, 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 programs under complex flow conditions, significantly improves the model's prediction accuracy under key conditions, and effectively integrates multi-source data with different precision characteristics and distribution patterns.

[0012] 3. This invention reduces the reliance on expensive experiments and time-consuming CFD simulations for building high-precision compressor models, shortens the R&D cycle, and reduces testing costs. The constructed high-precision real-time model 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

[0013] Figure 1 This is a flowchart of the present invention; Figure 2 This is the sampling result of a one-dimensional aerodynamic performance program. Figure 3 For surrogate model neural network structure; Figure 4 This relates to the structure and process of transfer learning. Detailed Implementation

[0014] The invention will now be described in more detail with reference to the accompanying drawings: Combination Figure 1-4 The specific process of this invention is as follows: Constructing a surrogate model driven by a one-dimensional compressor mechanism model: (1) Based on a one-dimensional compressor mechanism model, multiple sets of input-output data are generated within the target operating condition range to construct an initial sample library. The model inputs include compressor speed, rotatable guide vane angle, pressure ratio, and bleed air volume, while the model outputs include compressor efficiency and compressor flow rate. (2) Perform data preprocessing, normalize the input and output data, use filtering methods to eliminate noise, and divide the training set and test set in an 8:2 ratio; (3) A BP neural network was selected for sample training. The four nodes of the input layer correspond to the compressor speed, the angle of the rotatable guide vane, the pressure ratio, and the bleed air volume. The two nodes of the output layer correspond to the compressor efficiency and the compressor flow rate. The hidden layer is a two-layer network with ReLU activation function and mean squared error (MSE) loss function. L1 regularization and Dropout structure optimization were added to the network to prevent overfitting. A variable learning rate strategy was adopted to achieve higher model accuracy. (4) Optimize the model structure and parameters to complete the verification of the one-dimensional surrogate model of the compressor, compare the prediction error of the surrogate model and the one-dimensional program at the unseen operating point, and ensure that the error of the surrogate model is within an acceptable range; Conducting CFD simulation data-driven transfer learning optimization: (5) Conduct three-dimensional CFD simulations for key operating points to obtain high-precision compressor efficiency, compressor flow data and flow field distribution; (6) Freeze the weights of the first two layers of the proxy model, fine-tune only the output layer, and use CFD data to locally optimize the model so that the model learns the shared features of one-dimensional data and CFD data. (7) Transfer the learned model to the overall performance model of the gas turbine and test and verify it.

[0015] 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 the blade profile, thus enabling accurate calculation of the parameters of the outlet blade throat.

[0016] The calculation method for the one-dimensional mechanistic model of the 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 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.

[0017] The main theoretical formulas used in the HARIKA algorithm are as follows, for flow calculation: Where G is the compressor mass flow rate. This is the total pressure at the compressor inlet. This refers to the total temperature at the compressor inlet. The traffic reserve coefficient is an adjustable data input by the user.

[0018] Efficiency calculation: Stage pressure ratio calculation: Using these calculation formulas, combined with other aerodynamic formulas and empirical corrections, the characteristic curve of the compressor is calculated, and the compressor performance is predicted.

[0019] The neural network employs a modified BP neural network architecture, whose core principle lies in achieving a non-linear mapping from input parameters to output parameters through the error backpropagation algorithm. The network structure is defined as follows: Input layer input vector Where n is the rotational speed, For the guide vane angle, For pressure ratio, This represents the induced draft volume. The number of input layer nodes is 4, and the data is normalized to the interval [-1, 1].

[0020] The hidden layers employ a two-layer fully connected structure, with the number of neurons in each layer dynamically optimized through sensitivity analysis. The activation function used in the hidden layers is the Modified Linear Unit (ReLU). in This is the weight matrix. For bias vectors, This is the output of the previous layer.

[0021] Output layer Efficiency and throughput are respectively considered. A linear activation function is used: The regularization strategy chosen is L1 regularization, which adds the sum of the absolute values ​​of the weights to the loss function to suppress overfitting during network training. In the formula, λ is the penalty coefficient, which is optimized through cross-validation.

