Gas turbine efficiency prediction method

Abstract

The gas turbine performance prediction method provided in this application specifically relates to the field of industrial data prediction technology. This method first constructs a CFD simulation model based on the geometric model and physical structural parameters of the gas turbine combustor, and then calculates the predicted value of the first combustor outlet temperature using mechanistic methods. Next, it uses real-time monitoring data to extract key features through a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm. Then, it inputs these key features into a LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustor outlet temperature. The two predicted values ​​are then input into a Stacking ensemble model. Finally, the meta-learner in the model performs weighted fusion of the two predicted values, outputting the final predicted result of the combustor outlet temperature. This method solves the technical problem of synergistic optimization between the insufficient real-time performance of the CFD model and the limited prediction accuracy of the LightGBM model, thereby achieving efficient and accurate prediction of gas turbine performance and intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of industrial data prediction technology, and more specifically, to a method for predicting the performance of a gas turbine. Background Technology

[0002] Thermal power generation is a core sector of my country's energy supply, and the operating efficiency of its gas turbine units directly impacts the economy and stability of the power system. However, as complex energy conversion equipment operating at high temperatures and speeds, the real-time monitoring of key parameters such as compressor outlet temperature and combustion efficiency in gas turbines faces significant technical bottlenecks. Traditional industrial monitoring methods struggle to directly measure parameters of high-temperature, sealed components, while single-mechanism models or data-driven models have significant limitations in real-time response speed, feature correlation analysis, and multi-condition generalization capabilities. This results in insufficient accuracy in unit performance prediction, and the risk of unplanned shutdowns has long plagued the industry.

[0003] While existing technologies such as computational fluid dynamics (CFD) simulation can analyze the internal mechanisms of gas turbines through multiphysics coupling, they are limited by the computational power requirements for high-fidelity modeling, making it difficult to meet the requirements for minute-level real-time prediction. Data-driven methods based on LightGBM can rapidly model using SIS system monitoring data, but they suffer from difficulties in identifying characteristic time-delay correlations and weak physical interpretability. Traditional solutions often employ a single model architecture or a simple weighted fusion strategy, failing to effectively coordinate the spatiotemporal alignment of mechanism simulation and measured data, creating a significant technical gap between improving model accuracy and engineering practicality.

[0004] In summary, how to solve the technical problem of synergistic optimization between the insufficient real-time performance of CFD models and the limited prediction accuracy of LightGBM models is a technical bottleneck that urgently needs to be overcome to achieve high-precision real-time prediction of gas turbine performance. Summary of the Invention

[0005] The main objective of this invention is to provide a gas turbine performance prediction method to at least solve the technical problem of synergistic optimization between the insufficient real-time performance of CFD models and the limited prediction accuracy of LightGBM models, thereby achieving efficient and accurate prediction of gas turbine performance and intelligent operation and maintenance.

[0006] To achieve the above objectives, the present invention provides a method for predicting the performance of a gas turbine.

[0007] This invention provides a method for predicting the performance of a gas turbine, the method comprising:

[0008] A CFD simulation model is constructed based on the geometric model and physical structural parameters of the gas turbine combustion chamber. The combustion chamber outlet temperature of the gas turbine is calculated by mechanism to obtain the predicted value of the first combustion chamber outlet temperature.

[0009] Based on the real-time monitoring data of the gas turbine, key features are extracted using a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm.

[0010] The key features are input into the LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustion chamber outlet temperature.

[0011] The predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature are input into the Stacking ensemble model.

[0012] The meta-learner in the Stacking ensemble model performs inter-model weighted fusion of the predicted values ​​of the first and second combustion chamber outlet temperatures, and outputs the final predicted combustion chamber outlet temperature result.

[0013] Specifically, the CFD simulation model constructed based on the gas turbine's combustion chamber geometric model and physical structural parameters is used to perform a mechanistic calculation on the gas turbine's combustion chamber outlet temperature to obtain a predicted value for the first combustion chamber outlet temperature, including:

[0014] Based on the combustion chamber geometric model and physical structural parameters of the gas turbine, the fuel flow rate, air-fuel mixture ratio, and combustion reaction boundary conditions are set.

[0015] Based on the fuel flow rate, air-to-fuel ratio, and combustion reaction boundary conditions, the heat released during combustion is obtained through computational fluid dynamics simulation.

[0016] Based on the ratio of the heat released by combustion to the theoretical heat of the fuel, the predicted value of the outlet temperature of the first combustion chamber is calculated.

[0017] Specifically, the key features extracted from the real-time monitoring data of the gas turbine using a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm include:

[0018] Gradient calculation is performed on the real-time monitoring data to select the first gradient data sample and the second gradient data sample;

[0019] Based on the first gradient data sample and the second gradient data sample, a preliminary key feature set is generated;

[0020] The feature parameters in the preliminary key feature set are mutually exclusive feature bundled to generate the key features.

[0021] Specifically, the step of performing gradient calculation on the real-time monitoring data and filtering out the first gradient data sample and the second gradient data sample includes:

[0022] Gradient calculation is performed on the real-time monitoring data, and the first gradient data samples with the top 5%-15% absolute gradient values ​​are retained.

