Gas turbine heat rate soft measurement method and system based on machine learning model

Abstract

The invention relates to the technical field of gas turbine efficiency, in particular to a gas turbine heat rate soft measurement method and system based on a machine learning model, and the method comprises the steps: collecting the operation parameters of a gas turbine in real time; establishing a heat rate model, and training the heat rate model to obtain a trained heat rate model; and inputting the extracted feature vectors into the trained heat rate model to obtain a heat rate prediction value. According to the method, feature extraction is carried out on the operation parameters, feature vectors with higher correlation with the heat consumption rate can be screened out, interference of redundant information on the model is reduced, and the prediction performance of the model is improved. The heat rate model is based on a machine learning model, common characteristics among different working conditions can be better captured, dependence on specific working conditions is reduced, and the accuracy of a predicted value is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gas turbine efficiency, and particularly relates to a gas turbine heat rate soft measurement method and system based on a machine learning model. BACKGROUND

[0002] As the core equipment of energy power generation, aviation propulsion and industrial driving, the heat rate of a gas turbine is a key indicator for measuring the energy efficiency and economy of a unit. Traditional heat rate measurement mainly relies on direct physical sensor monitoring (such as exhaust temperature sensor, fuel flow meter) and thermodynamic formula calculation, but there are significant defects in actual industrial applications. The mechanism-based thermodynamic model needs to assume steady-state operating conditions, and it is difficult to describe the nonlinear thermodynamic process under dynamic variable conditions (such as rapid load adjustment and environmental disturbance), resulting in accumulated prediction deviation. The traditional method relies on offline data correction and manual experience calibration, which cannot meet the needs of intelligent power plants for real-time monitoring and optimization of energy efficiency.

[0003] In recent years, data-driven soft measurement technology has been gradually applied to heat rate estimation, but existing soft measurement methods generally have weak model interpretability and lack of uncertainty quantification. SUMMARY

[0004] (I) Invention purpose

[0005] The purpose of the present application is to provide a gas turbine heat rate soft measurement method and system based on a machine learning model, which can filter out feature vectors with stronger correlation with heat rate by feature extraction of operating parameters, reduce the interference of redundant information on the model, and better capture the common characteristics between different operating conditions by using a heat rate model based on a machine learning model, reduce the dependence on specific operating conditions, and improve the accuracy of the predicted value.

[0006] (II) Technical solutions

[0007] To solve the above problems, the present application provides a gas turbine heat rate soft measurement method based on a machine learning model, comprising:

[0008] Real-time acquisition of operating parameters of a gas turbine; the operating parameters include temperature parameters, pressure parameters, flow parameters and vibration parameters;

[0009] Feature extraction of the operating parameters to obtain a feature vector, the feature vector including time domain features and frequency domain features, the extraction including time domain feature extraction of the temperature parameters, pressure parameters and flow parameters, and frequency domain feature extraction of the vibration parameters;

[0010] Establishing a heat rate model, training the heat rate model, and obtaining a trained heat rate model;

[0011] Input the extracted features into the trained heat consumption rate model to obtain a heat consumption rate prediction value.

[0012] In another aspect of the present application, preferably, feature extraction is performed on the operating parameters to obtain a feature vector, including:

[0013] The operating parameters are divided according to a preset time window to obtain a plurality of segments of segmented data;

[0014] The instantaneous maximum value of each segment of segmented data is identified respectively, and the instantaneous maximum value is an extreme value of the operating parameter;

[0015] The segmented data is respectively subjected to time domain feature extraction and frequency domain feature extraction;

[0016] The time domain features, the frequency domain features and the extreme value of the operating parameter are used to form a feature vector.

[0017] In another aspect of the present application, preferably, the time domain feature extraction includes:

[0018] Quantile statistics calculation is performed on the operating parameters to obtain quantile statistics;

[0019] A sliding comparison window method is used to calculate the standard deviation and the peak value of the temperature parameter, the pressure parameter and the flow parameter respectively, and a dynamic fluctuation feature sequence is generated;

[0020] Kurtosis index calculation is performed on the pressure parameter to obtain a kurtosis index;

[0021] The quantile statistics, the dynamic fluctuation feature sequence and the kurtosis index are spliced according to the type of the operating parameter to obtain the time domain features of the operating parameter.