[0022] During training, the Dropout mechanism randomly discards hidden layer neurons with a probability of p=0.5. During testing, all neurons are enabled and their weights are scaled.

[0023] Transfer learning freezes the weights of the first two layers, preserving the general flow characteristics from one-dimensional data training and avoiding overfitting due to insufficient CFD data. The output layer is fine-tuned using high-precision CFD data to adjust its weights, correcting systematic biases in the one-dimensional mechanistic model under complex flow conditions.

[0024] Let the parameters of the neural network be... W1 is the weight matrix from the input layer to the first hidden layer, W2 is the weight matrix from the first hidden layer to the second hidden layer, and W3 is the weight matrix from the second hidden layer to the output layer.

[0025] During transfer learning, W1 and W2 are frozen, and only W3 is optimized. The loss function is then modified as follows: The hyperparameters are determined through cross-validation. Define the dynamic fusion weights: The final prediction output is a weighted fusion: .

[0026] Example: One-dimensional mechanism model-driven surrogate model construction: Step 1: Sample generation and data preprocessing. A calculation program based on the aerodynamic performance mechanism model of a one-dimensional compressor in a marine gas turbine is used. The input parameters are relative speed (N, speed range 0.4~1.0) and rotatable guide vane angle (…). The range is -40° to -25°. Between 0.8 and 0.9 RPM, the rotational angle shows a linear relationship with the RPM, while for speeds below 0.8 and 0.9, the upper and lower limits of the rotational angle remain unchanged. (Pressure ratio) ), expiratory air volume (Q, range 0~3.5% of total flow); output parameter is efficiency ( The sample points are shown in the figure, which represent the flow rate (G) and the sample points.

[0027] 120,000 samples were generated within the operating range, covering typical scenarios such as the design point. The sample selection criteria were based on the actual control conditions of the gas turbine, with the guide vane angle maintained at -35° at low speeds and 3° at high speeds. 80% of the samples were used as the training set, and 20% as the test set.

[0028] The generated samples are preprocessed using Min-Max normalization to map the input and output parameters to the [0,1] interval. Noisy data is filtered to eliminate outliers.

[0029] Step 2: Neural Network Construction and Training. Since the one-dimensional performance model of the compressor generates steady-state data, a backpropagation (BP) neural network is selected for training. The network structure is determined as follows: The input layer has 4 nodes, corresponding to compressor speed, guide vane angle, bleed air volume, and pressure ratio, respectively; the hidden layer has two layers to avoid excessive network complexity while balancing model expressiveness and training efficiency, with ReLU selected as the activation function for both layers; the output layer has two nodes, corresponding to compressor efficiency and compressor flow rate, respectively; the optimizer is Adam, and L1 regularization and Dropout are added to the network to prevent overfitting during training.

[0030] Step 3: Model Parameter Optimization. Sliding window analysis can be used to compare the model's predicted output with the actual output of the original samples under different input samples, thus measuring the network training effect and guiding parameter optimization. The structures and parameters that need optimization include: initial learning rate, number of neurons in each hidden layer, number of training iterations, training batch size, L1 regularization coefficient, and Dropout coefficient.

[0031] like Figure 1 As shown, when the prediction errors of both the training set and the test set are within an acceptable range, the surrogate model driven by the constructed one-dimensional compressor mechanism model is considered to have met the accuracy requirements.

[0032] CFD simulation data-driven transfer learning optimization: Step 1: Generate and align CFD simulation data. ANSYS CFX is selected as the numerical calculation software, and SST k-ω is used as the turbulence model. Key operating points are selected for steady-state solutions to obtain steady-state efficiency, flow rate, and flow field distribution.