[0023] Randomly select 10%-30% of the second gradient data samples from the remaining real-time monitoring data;

[0024] A weight compensation coefficient is introduced for the second gradient data sample. ,in b is the retention ratio of the first gradient samples, and b is the extraction ratio of the second gradient samples.

[0025] Specifically, the step of generating the key features by mutually exclusive feature bundling of the feature parameters in the preliminary key feature set includes:

[0026] Bind multiple pressure sensor data in the aforementioned feature parameters into a pressure feature group;

[0027] Bind multiple temperature sensor data in the aforementioned feature parameters into a temperature feature group;

[0028] The pressure feature set and temperature feature set are used as the key features.

[0029] Specifically, the step of inputting the key features into the LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustion chamber outlet temperature includes:

[0030] The key features are standardized to obtain a standardized feature vector;

[0031] Based on the standardized feature vector, multiple gradient boosting decision trees are constructed. Each gradient boosting decision tree is split through a leaf node growth strategy, with the splitting rule being to maximize information gain.

[0032] Based on the prediction results of the gradient boosting decision tree, the predicted value of the second combustion chamber outlet temperature is output through weighted summation within the model.

[0033] Specifically, the step of performing inter-model weighted fusion of the predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature through the meta-learner in the Stacking ensemble model to output the final combustion chamber outlet temperature prediction result includes:

[0034] Based on the real-time operating parameters of the gas turbine, the weighting coefficient α of the predicted value of the first combustion chamber outlet temperature is dynamically calculated, where the value of α ranges from 0.2 to 0.4.

[0035] Based on the weighting coefficient α, the predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature are fused using a weighted formula:

[0036]

[0037] in, This is the predicted value for the outlet temperature of the first combustion chamber. This is the predicted value for the outlet temperature of the second combustion chamber;

[0038] The weighted fusion result is output as the final combustion chamber outlet temperature prediction result.

[0039] This application provides a gas turbine performance prediction method that integrates mechanistic modeling and data-driven techniques to improve the accuracy of combustion chamber outlet temperature prediction. Specifically, a CFD simulation model is constructed based on the gas turbine combustion chamber's geometric model and physical structural parameters. Mechanistic calculations are performed to obtain the first predicted combustion chamber outlet temperature. Simultaneously, real-time monitoring data of the gas turbine is collected, and key features are extracted using gradient one-sided sampling and mutually exclusive feature binding algorithms. These features are then input into a LightGBM prediction model for data-driven calculation to obtain the second predicted combustion chamber outlet temperature. Subsequently, the two predicted values ​​are input into a Stacking ensemble model, and a meta-learner is used to perform weighted fusion between the two models. Taking into account both mechanistic characteristics and data features, an accurate and reliable combustion chamber outlet temperature prediction result is finally output. This method solves the technical problem of synergistic optimization between the insufficient real-time performance of CFD models and the limited prediction accuracy of LightGBM models, thereby achieving efficient and accurate prediction of gas turbine performance and intelligent operation and maintenance. Attached Figure Description

[0040] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0041] Figure 1 A flowchart illustrating the gas turbine performance prediction method provided in this application;

[0042] Figure 2 Flowchart for training the LightGBM prediction model;

[0043] Figure 3 This is a schematic diagram of the Stacking integrated model architecture and training.

[0044] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0047] In this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0048] The gas turbine performance prediction method provided in this application is based on the integration of the advantages of mechanistic modeling and data-driven approaches to improve the prediction accuracy of combustion chamber outlet temperature. First, a CFD simulation model is constructed using the combustion chamber's geometric model and physical structural parameters. Mechanistic calculations are then performed based on combustion mechanisms to obtain a first predicted value. Next, key features are extracted from real-time monitoring data using gradient unilateral sampling and mutually exclusive feature binding algorithms. These features are then processed by the LightGBM model to obtain a second predicted value. Finally, the meta-learner in the Stacking ensemble model fuses the two predicted values ​​to output the final result.

[0049] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0050] Figure 1 This is a flowchart illustrating the gas turbine performance prediction method provided in this application, intended to provide a detailed explanation of the gas turbine performance prediction method. Figure 1 As shown in this embodiment, the gas turbine performance prediction method includes:

[0051] S101: Based on the geometric model and physical structural parameters of the gas turbine's combustion chamber, a CFD simulation model is constructed, and the combustion chamber outlet temperature of the gas turbine is calculated mechanically to obtain the predicted value of the first combustion chamber outlet temperature.

[0052] The CFD simulation model, constructed based on the gas turbine's combustion chamber geometric model and physical structural parameters, performs a mechanistic calculation on the gas turbine's combustion chamber outlet temperature to obtain a predicted value for the first combustion chamber outlet temperature, including:

[0053] Based on the combustion chamber geometric model and physical structural parameters of the gas turbine, the fuel flow rate, air-fuel mixture ratio, and combustion reaction boundary conditions are set.

[0054] Based on the fuel flow rate, air-to-fuel ratio, and combustion reaction boundary conditions, the heat released during combustion is obtained through computational fluid dynamics simulation.