[0022] In another aspect of the present application, preferably, the frequency domain feature extraction includes:

[0023] Windowed Fourier transform is performed on the vibration parameter to obtain the frequency domain features of the vibration parameter, wherein a Blackman-Harris window function is used for calculation in the windowed Fourier transform.

[0024] In another aspect of the present application, preferably, the kurtosis index is calculated by using the following formula:

[0025]

[0026] wherein K represents the kurtosis index, N represents the total number of data of the pressure parameter, μ represents the mean value of the pressure parameter, σ represents the standard deviation of the pressure parameter, and xi represents the i th data of the pressure parameter. i

[0027] In another aspect of the present application, preferably, ​

[0028] The heat consumption rate model is based on a machine learning model, and the heat consumption rate model adopts a double-layer stacked integration framework.

[0029] The first layer of the heat consumption rate model is a base learner, which performs preliminary prediction on the input feature vector to generate a base prediction result.

[0030] The second layer of the heat consumption rate model is a meta-learner, which performs nonlinear fusion on the base prediction result to output a heat consumption rate prediction value.

[0031] In another aspect of the present application, preferably, the base learner includes a gradient boosting decision tree and an extreme learning machine.

[0032] The gradient boosting decision tree is used to perform first prediction on the feature vector to obtain a first prediction result.

[0033] The extreme learning machine is used to perform second prediction on the feature vector to obtain a second prediction result.

[0034] The first prediction result and the first prediction result are two columns of base prediction features, and the two columns of base prediction features are the base prediction result.

[0035] The meta-learner is based on a Bayesian regularization neural network.

[0036] In another aspect of the present application, preferably, the training includes model training using an adaptive moment estimation optimizer, and the optimal parameter combination of the heat consumption rate model is determined by maximizing the evidence method.

[0037] In another aspect of the present application, preferably, the method further includes optimizing the feature vector to obtain an optimized feature, inputting the optimized feature into the trained heat consumption rate model, and obtaining a heat consumption rate prediction value.

[0038] Optimizing the feature vector to obtain an optimized feature includes:

[0039] Calculating the covariance matrix of the feature vector.

[0040] Performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues and decomposition feature vectors.

[0041] Selecting a plurality of decomposition feature vectors corresponding to the largest eigenvalues in the front of the order to form a dimension reduction matrix.

[0042] Multiplying the feature vector by the dimension reduction matrix to obtain the optimized feature.

[0043] In another aspect of the present application, preferably, a gas turbine heat consumption rate soft measurement system based on a machine learning model includes:

[0044] The collection module collects the operation parameters of the gas turbine in real time; the operation parameters include temperature parameters, pressure parameters, flow parameters and vibration parameters;

[0045] The extraction module extracts features of the operation parameters to obtain feature vectors, the feature vectors include time domain features and frequency domain features, the extraction includes time domain feature extraction on the temperature parameters, pressure parameters and flow parameters, and frequency domain feature extraction on the vibration parameters;

[0046] The establishment module establishes a heat consumption rate model, trains the heat consumption rate model, and obtains a trained heat consumption rate model;

[0047] The prediction module inputs the extracted features into the trained heat consumption rate model to obtain a heat consumption rate prediction value.

[0048] (Three) beneficial effects

[0049] The above technical solutions of the present application have the following beneficial technical effects:

[0050] The present application can better adapt to the complex working condition changes of the gas turbine by collecting the operation parameters of the gas turbine in real time and extracting the features thereof, compared with the traditional mechanism modeling method, thereby improving the prediction accuracy of the heat consumption rate. Feature extraction is performed on the operation parameters, which can filter out feature vectors with stronger correlation with the heat consumption rate, reduce the interference of redundant information on the model, and further improve the prediction performance of the model. The heat consumption rate model of the present application is based on a machine learning model, which can better capture the common features between different working conditions, reduce the dependence on specific working conditions, and improve the accuracy of the prediction value. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is the overall flowchart of one embodiment of the present application. DETAILED DESCRIPTION

[0052] To make the purpose, technical scheme and advantages of the present application clearer, further detailed description will be made below in combination with specific embodiments and with reference to the drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present application. In addition, in the following description, the description of known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present application.