[0033] Step Two: Data Fusion Architecture Design, constructing a dual-channel data fusion framework: The basic channel inherits the trained one-dimensional data proxy model (BP neural network), retaining its generalization ability for common operating conditions. The correction channel adds a three-dimensional CFD feature correction network, using transfer learning technology to map high-precision CFD calculation results to the one-dimensional model residual space. The network topology adopts a parallel dual-branch structure, and the final output is: Step 3: The core objective of this step is to use a small amount of high-precision CFD simulation data to correct the prediction bias of the surrogate model driven by the one-dimensional mechanistic model under complex three-dimensional flow conditions, while retaining the model's good generalization ability under normal conditions. The specific operations are as follows: The one-dimensional mechanistic model trained in step two is used as the base model to drive the surrogate model. All weights (W1, W2) and bias terms of the first two layers of this model (from the input layer to the first hidden layer, and from the first hidden layer to the second hidden layer) are frozen. These two layers have already learned the compressor performance parameters (speed N, guide vane angle) during training on a large amount of one-dimensional data. Pressure ratio This study explores the deep, general nonlinear mapping between air intake (Q) and fundamental flow characteristics. Freezing these layers aims to preserve this general knowledge extracted from large-sample one-dimensional data, avoiding overfitting or forgetting fundamental principles due to relearning on small-sample CFD data.

[0034] Only the weights (W3) and biases of the last layer (from the second hidden layer to the output layer) are unfrozen and optimized. This layer is responsible for mapping the high-level features extracted by the preceding networks to the final predicted output (efficiency). (and flow G). Use the CFD dataset generated in step one as the new training data.

[0035] Define the transfer learning loss function. Considering the potential for small errors in CFD data and the limited sample size, and taking into account the reliability of the one-dimensional surrogate model in non-critical conditions, a weighted mean squared error loss function can be used: By optimizing various parameters and iteratively learning, the compressor performance model can be transferred from one dimension to three dimensions.

[0036] The above embodiments were implemented using Python scripts of MWORKS.Sysplorer software to generate a compressor transfer learning model, which was then integrated into the overall performance model of the gas turbine for testing and verification.

[0037] This invention aims to address the key bottlenecks in accuracy, efficiency, and generalization of traditional modeling techniques. By innovatively integrating the high efficiency of one-dimensional mechanistic modeling programs, the high accuracy of CFD simulations, and the adaptability of data-driven models, this invention achieves a breakthrough improvement in compressor performance prediction.

[0038] Compared with existing technologies, the innovation of this invention lies in constructing a dynamic correction mechanism based on transfer learning, which significantly improves the prediction accuracy of the surrogate model under key operating conditions using a small amount of CFD simulation data; and developing a lightweight network architecture that achieves real-time simulation while maintaining engineering interpretability, laying the foundation for virtual and real control of gas turbines.

[0039] This invention provides core technological support for the digital design, intelligent operation and maintenance, and performance optimization of marine gas turbines. It can significantly shorten the R&D cycle and reduce testing costs, and has significant strategic value in promoting the intelligent upgrading of marine power systems. The widespread application of this invention will effectively promote the digital transformation of the shipbuilding industry and provide key technological guarantees for the development of green and low-carbon ships.

Claims

1. A method for modeling the performance of marine gas turbine compressors based on transfer learning, characterized by: Includes the following steps: (1) Based on a one-dimensional compressor mechanism model, input-output data are generated within the target operating condition range to construct an initial sample library; wherein, the model input includes compressor speed, rotatable guide vane angle, pressure ratio, and bleed air volume, and the model output includes compressor efficiency and compressor flow rate; (2) Perform data preprocessing, normalize the input and output data, use filtering methods to eliminate noise, and divide the training set and test set in an 8:2 ratio; (3) A BP neural network was selected for sample training. The four nodes of the input layer correspond to the compressor speed, the angle of the rotatable guide vane, the pressure ratio, and the bleed air volume. The two nodes of the output layer correspond to the compressor efficiency and the compressor flow rate. The hidden layer is a two-layer network with ReLU activation function and mean square error (MSE) loss function. L1 regularization and Dropout structure optimization were added to the network to prevent overfitting. A variable learning rate strategy was adopted to achieve higher model accuracy. (4) Optimize the model structure and parameters to complete the verification of the one-dimensional surrogate model of the compressor, compare the prediction errors of the surrogate model and the one-dimensional program at unseen operating points, and ensure that the error of the surrogate model is within an acceptable range; carry out CFD simulation data-driven transfer learning optimization: (5) Conduct three-dimensional CFD simulations for the required operating points to obtain compressor efficiency, compressor flow data and flow field distribution; (6) Freeze the weights of the first two layers of the proxy model, fine-tune the output layer, and use CFD data to perform local optimization of the model so that the model learns the shared features of one-dimensional data and CFD data. (7) Transfer the learned model to the overall performance model of the gas turbine and test and verify it.