[0055] Based on the ratio of the heat released by combustion to the theoretical heat of the fuel, the predicted value of the outlet temperature of the first combustion chamber is calculated.

[0056] The specific implementation includes:

[0057] 1. Construct a three-dimensional geometric model of the combustion chamber:

[0058] The geometric parameters of the combustion chamber were extracted from the CAD drawings provided by the gas turbine manufacturer. These parameters included the inner diameter of the annular structure (550 mm), the outer diameter (680 mm), the axial length of the flame tube (320 mm), the transition section cone angle (15°), and the annular array layout of 24 fuel nozzles with diameters of 8 mm and a spacing of 45 mm. A three-dimensional solid model of the combustion chamber was created using ANSYS SpaceClaim software and exported as a geometric file in STEP AP214 format.

[0059] 2. Define physical structure parameters:

[0060] Import the geometry file into ANSYS Fluent 2022R1, and set the combustion chamber wall material to Inconel 718 nickel-based superalloy, with a density of 8190 kg / m³, a thermal conductivity of 11.2 W / (m·K), and a specific heat capacity of 435 J / (kg·K). The fuel pipeline material is defined as SA-240 316L stainless steel, and the inner wall of the airflow channel is sprayed with a 0.5 mm thick yttrium-stabilized zirconia (YSZ) thermal barrier coating.

[0061] 3. Set boundary conditions for the CFD simulation model:

[0062] The fuel nozzle inlet adopts the mass flow inlet condition: methane fuel flow rate Q_fuel = 2.3 kg / s, with a dynamic adjustment range of 1.8-2.8 kg / s (corresponding to 50-100% gas turbine load).

[0063] The air inlet at the compressor outlet is set as a velocity inlet condition: flow velocity V_air = 120 m / s, turbulence intensity 5%, temperature T_air = 723 K;

[0064] The combustion chamber outlet is set to pressure outlet condition: back pressure P_out = 2.35 MPa;

[0065] The wall boundary adopts a no-slip thermal insulation condition with a surface roughness Ra=3.2 μm.

[0066] 4. Configure the combustion reaction model:

[0067] We selected the finite-rate / eddy-dissipation combustion model and defined the methane-air chemical reaction mechanism as the GRI-Mech 3.0 detailed mechanism, which includes 53 components and 325 elementary reactions.

[0068] The turbulence model uses a Realizable k-epsilon model with Enhanced Wall Treatment activated.

[0069] The radiation model uses the Discrete Ordinates Model, and the combustion gas absorption coefficient uses the Weighted Ash Gas Model (WSGG).

[0070] 5. Execute the transient CFD simulation model:

[0071] A hexahedral structured mesh was generated using ICEM CFD, with 5 prism layers set in the near-wall region. The height of the first layer was 0.01 mm, the total number of meshes was 3.8 million, and the Jacobian matrix quality was >0.3.

[0072] The transient solver time step was set to Δt = 1e-4 seconds, the total simulation time was 0.5 seconds, and the SIMPLE algorithm was used to solve the Navier-Stokes equations.

[0073] Monitor the mass-weighted average temperature of the combustion chamber outlet section until the residual curve converges to the order of 10^-5.

[0074] 6. Calculate the heat released by combustion:

[0075] Extracting fuel consumption rate from simulation results =2.28 kg / s, oxygen consumption ΔO2=4.56 kg / s;

[0076] Based on the lower heating value of methane, LHV = 50.0 MJ / kg, calculate the actual heat released, Q_actual = ×LHV=2.28×50=114 MW;

[0077] Theoretical heat of complete combustion:

[0078] Q_theoretical= ×ΔH_combustion=2.28×55.5=126.54MW;

[0079] Wherein, ΔH_combustion is the molar enthalpy of combustion of methane;

[0080] Combustion efficiency η = Q_actual / Q_theoretical × 100% = 114 / 126.54 × 100% = 90.1%.

[0081] 7. Calculate the predicted outlet temperature:

[0082] Applying the energy conservation equation:

[0083]

[0084] in: =45 kg / s (compressor air mass flow rate)

[0085] =1.005 kJ / (kg·kg·kJ) (Specific heat capacity of air at constant pressure)

[0086] =723 K (compressor outlet temperature)

[0087] Substitute into the calculation: =723+45×1.0050.901×126.54×10=723+253=976 K;

[0088] Output the predicted value of the first combustion chamber outlet temperature =976±12 K (error from grid sensitivity analysis).

[0089] This step accurately analyzed the turbulent combustion, heat transfer, and mass transfer processes within the combustion chamber using a 3D CFD simulation model. A structured mesh and a Realizable k-epsilon model were employed to control the temperature prediction error within ±12K, achieving a 60% improvement in accuracy compared to traditional zero-dimensional models. By setting parameterized boundary conditions (such as dynamic adjustment of fuel flow), rapid simulations under different operating conditions were achieved (single-condition calculation time <4 hours). The quantitative calculation of combustion efficiency η provided a physical constraint benchmark for subsequent data-driven models, ensuring that the prediction results conform to fundamental thermodynamic principles.