[0053] Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0054] In the description of the present application, it should be noted that the terms "first", "second", "third" are only for the purpose of description and cannot be understood as indicating or implying relative importance.

[0055] In addition, the technical features involved in the different embodiments of the application described below can be combined with each other as long as there is no conflict.

[0056] Embodiment one

[0057] A gas turbine heat consumption rate soft measurement method based on a machine learning model, Figure 1 The overall flowchart of one embodiment of the present application is shown as Figure 1 As shown, it comprises:

[0058] Real-time acquisition of operating parameters of the gas turbine; the operating parameters include temperature parameters, pressure parameters, flow parameters and vibration parameters, and the acquisition mode is by means of various sensors, such as temperature sensors for measuring inlet and outlet air temperatures, pressure sensors for measuring inlet air pressure, etc. The temperature parameters can include compressor inlet temperature, the temperature sensor is arranged at the compressor inlet to collect the temperature of the ambient air before entering the compressor; combustion chamber outlet temperature, arranged at the connection between the combustion chamber and the turbine; turbine exhaust temperature, collecting the turbine outlet temperature, which is the exhaust gas temperature after the gas does work, reflecting the heat energy utilization efficiency. The pressure parameters are respectively the compressor outlet pressure, the combustion chamber pressure and the turbine inlet pressure; the sensor for collecting the compressor outlet pressure is arranged at the connection between the compressor and the combustion chamber, which is the pressure of the compressed air before entering the combustion chamber, determining the compression work and the combustion efficiency. The sensor for collecting the combustion chamber pressure is arranged in the middle of the combustion chamber, and the sensor for collecting the turbine inlet pressure is arranged in front of the turbine inlet guide vane. The flow parameters can include fuel flow, compressor air flow, cooling air flow, etc., the fuel flow is the amount of fuel input into the combustion chamber per unit time of the fuel supply pipeline, directly determining the energy input; the sensor for collecting the compressor air flow can be arranged at the compressor inlet pipeline to collect the air mass flow entering the compressor; the sensor for collecting the cooling air flow can be arranged in the turbine blade cooling channel. The vibration parameters can include compressor bearing vibration and turbine bearing vibration, the vibration parameters are obtained by accelerometers and displacement sensors, and the vibration sensor is arranged on the compressor rotor bearing; the turbine bearing vibration, the vibration sensor is arranged on the turbine rotor bearing, and the aggravation of vibration will cause the heat consumption rate to rise. In this embodiment, the temperature parameter is the turbine exhaust temperature, the pressure parameter is the turbine inlet pressure, the flow parameter is the fuel flow and the cooling air flow, and the vibration parameter is the turbine bearing vibration.

[0059] The operating parameters are divided according to a preset time window to obtain a plurality of segment data;

[0060] In the division of data, a sliding window is adopted. Assuming that a time window T and a sliding step t are set, starting from the starting time of the sequence of operating parameters, the data is intercepted in a window with a length of T, and the window is slid by t time units after each interception to avoid data omission and obtain characteristic information of different time segments, making the data processing more flexible and capable of adapting to the dynamic changes of the parameters in the operation of the gas turbine. The divided data segments provide a basic unit for subsequent feature recognition.

[0061] The segmented data is respectively subjected to feature extraction; the feature extraction is a process of mining information having a significant influence on the heat rate from the original operating parameters. In this embodiment, the feature extraction includes:

[0062] The instantaneous maximum value of the segmented data is respectively identified, and the instantaneous maximum value is an extreme value of the operating parameter; the instantaneous maximum value (extreme value feature) of the operating parameter is an important characteristic quantity. It reflects the highest value of a parameter in the gas turbine in an instant during the operation. For example, the inlet pressure of the gas turbine may have an instantaneous peak value under certain working conditions, which may be caused by instantaneous disturbance of the inlet system, sudden change of load, etc. The instantaneous maximum value is identified by comparing each segmented data point by point. In the data processing process, a variable is set to store the maximum value found at present, and the initial value can be set as the first value of the data segment. Then, starting from the second data point, the current maximum value is compared with each data point in turn. If the value of the current data point is greater than the stored maximum value, the maximum value variable is updated. After traversing each data segment, the instantaneous maximum value in each data segment is obtained.