2. The method for modeling the performance of marine gas turbine compressors based on transfer learning according to claim 1, characterized in that: In step (1), the compressor characteristics calculation of the compressor mechanism model adopts the one-dimensional HARIKA algorithm. The HARIKA algorithm is based on the stage superposition method, performs empirical corrections along the blade height, and also performs empirical corrections on the characteristic parameters of each blade profile, thereby calculating the parameters of the outlet blade throat: The calculation method for the one-dimensional mechanistic model of compressor performance is as follows: Based on the known compressor geometry, inlet total temperature, total pressure, and circumferential velocity parameters, solve for the velocity triangle on the average radius of the rotor inlet section; use empirical methods to solve for the total pressure ratio and adiabatic efficiency of this stage; use the same method to further obtain the velocity triangle at the stage outlet, and use this outlet parameter as the inlet parameter of the next stage to calculate the outlet parameter of the next stage at this flow point; repeat this process until the outlet parameter of the last stage of the compressor is obtained, then change the flow rate and speed, and repeat the above process until the performance parameters of the entire compressor are obtained; The theoretical formula used in the HARIKA algorithm for flow calculation is as follows: Where G is the compressor mass flow rate. This is the total pressure at the compressor inlet. This refers to the total temperature at the compressor inlet. The traffic reserve coefficient is an adjustable data input by the user. Efficiency calculation: Stage pressure ratio calculation: By combining aerodynamic formulas and empirical corrections, the characteristic curve of the compressor is calculated, and the compressor performance is predicted.

3. The method for modeling the performance of marine gas turbine compressors based on transfer learning according to claim 1, characterized in that: The neural network adopts an improved BP neural network architecture, and achieves a nonlinear mapping from input parameters to output parameters through the error backpropagation algorithm. The network structure is defined as follows: Input layer input vector Where n is the rotational speed, For the guide vane angle, For pressure ratio, To determine the air intake, the number of input layer nodes is 4, and the data is normalized to the range [-1,1]. The hidden layers employ a two-layer fully connected structure, with the number of neurons in each layer dynamically optimized through sensitivity analysis. The activation function for the hidden layers is the modified linear unit ReLU. in This is the weight matrix. For bias vectors, For output from the previous layer; Output layer For efficiency and flow rate, linear activation functions are used respectively: The regularization strategy chosen is L1 regularization, which adds the sum of the absolute values ​​of the weights to the loss function to suppress overfitting during network training. In the formula, λ is the penalty coefficient, which is optimized through cross-validation; During training, the Dropout mechanism randomly discards hidden layer neurons with a probability of p=0.

5. During testing, all neurons are enabled and their weights are scaled.

4. The method for modeling the performance of marine gas turbine compressors based on transfer learning according to claim 1, characterized in that: In step (6), the transfer learning freezes the weights of the first two layers, retains the general flow characteristics of the one-dimensional data training, fine-tunes the output layer, uses CFD data to adjust the weights of the output layer, and corrects the systematic bias of the one-dimensional mechanism model under complex flow conditions. Let the parameters of the neural network be... W1 is the weight matrix from the input layer to the first hidden layer, W2 is the weight matrix from the first hidden layer to the second hidden layer, and W3 is the weight matrix from the second hidden layer to the output layer. During transfer learning, W1 and W2 are frozen, and only W3 is optimized. The loss function is then modified as follows: The hyperparameters are determined through cross-validation, and the dynamic fusion weights are defined as follows: The final prediction output is a weighted fusion: 。

Images