[0090] S102: Based on the real-time monitoring data of the gas turbine, key features are extracted using a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm;

[0091] The key features extracted from the real-time monitoring data of the gas turbine using a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm include:

[0092] Gradient calculation is performed on the real-time monitoring data to select first gradient data samples and second gradient data samples. Specifically, the gradient calculation and selection of first and second gradient data samples includes: performing gradient calculation on the real-time monitoring data and retaining the first gradient data samples with the top 5%-15% of the absolute gradient values; randomly selecting 10%-30% of the second gradient data samples from the remaining real-time monitoring data; and introducing a weight compensation coefficient to the second gradient data samples. ,in b is the retention ratio of the first gradient samples, and b is the extraction ratio of the second gradient samples.

[0093] Based on the first gradient data sample and the second gradient data sample, a preliminary key feature set is generated;

[0094] The key features are generated by binding mutually exclusive features to the feature parameters in the preliminary key feature set. Specifically, binding mutually exclusive features to the feature parameters in the preliminary key feature set to generate the key features includes: binding multiple pressure sensor data in the feature parameters into a pressure feature group; binding multiple temperature sensor data in the feature parameters into a temperature feature group; and using the pressure feature group and the temperature feature group as the key features.

[0095] Specific embodiments of step S102 include:

[0096] 1. Obtain real-time monitoring dataset:

[0097] 1.1 Data from 15 types of sensors were collected through the gas turbine SIS system, including: pressure sensor group: combustion chamber inlet pressure P_in (0-3MPa), compressor outlet pressure P_out (0-2.5MPa), turbine inlet pressure P_turbine (0-2.8MPa), with a sampling frequency of 100Hz;

[0098] Temperature sensor group: compressor outlet temperature T_comp (300-800K), fuel temperature T_fuel (280-320K), cooling air temperature T_cool (400-600K), sampling frequency 50Hz;

[0099] Flow meters: fuel flow rate Q_fuel (1.5-3.0 kg / s), air flow rate Q_air (30-60 kg / s), sampling frequency 10 Hz;

[0100] 1.2 Constructing a time series dataset X∈ (N×15), time window N=300 (corresponding to 3 seconds of data).

[0101] 2. Perform gradient-based one-side sampling (GOS): Calculate the absolute value of the gradient:

[0102] Calculate the first gradient of the dataset X using the loss function of the LightGBM prediction model:

[0103]

[0104] Where L is the mean squared error loss function, This is the actual temperature. This is the current model's predicted value.

[0105] Filtering first gradient samples:

[0106] according to Sort in descending order, keeping the first few. =10% of the high gradient samples (N1=300×10%=30 samples), forming the set S_high;

[0107] Randomly select second gradient samples:

[0108] From the remaining 90% of the data, randomly select b=20% of the samples (N2=300×90%×20%=54 samples) to form the set S_low;

[0109] Introducing a weighting compensation coefficient:

[0110] Apply a weight compensation factor (1-) to the S_low sample. ) / b=(1-0.1) / 0.2=4.5, ensuring that the contribution of low gradient samples in the loss calculation is balanced.

[0111] 3. Generate a preliminary key feature set:

[0112] 3.1 Merge the S_high and S_low samples to obtain the total number of samples N_total = 30 + 54 = 84;

[0113] 3.2 Mutual information calculation was performed on the 15-dimensional original features, and features with a correlation > 0.7 with the combustion chamber outlet temperature were selected:

[0114] Key features include: P_in, P_turbine, T_comp, Q_fuel, and Q_air;

[0115] 3.3 Constructing the preliminary key feature matrix F∈ (84×5).

[0116] 4. Perform exclusive feature bundling (EFB):

[0117] 4.1 Construction of pressure feature set:

[0118] P_in (that is) ), P_turbine (i.e. Perform principal component analysis (PCA) and retain principal components with a variance contribution rate > 95%.

[0119]

[0120] Temperature feature set construction:

[0121] For T_comp (that is) Perform polynomial expansion (quadratic terms and interaction terms) to generate new features:

[0122]

[0123] Traffic feature processing:

[0124] Q_fuel (that is) ) and Q_air (that is) Combined into equivalent ratio characteristics:

[0125]

[0126] Where 0.067 is the methane-air stoichiometric ratio.

[0127] 5. Generate final key features:

[0128] 5.1 The output key feature set contains 3 feature groups:

[0129] Pressure characteristic group F_pressure (i.e. (Dimension 1);

[0130] Temperature feature group F_temp (i.e. (Dimension 1);

[0131] Equivalent ratio characteristic F_ratio (i.e.) (Dimension 1);

[0132] 5.2 Constructing the final feature matrix F_final∈ (84×3) is used as the input to the LightGBM prediction model.

[0133] This step compresses the data volume from 300 samples to 84 samples using the GOSS algorithm (a 72% reduction), and combines it with EFB to reduce the feature dimension from 15 dimensions to 3 dimensions, shortening the training time of the LightGBM prediction model by 65%. The gradient weight compensation mechanism reduces the prediction error in low-gradient regions by 18.7% (comparative experiments show MAE decreasing from ±23K to ±18.7K). PCA processing in mutually exclusive feature bundling increases the variance explained by the stress feature group to 97.3%, effectively eliminating sensor collinearity. Compared to the traditional Pearson correlation coefficient screening method, this embodiment improves the correlation between features and the target variable from 0.62 to 0.81, providing a high signal-to-noise ratio input for subsequent prediction models.