[0063] The segmented data is respectively subjected to time domain feature extraction and frequency domain feature extraction; the time domain feature can intuitively reflect the change characteristics of the operating parameter in the time dimension, and the frequency domain feature is obtained by transforming the time domain data and represents the distribution of the operating parameter signal in different frequency components.

[0064] According to the time domain feature, the frequency domain feature and the extreme value of the operating parameter, a feature is formed. The time domain feature, the frequency domain feature and the extreme value of the operating parameter are spliced into a high-dimensional vector to form a feature vector.

[0065] Further, in this embodiment, the time domain feature of the temperature parameter, the pressure parameter and the flow parameter includes:

[0066] The quantile statistics of the temperature parameter, the pressure parameter and the flow parameter are calculated respectively to obtain quantile statistics. The quantile statistics are the 25%, 50% (i.e. the median) and 75% quantiles. The 25% quantile can reflect that 25% of all the temperature data is lower than the value, which can be used to judge the lower temperature level of the gas turbine during operation. The median is in the middle position of the data sequence, which can avoid the interference of extreme values and more stably represent the intermediate level of the temperature, which is of great significance for analyzing the temperature of the gas turbine under normal working conditions. The 75% quantile can be used to understand the distribution of the higher temperature. Through these quantile statistics, the distribution characteristics of the temperature parameter in different intervals during the operation of the gas turbine can be comprehensively grasped, which is helpful for analyzing the variation range of the temperature and the probability of abnormal temperature. For example, if the 75% quantile of the temperature continuously rises in a certain period of time, it may mean that the gas turbine has a temperature rising trend under some working conditions.

[0067] The standard deviation and the peak value of the temperature parameter, the pressure parameter and the flow parameter are calculated respectively by using the sliding comparison window method to generate a dynamic fluctuation feature sequence. The standard deviation and the peak value of the temperature parameter are calculated by using the sliding comparison window method, which can capture the dynamic fluctuation of the temperature in real time. It is assumed that a sliding window size M is set, and the window slides point by point on the temperature data sequence. In each window, the standard deviation of the temperature data is calculated first. The standard deviation reflects the dispersion degree of the temperature data relative to the mean value. The larger the standard deviation, the more intense the temperature fluctuation in the time period. For example, in a period of time after the start of the gas turbine, the temperature fluctuation is large due to the gradual stabilization of the combustion process, and the standard deviation calculated by the sliding window will present a high value. As the operation gradually stabilizes, the standard deviation will gradually decrease. At the same time, the peak value of the temperature is found in each window, which can directly show the highest value of the temperature in the time period. The standard deviation and the peak value calculated in each window are arranged in time sequence to generate a dynamic fluctuation feature sequence of the temperature parameter. This sequence can clearly show the dynamic change characteristics of the temperature during operation.

[0068] The kurtosis index of the pressure parameter is calculated to obtain the kurtosis index. For the pressure parameter, the kurtosis index can reflect the degree of peak of the data distribution. When the gas turbine operates normally, the distribution of the pressure parameter has a certain regularity, and the kurtosis index is in a relatively stable range. For example, in the stable inlet pressure data, the kurtosis index reflects the deviation degree of the distribution compared with the normal distribution. In this embodiment, the kurtosis index is calculated by using the following formula:

[0069]

[0070] wherein K represents the kurtosis index, N represents the total number of data of the pressure parameter, μ is the mean value of the pressure parameter, σ is the standard deviation of the pressure parameter, and x represents the data of the pressure parameter.i The i-th data representing a pressure parameter.

[0071] The quantile statistics, the dynamic fluctuation feature sequence and the kurtosis index are spliced according to the running parameter type to obtain a time domain feature, the quantile statistics can be arranged first, and then the standard deviation and the peak value calculated in each window are connected in turn. For the pressure parameter, in addition to the quantile statistics and the dynamic fluctuation feature sequence, the kurtosis index also needs to be included.