[0134] S103: Input the key features into the LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustion chamber outlet temperature;

[0135] The step of inputting the key features into the LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustion chamber outlet temperature includes:

[0136] The key features are standardized to obtain a standardized feature vector;

[0137] Based on the standardized feature vector, multiple gradient boosting decision trees are constructed. Each gradient boosting decision tree is split through a leaf node growth strategy, with the splitting rule being to maximize information gain.

[0138] Based on the prediction results of the gradient boosting decision tree, the predicted value of the second combustion chamber outlet temperature is output through weighted summation within the model.

[0139] Specific embodiments of step S103 include:

[0140] 1. Standardization of key features:

[0141] 1.1 The input features are S102, generating a 3D key feature matrix F_final∈ (84×3), including pressure characteristic group, temperature characteristic group and equivalent ratio characteristic;

[0142] 1.2 The Z-score standardization method is used to calculate the standardized values ​​of each feature dimension according to the formula:

[0143]

[0144] in:

[0145] : The j-th original feature value (j=1,2,3 correspond to pressure, temperature, and equivalence ratio features, respectively);

[0146] : The mean of the j-th dimension feature in the training set (calculated using historical 30-day data, e.g., μ_pressure = 2.12MPa, μ_temperature = 752K, μ_equivalence ratio = 1.05);

[0147] : Standard deviation of the j-th dimension feature of the training set (e.g., σ_pressure = 0.35MPa, σ_temperature = 28K, σ_equivalence = 0.12);

[0148] 1.3 Output the standardized feature vector Z∈ (84×3), with a mean of 0 and a standard deviation of 1 for each dimension.

[0149] 2. Gradient boosting decision tree construction:

[0150] 2.1 Setting the parameters of the LightGBM prediction model:

[0151] Number of trees: num_trees = 100;

[0152] Learning rate = 0.1;

[0153] The maximum number of leaf nodes is max_leaves=31;

[0154] The minimum data size in a leaf node is min_data_in_leaf = 5;

[0155] 2.2 Single Tree Splitting Rules:

[0156] Calculate the information gain of candidate split points:

[0157]

[0158] in:

[0159] : The first gradient of sample i (from the GOSS calculation result of S102);

[0160] : The second gradient of sample i (the second derivative, which is a constant 1 for the MSE loss function);

[0161] λ=1: L2 regularization coefficient;

[0162] The split point with the highest information gain is selected, and the histogram-based algorithm is used to accelerate the calculation. Each feature is divided into 256 bins.

[0163] 3. Leaf node growth strategy:

[0164] A leaf-wise growth strategy is adopted, and each split selects the node with the largest gain among all current leaf nodes for splitting.

[0165] The maximum depth of a single tree is limited to 7 levels to prevent overfitting;

[0166] The output of each tree is the weight of its leaf nodes. Calculated using Newton's method:

[0167]

[0168] Where k represents the k-th leaf node.

[0169] 4. Intra-model weighted prediction:

[0170] For an input sample x, the prediction value of the t-th tree is the weight of its leaf node. ;

[0171] Weighted output of all trees:

[0172]

[0173] in:

[0174] : The predicted value after the m-th iteration;

[0175] η=0.1: Learning rate;

[0176] After 100 tree iterations, the predicted value of the second combustion chamber outlet temperature is output:

[0177]

[0178] in =85K =920K is the standard deviation and mean of the training set temperature (inverse standardization parameter).

[0179] 5. LightGBM Prediction Model Construction

[0180] 5.1 Model Parameter Configuration

[0181] Number of gradient boosting trees: 150 (verified by grid search, range 100-200)

[0182] Learning rate: 0.08 (determined by early stopping; training is terminated when the validation set loss decreases by less than 1% for 5 consecutive iterations).

[0183] Maximum depth of a single tree: 6 layers (matching the physical parameters of the combustion chamber)

[0184] Minimum number of data points for leaf nodes: 8 samples (to prevent overfitting with small samples).

[0185] Number of feature bins: 512 (balancing computational accuracy and efficiency)

[0186] 5.2 Regularization Settings

[0187] L2 regularization coefficient λ = 1.2 (controls the weight magnitude of leaf nodes)

[0188] Feature sampling rate = 0.7 (70% of features are randomly selected for splitting in each iteration)

[0189] 5.3 Training optimization methods, such as Figure 2 As shown, Figure 2 Flowchart for training the LightGBM prediction model.

[0190] This step eliminates feature dimension differences through Z-score normalization, improving the convergence speed of the LightGBM prediction model by 40% (comparative experiments show that the number of training iterations decreased from 200 to 120). Employing a leaf-wise growth strategy and histogram algorithm, the training time for a single tree was reduced from 15ms to 4ms, and the total training time for 100 trees was controlled within 0.4 seconds, meeting real-time requirements. L2 regularization (λ=1.2) was introduced in the information gain calculation, reducing the overfitting rate of the model on the test set from 12.3% to 3.8%. The weighted prediction mechanism achieved a mean absolute error (MAE) of ±14.5K for the predicted second combustion chamber outlet temperature, a 22.7% reduction compared to the traditional XGBoost model. Inverse normalization restores the physical dimensions, ensuring that the prediction results can be directly used in the gas turbine control system.

[0191] S104: Input the predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature into the Stacking ensemble model.

[0192] Specific embodiments of step S104 include:

[0193] 1. Construct the input dataset for the Stacking ensemble model:

[0194] 1.1 Input data source:

[0195] Predicted temperature at the outlet of the first combustion chamber Simulation results from the CFD simulation model of S101, time series data (sampling interval 1 second), format is [ , ,..., Temperature range: 900-1050K;

[0196] Predicted temperature at the outlet of the second combustion chamber Output from the LightGBM prediction model of S103, time series data (sampling interval 0.5 seconds), format is [ , ,..., ].

[0197] 1.2 Time alignment processing:

[0198] right Perform linear interpolation to unify the sampling frequency to 1 second, and generate an aligned dataset. .

[0199] 2. Define the Stacking integration model structure:

[0200] 2.1 Base Learner Layer (Level-0):

[0201] Input features: and ;

[0202] Output: A two-column matrix of predicted values .

[0203] 2.2 Meta-learner layer (Level-1):

[0204] Input features: Expanded feature matrix (Including original predicted values ​​and derived features);

[0205] Model selection: Ridge Regression model, regularization coefficient λ=1.0.

[0206] 3. Generate training data for the meta-learner:

[0207] 3.1 Feature Expansion:

[0208] 3.1.1 to Add the following statistical features:

[0209] Sliding window mean: window size = 5 seconds, step size = 1 second, calculation and The mean;

[0210] First-order difference: ;

[0211] Operating parameters: Real-time load command (50-100% range, 1% resolution).

[0212] 3.1.2 Generating the Extended Feature Matrix The dimension is N×5.

[0213] 3.2 Data partitioning:

[0214] Using 5-fold cross validation, The data is divided into 5 subsets in chronological order, with each subset containing 20% ​​of the data in a continuous sequence to ensure the integrity of the time series.

[0215] 4. Data standardization processing:

[0216] 4.1 pairs Z-score normalization is performed on each column of features:

[0217]

[0218] in:

[0219] : The mean of the j-th column feature in the training set (calculated based on 30 days of historical data);

[0220] : The standard deviation of the j-th column feature in the training set;

[0221] 4.2 Output the normalized matrix .

[0222] 5. Input Stacking ensemble model:

[0223] 5.1 The standardized matrix With actual temperature label (Measured outlet temperatures from the gas turbine SCADA system) are aligned by timestamps to construct the final input dataset. ;

[0224] Input to the meta-learner training interface of the Stacking ensemble model, ready to execute the weighted fusion of S105.

[0225] 5.2 Stacking ensemble model structure and training, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the Stacking integrated model architecture and training.

[0226] Stacking ensemble models use the predictions of multiple models as new features input into another model (called a meta-learner) for final prediction. Essentially, it uses a new model (sub-learner) to learn how to combine base learners. If Bagging is viewed as a linear combination of multiple base classifiers, then Stacking is a non-linear combination, allowing learners to be flexibly stacked layer by layer to form a mesh structure. By combining the advantages of multiple models, it effectively reduces potential bias and variance, improving prediction accuracy and stability. Therefore, the Stacking method not only leverages the strengths of different models to improve generalization ability but also offers high flexibility, combining various types of base models and strong adaptability. Since high-fidelity mechanism simulation CFD models and high-precision, lightweight data-driven LightGBM prediction models possess advantages at different levels, training diverse data layers (simulation data + monitoring data), and achieving lightweight deployment with parallel computing power and real-time prediction response, the Stacking ensemble method can maximize the accuracy of ensemble model predictions. Therefore, this invention will use a stacking approach to integrate the mechanism simulation CFD simulation model and the data-driven LightGBM prediction model, forming a prediction of the performance of thermal power gas turbines based on the CFD simulation model, the LightGBM prediction model and the stacking integrated model.

[0227] The Stacking ensemble model training process includes: forming training and testing sets based on operating conditions; performing K-fold cross-training CFD simulation analysis to obtain predictions; training the LightGBM model for prediction using K-fold cross-training; and training a combination of the CFD simulation model and the LightGBM prediction model using the Stacking meta-learner.

[0228] The Stacking ensemble model architecture and training are as follows: Figure 3 As shown.

[0229] This step expands the single-sample features from 2D to 5D through time alignment and feature expansion (moving mean, difference, operating parameters), increasing the input information of the meta-learner by 150% and reducing the cross-validation mean squared error (CV-MSE) by 28.5%. Employing 5-fold time series cross-validation avoids the time discontinuity problem caused by traditional random partitioning, improving the model's predictive stability on the test set by 19%. Standardization unifies the dimensions of each feature, improving the comparability of the weight coefficients in the ridge regression model and providing a robust input data foundation for subsequent dynamic weighted fusion. Experiments show that this step reduces the overfitting rate of the Stacking ensemble model from 8.7% to 2.3%, ensuring the engineering practicality of the final temperature prediction results.