[0072] Further, in the embodiment, the running parameter includes a vibration parameter.

[0073] Identifying the frequency domain feature of the vibration parameter includes:

[0074] Performing windowed Fourier transform on the vibration parameter to obtain the frequency domain feature of the vibration parameter, wherein a Blackman-Harris window function is used for calculation in the windowed Fourier transform. The Blackman-Harris window function is represented by the following formula:

[0075]

[0076] Wherein, r represents the window function number, and R represents the window length. The Blackman-Harris window effectively suppresses the spectral aliasing phenomenon of the wideband vibration signal of the gas turbine.

[0077] A heat consumption rate model is established, the heat consumption rate model is trained to obtain a trained heat consumption rate model, and the extracted features are input into the trained heat consumption rate model to obtain a heat consumption rate prediction value. In the embodiment, the heat consumption rate model is based on a machine learning model, and the heat consumption rate model uses a double-layer stacked integration framework.

[0078] The first layer of the heat consumption rate model is a base learner, the base learner performs preliminary prediction on the input features to generate a base prediction result, and the base learner includes a gradient boosting decision tree and an extreme learning machine.

[0079] The gradient boosting decision tree is used to make a first prediction on the features, and a first prediction result is obtained; the gradient boosting decision tree uses an iterative method to construct a model. First, an initial decision tree is constructed based on the input features to make a preliminary prediction of the heat consumption rate. The decision tree will generate a prediction error, and then a new decision tree is constructed for the error to fit the residual of the previous round of prediction. Each newly constructed decision tree learns from the errors of the previous round of prediction, and through the continuous accumulation of these newly constructed decision trees, the fitting ability for the input features is gradually improved. After multiple iterations, the first prediction result of the input features is finally generated. For example, when facing the complex relationship between steam flow variation and heat consumption rate, the gradient boosting decision tree can construct multiple decision trees to mine the influence of flow variation on heat consumption rate in different intervals, and thus give a more accurate first prediction result.

[0080] The extreme learning machine is used to make a second prediction on the features, and a second prediction result is obtained; the connection weights between the input layer and the hidden layer of the extreme learning machine and the threshold values of the hidden layer neurons are randomly generated and fixed during the training process. After the input features are processed by the random weights and threshold values, the hidden layer output is calculated directly through an activation function, and then the output layer weights are calculated through a least square method to achieve the purpose of fast learning. The input features can be quickly processed to generate the second prediction result. Taking the influence of temperature and pressure on the heat consumption rate as an example, the extreme learning machine can quickly calculate the prediction value related to the heat consumption rate under the randomly set weights and threshold values to form the second prediction result.

[0081] Based on the first prediction result and the second prediction result, two columns of base prediction features are generated, and the two columns of base prediction features are base prediction results; the bias that may occur on a specific data subset by a single model can be effectively avoided, and the reliability and stability of the base prediction result can be enhanced. The base prediction result is the initial heat consumption rate obtained by the two base learners of the gradient boosting decision tree and the extreme learning machine based on the input feature vector, respectively.

[0082] The second layer of the heat consumption rate model is a meta-learner that performs nonlinear fusion on two columns of base prediction features in the base prediction result, and the meta-learner outputs a heat consumption rate prediction value based on a Bayesian regularization neural network. The Bayesian regularization neural network automatically adjusts the size of the weight during the training process to avoid overfitting. For the base prediction result, the Bayesian regularization neural network assigns different weights to each input base prediction feature, and these weights are not fixed but are continuously optimized during the training process according to the Bayesian theory. For example, for base prediction features that perform better in gradient boosting decision trees than in extreme learning machines under different operating conditions, the Bayesian regularization neural network assigns higher weights to them. Through nonlinear transformation and weighted combination of the base prediction result, the Bayesian regularization neural network can dig out deeper information hidden behind the base prediction result, and finally output more accurate and reliable heat consumption rate prediction values. This prediction value combines the advantages of gradient boosting decision trees and extreme learning machines in base learners, overcomes the limitations of single learners, and greatly improves the accuracy and stability of heat consumption rate prediction.