[0230] S105: The predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature are weighted and fused between models using the meta-learner in the Stacking ensemble model, and the final combustion chamber outlet temperature prediction result is output.

[0231] The step of performing inter-model weighted fusion of the predicted values ​​of the first and second combustion chamber outlet temperatures using the meta-learner in the Stacking ensemble model to output the final combustion chamber outlet temperature prediction result includes:

[0232] Based on the real-time operating parameters of the gas turbine, the weighting coefficient α of the predicted value of the first combustion chamber outlet temperature is dynamically calculated, where the value of α ranges from 0.2 to 0.4.

[0233] Based on the weighting coefficient α, the predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature are fused using a weighted formula:

[0234]

[0235] in, This is the predicted value for the outlet temperature of the first combustion chamber. This is the predicted value for the outlet temperature of the second combustion chamber;

[0236] The weighted fusion result is output as the final combustion chamber outlet temperature prediction result.

[0237] Specific embodiments of step S105 include:

[0238] 1. Calculation of dynamic weighting coefficient α:

[0239] 1.1 Input real-time operating parameters:

[0240] The gas turbine load rate L (50%-100%) is obtained in real time through the Modbus protocol of the control system.

[0241] Compressor outlet pressure P_comp (1.8-2.5MPa), sampling frequency 10Hz;

[0242] Fuel flow rate Q_fuel (i.e. (1.8-2.8kg / s), resolution 0.1kg / s.

[0243] 1.2 Calculation Rules:

[0244] The value of α is determined using piecewise linear interpolation:

[0245]

[0246] For example, when L=80% and Q_fuel=2.3kg / s, α=0.2+0.2×(80-60) / 40=0.3.

[0247] 2. Execution of the weighted fusion formula:

[0248] 2.1 Input predicted values:

[0249] Predicted temperature at the outlet of the first combustion chamber (From S101, range 900-1050K, error ±12K);

[0250] Predicted temperature at the outlet of the second combustion chamber (From S103, range 890-1030K, error ±14.5K).

[0251] 2.2 Fusion Computing:

[0252] The final predicted value is calculated in real time according to the formula:

[0253]

[0254] For example: when α = 0.3, =976K At 962K: =0.3×976+0.7×962=292.8+673.4=966.2 K.

[0255] 3. Meta-learner online updates:

[0256] 3.1 Data source:

[0257] The historical dataset contains 10,000 groups ( , , )data, The measured temperature of the SCADA system.

[0258] 3.2 Model Training:

[0259] Using linear regression with L2 regularization, the objective function is:

[0260]

[0261] Where λ=1.2, the optimal α=0.32 is obtained by solving (the deviation from the dynamic calculation result is <±0.02).

[0262] 3.3 Update frequency:

[0263] The α calculation rules are updated every 24 hours based on new data increments to ensure that the weighting coefficients adapt to equipment aging.

[0264] 4. Output the final prediction result:

[0265] Will Write the data to the OPC UA server of the gas turbine control system, with the data type being 32-bit floating-point numbers;

[0266] Trigger alarm mechanism: If If the load exceeds the safety threshold of 1050K, the load reduction protection program will be activated.

[0267] Data storage: Record timestamps, The α value is stored in the historical database for performance analysis.

[0268] This step reduces the prediction error variance of the mechanistic model (i.e., the CFD simulation model) and the data-driven model (i.e., the LightGBM prediction model) by 42.7% through dynamic weighting α (0.2-0.4) (experimental data: the error variance of a single model is 186K², and after fusion it is 106K²). The piecewise linear interpolation rule shortens the adjustment response time of α under varying operating conditions to 0.1 seconds, improving the adaptability of operating conditions by 35% compared to the fixed weight strategy. The online-updated ridge regression model controls the long-term prediction drift to ±2.3K / year, and combined with the real-time protection mechanism, extends the life of high-temperature components of the gas turbine by 8-12%. The final prediction results can be directly used for closed-loop control, reducing the combustion chamber outlet temperature fluctuation range from ±25K to ±8K, and improving the combined cycle efficiency by 1.2-1.8%.

[0269] This embodiment provides a gas turbine performance prediction method, focusing on combustion chamber outlet temperature prediction. By integrating mechanistic modeling and data-driven techniques, it effectively improves prediction accuracy and reliability. In terms of mechanistic modeling, a CFD simulation model is constructed based on the precise geometric model and physical structural parameters of the gas turbine combustion chamber. Taking into full account combustion principles and fluid dynamics characteristics, a mechanistic calculation is performed on the combustion chamber outlet temperature to obtain a first predicted value, which theoretically reflects the temperature change pattern. In the data-driven stage, real-time monitoring data of the gas turbine is collected. Using a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm, key features affecting the temperature are accurately extracted and input into the LightGBM prediction model. Leveraging its efficient data processing and learning capabilities, a second predicted value for the combustion chamber outlet temperature is obtained. Finally, the two predicted values ​​are input into a Stacking ensemble model. A meta-learner is used for weighted fusion between models, comprehensively considering mechanistic characteristics and data features, to output the final prediction result, providing strong support for gas turbine performance evaluation.

[0270] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0271] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predicting the efficiency of a gas turbine, characterized in that, The method includes: A CFD simulation model is constructed based on the geometric model and physical structural parameters of the gas turbine combustion chamber. The combustion chamber outlet temperature of the gas turbine is calculated by mechanism to obtain the predicted value of the first combustion chamber outlet temperature. Based on the real-time monitoring data of the gas turbine, key features are extracted using a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm. The key features are input into the LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustion chamber outlet temperature. The predicted values ​​of the first and second combustion chamber outlet temperatures are input into the Stacking ensemble model. The training process of the Stacking ensemble model includes: forming a training set and a test set based on the operating conditions; performing K-fold cross-training CFD simulation analysis to obtain predictions; training the LightGBM model prediction using K-fold cross-training; and training the Stacking meta-learner to train a combination of the CFD simulation model and the LightGBM prediction model. The predicted values ​​of the first and second combustion chamber outlet temperatures are weighted and fused between models using the meta-learner in the Stacking ensemble model, and the final combustion chamber outlet temperature prediction result is output in real time. This includes: dynamically calculating the weight coefficient α of the predicted value of the first combustion chamber outlet temperature based on the real-time operating parameters of the gas turbine; and fusing the predicted values ​​of the first and second combustion chamber outlet temperatures using a weighting formula based on the weight coefficient α.

2. The gas turbine efficiency prediction method according to claim 1, characterized in that, The CFD simulation model is constructed based on the gas turbine's combustion chamber geometric model and physical structural parameters. Mechanistic calculations are performed on the gas turbine's combustion chamber outlet temperature to obtain a predicted value for the first combustion chamber outlet temperature, including: Based on the combustion chamber geometric model and physical structural parameters of the gas turbine, the fuel flow rate, air-fuel mixture ratio, and combustion reaction boundary conditions are set. Based on the fuel flow rate, air-to-fuel ratio, and combustion reaction boundary conditions, the heat released during combustion is obtained through computational fluid dynamics simulation. Based on the ratio of the heat released by combustion to the theoretical heat of the fuel, the predicted value of the outlet temperature of the first combustion chamber is calculated.

3. The gas turbine efficiency prediction method according to claim 1, characterized in that, The real-time monitoring data based on the gas turbine is used to extract key features through a gradient one-sided sampling algorithm and a mutually exclusive feature binding algorithm, including: Gradient calculation is performed on the real-time monitoring data to select the first gradient data sample and the second gradient data sample; Based on the first gradient data sample and the second gradient data sample, a preliminary key feature set is generated; The feature parameters in the preliminary key feature set are mutually exclusive feature bundled to generate the key features.

4. The gas turbine performance prediction method according to claim 3, characterized in that, The step of performing gradient calculation on the real-time monitoring data to select first gradient data samples and second gradient data samples includes: Gradient calculation is performed on the real-time monitoring data, and the first gradient data samples with the top 5%-15% absolute gradient values ​​are retained. Randomly select 10%-30% of the second gradient data samples from the remaining real-time monitoring data; A weight compensation coefficient is introduced for the second gradient data sample. ,in b is the retention ratio of the first gradient samples, and b is the extraction ratio of the second gradient samples.

5. The gas turbine performance prediction method according to claim 3, characterized in that, The step of generating the key features by mutually exclusive feature bundling of feature parameters in the preliminary key feature set includes: Bind multiple pressure sensor data in the aforementioned feature parameters into a pressure feature group; Bind multiple temperature sensor data in the aforementioned feature parameters into a temperature feature group; The pressure feature set and temperature feature set are used as the key features.

6. The gas turbine efficiency prediction method according to claim 1, characterized in that, The step of inputting the key features into the LightGBM prediction model for data-driven calculation to obtain the predicted value of the second combustion chamber outlet temperature includes: The key features are standardized to obtain a standardized feature vector; Based on the standardized feature vector, multiple gradient boosting decision trees are constructed. Each gradient boosting decision tree is split through a leaf node growth strategy, with the splitting rule being to maximize information gain. Based on the prediction results of the gradient boosting decision tree, the predicted value of the second combustion chamber outlet temperature is output by weighted summation within the LightGBM prediction model.

7. The gas turbine efficiency prediction method according to claim 1, characterized in that, The step of performing inter-model weighted fusion of the predicted values ​​of the first and second combustion chamber outlet temperatures using the meta-learner in the Stacking ensemble model to output the final combustion chamber outlet temperature prediction result includes: Based on the real-time operating parameters of the gas turbine, the weighting coefficient α of the predicted value of the first combustion chamber outlet temperature is dynamically calculated, where the value of α ranges from 0.2 to 0.

4. Based on the weighting coefficient α, the predicted values ​​of the first combustion chamber outlet temperature and the second combustion chamber outlet temperature are fused using a weighted formula: ； in, This is the predicted value for the outlet temperature of the first combustion chamber. This is the predicted value for the outlet temperature of the second combustion chamber; The weighted fusion result is output as the final predicted combustion chamber outlet temperature.

Images