[0083] Further, in the embodiment, the training includes model training using an adaptive matrix estimation optimizer to determine the optimal parameter combination of the heat consumption rate model by maximizing the evidence method.

[0084] Further, the embodiment also includes optimizing the features to obtain optimized features, inputting the optimized features into the trained heat consumption rate model, and obtaining a heat consumption rate prediction value.

[0085] The method for obtaining the optimized features includes:

[0086] The covariance matrix of the features is calculated, the covariance matrix is subjected to eigenvalue decomposition to obtain eigenvalues and eigenvectors, the eigenvectors corresponding to a plurality of largest eigenvalues in the front are selected to form a dimension reduction matrix, and the features are multiplied by the dimension reduction matrix to obtain the optimized features. The original high-dimensional feature data is mapped to a low-dimensional space by selecting the eigenvectors corresponding to a plurality of largest eigenvalues in the front to form a dimension reduction matrix. This can significantly reduce the number of input features of the subsequent heat consumption rate model, reduce the computational complexity in the prediction process, and improve the computational efficiency.

[0087] Compared with the traditional mechanism modeling method, the present application can better adapt to the complex operating condition changes of the gas turbine by collecting the operating parameters of the gas turbine in real time and extracting the features thereof, thereby improving the prediction accuracy of the heat consumption rate. Feature extraction on the operating parameters can screen out feature vectors with stronger correlation with the heat consumption rate, reduce the interference of redundant information on the model, and further improve the prediction performance of the model. The heat consumption rate model of the present application is based on a machine learning model, which can better capture the common features between different operating conditions, reduce the dependence on specific operating conditions, and improve the accuracy of the prediction value.

[0088] Embodiment two

[0089] A gas turbine heat rate soft measurement system based on a machine learning model, comprising:

[0090] A collection module: real-time collection of operating parameters of the gas turbine; the operating parameters include temperature parameters, pressure parameters, flow parameters and vibration parameters;

[0091] An extraction module: feature extraction is performed on the operating parameters to obtain a feature vector, the feature vector includes time domain features and frequency domain features, the extraction includes time domain feature extraction on the temperature parameters, pressure parameters and flow parameters, and frequency domain feature extraction on the vibration parameters;

[0092] An establishment module: a heat rate model is established, the heat rate model is trained, and a trained heat rate model is obtained;

[0093] A prediction module: the extracted features are input into the trained heat rate model to obtain a heat rate prediction value.

[0094] It should be understood that the above specific embodiments of the present application are only used for illustrative or explanatory purposes of the principles of the present application, and do not constitute a limitation on the present application. Therefore, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the present application shall be included in the protection scope of the present application. In addition, the appended claims of the present application are intended to cover all variations and modifications falling within the scope and boundary of the appended claims, or the equivalent forms of such scope and boundary.

[0095] The present application has been described above with reference to the embodiments of the present application. However, these embodiments are only for illustrative purposes, and are not intended to limit the scope of the present application. The scope of the present application is defined by the appended claims and their equivalents. Without departing from the scope of the present application, those skilled in the art can make various substitutions and modifications, which should fall within the scope of the present application.

[0096] Although the embodiments of the present application have been described in detail, it should be understood that various changes, replacements and modifications can be made to the embodiments of the present application without departing from the spirit and scope of the present application.

[0097] Obviously, the above embodiments are only examples for clear illustration, and are not a limitation on the embodiments. Based on the above description, other different forms of changes or modifications can be made by those skilled in the art. Here, it is not necessary or possible to exhaust all the embodiments. The obvious changes or modifications derived therefrom are still within the protection scope of the present application.

Claims

1. A soft measurement method for gas turbine heat rate based on machine learning model, characterized in that: include: Real-time collection of gas turbine operating parameters; the operating parameters include temperature parameters, pressure parameters, flow parameters and vibration parameters; Performing feature extraction on the operating parameters to obtain a feature vector, wherein the feature vector includes a time domain feature and a frequency domain feature, wherein the extraction includes performing time domain feature extraction on the temperature parameter, the pressure parameter, and the flow parameter, and performing frequency domain feature extraction on the vibration parameter; establishing a heat rate model, training the heat rate model, and obtaining a trained heat rate model; The extracted features are input into the trained heat rate model to obtain a heat rate prediction value.

2. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 1, characterized in that: Feature extraction is performed on the operating parameters to obtain a feature vector, including: Dividing the operating parameters according to a preset time window to obtain a plurality of segmented data; Identifying the instantaneous maximum value of each segmented data respectively, wherein the instantaneous maximum value is the extreme value of the operating parameter; Performing time domain feature extraction and frequency domain feature extraction on the segmented data respectively; A feature vector is formed according to the time domain features, frequency domain features and extreme values ​​of the operating parameters.

3. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 2, characterized in that: The time domain feature extraction includes: Calculate quantile statistics for the running parameters to obtain quantile statistics; Using a sliding comparison window method, the standard deviation and peak value of the temperature parameter, pressure parameter and flow parameter are calculated respectively to generate a dynamic fluctuation characteristic sequence; Calculate the kurtosis index of the pressure parameter to obtain the kurtosis index; The quantile statistics, dynamic fluctuation characteristic sequence and kurtosis index are spliced ​​according to the operating parameter type to obtain the time domain characteristics of the operating parameter.

4. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 2, characterized in that: The frequency domain feature extraction includes: Performing a windowed Fourier transform on the vibration parameters to obtain frequency domain features of the vibration parameters, wherein a Blackman-Harris window function is used for calculation in the windowed Fourier transform.

5. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 3, characterized in that: The kurtosis index is calculated using the following formula: Among them, K represents the kurtosis index, N represents the total number of pressure parameter data, μ is the mean of the pressure parameter, σ is the standard deviation of the pressure parameter, and x i Indicates the i-th data of the pressure parameter.

6. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 1, characterized in that: The heat rate model is based on a machine learning model and adopts a double-layer stacked integration framework; The first layer of the heat rate model is a base learner, which performs preliminary prediction on the input feature vector and generates a base prediction result; The second layer of the heat rate model is a meta-learner, which performs nonlinear fusion on the base prediction results and outputs a heat rate prediction value.

7. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 6, characterized in that: The base learners include gradient boosting decision trees and extreme learning machines; Performing a first prediction on the feature vector using the gradient boosting decision tree to obtain a first prediction result; Performing a second prediction on the feature vector by the extreme learning machine to obtain a second prediction result; For the first prediction result and the first prediction result are two columns of basic prediction features, and the two columns of basic prediction features are basic prediction results; The meta-learner is based on a Bayesian regularized neural network.

8. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 1, characterized in that: The training includes adopting an adaptive moment estimation optimizer to perform model training, and determining a parameter combination of an optimal heat rate model by using a maximum evidence method.

9. The method for soft measurement of heat rate of a gas turbine based on a machine learning model according to claim 1, characterized in that: The method further includes optimizing the feature vector to obtain optimized features, inputting the optimized features into a trained heat rate model to obtain a heat rate prediction value; Optimizing the feature vector to obtain optimized features includes: Calculating the covariance matrix of the eigenvectors; Perform eigendecomposition on the covariance matrix to obtain eigenvalues ​​and decomposition eigenvectors; Select the decomposed eigenvectors corresponding to the largest eigenvalues ​​in the top order to form a dimensionality reduction matrix; The feature vector is multiplied by the dimension reduction matrix to obtain optimized features.

10. A gas turbine heat rate soft measurement system based on machine learning model, characterized in that: include: Acquisition module: real-time acquisition of gas turbine operating parameters; the operating parameters include temperature parameters, pressure parameters, flow parameters and vibration parameters; Extraction module: extracting features from the operating parameters to obtain feature vectors, wherein the feature vectors include time domain features and frequency domain features. The extraction includes extracting features from the temperature parameters, pressure parameters, and flow parameters in the time domain, and extracting features from the vibration parameters in the frequency domain. Establishing module: establishing a heat rate model, training the heat rate model, and obtaining a trained heat rate model; Prediction module: inputs the extracted features into the trained heat rate model to obtain the heat rate prediction value.

Images