Mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion

Abstract

The invention provides a mixer outlet temperature prediction method based on a GA-Stacking framework and error compensation fusion. The method comprises the following steps: acquiring parameters of two paths of inlet airflow of a mixer; carrying out data preprocessing on the obtained samples, carrying out dimension reduction on the data by using an ElasticNet algorithm, and retaining input features of which the compression coefficient is not 0; building a mixer outlet temperature prediction model based on a GA-Stacking framework and error compensation; training the mixer outlet temperature prediction model, and predicting a test set sample by using the optimized mixer outlet temperature prediction model; and evaluating the prediction precision and the training effect of the mixer outlet temperature prediction model by comparing the error between the predicted value of the mixer outlet temperature prediction model and the real outlet temperature value. The prediction precision can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of machine learning and industrial process control, and particularly relates to a mixer outlet temperature prediction method based on a GA-Stacking framework and error compensation fusion. Background Art

[0002] The development of aero-engines requires a large number of tests. Conducting high-altitude tests on a high-altitude simulation test bench is one of the effective ways to verify the performance of aero-engines. The prerequisite for conducting high-altitude simulation tests is the accurate simulation of high-altitude environmental conditions. During the process of simulating high-altitude intake environmental conditions, the accurate simulation of the intake temperature is an important link. However, due to reasons such as energy loss, uneven mixing, and slow mixing in the air mixing process, there is a large difference between the actual outlet temperature and the theoretically calculated temperature.

[0003] Currently, the prediction of the outlet temperature of an air mixer usually relies on traditional mechanism modeling methods. These traditional methods are generally based on the principles of fluid mechanics and predict the outlet temperature by modeling physical quantities such as air flow, temperature, and pressure. However, due to the obvious non-linear relationship between various parameters in the intake flow, it is often difficult to handle complex multi-variable problems solely relying on mechanism models. Moreover, the gas mixing process has obvious hysteresis, and the outlet temperature has a strong correlation with time, which makes traditional methods face great limitations in terms of accuracy and real-time performance. Although it is a common practice to modify traditional mechanism models to improve the prediction effect, due to the complexity and non-linearity of the gas mixing process, simple model modification often fails to achieve satisfactory results. Therefore, there is an urgent need to provide a new method for predicting the outlet temperature of a mixer to improve the simulation accuracy of intake conditions. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a mixer outlet temperature prediction method based on a GA-Stacking framework and error compensation fusion, which corrects and predicts the temperature at the outlet of the mixer. The method aims to optimize the Stacking structure through the GA algorithm and combine an error compensation model to improve the prediction accuracy of the mixer outlet temperature. Specifically, the GA algorithm is used to optimize the first-layer base learners of the Stacking structure, adaptively select the optimal combination of base learners to fuse multi-features, and at the same time introduce historical temperature information, establish a physical constraint loss function, and use the GBDT algorithm to learn the temporal characteristics of the temperature change of the mechanism model as a compensation input to ensure that the model can better adapt to multi-dimensional historical data, thereby improving the accuracy of the prediction results.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting the outlet temperature of a mixer based on the GA-Stacking framework and error compensation fusion, comprising the following steps:

[0007] Step 1: Obtain the parameters of the two inlet airflows of the mixer, including the inlet temperature of the normal-temperature air , the inlet pressure , the inlet flow rate , the inlet temperature of the low-temperature air , the inlet pressure , the inlet flow rate ; Obtain the parameters of the outlet airflow of the mixer, including the outlet pressure of the mixer ; Obtain the parameters that may affect the mixing process, including the secondary flow temperature , the pressure , the flow rate , the inlet pressure of the front cabin and the inlet temperature ;

[0008] Step 2: Perform data preprocessing on the obtained samples, use the ElasticNet algorithm to reduce the dimension of the data, and retain the input features with non-zero compression coefficients;

[0009] Step 3: Build a prediction model for the outlet temperature of the mixer based on the GA-Stacking framework and error compensation;

[0010] Step 4: Train the prediction model for the outlet temperature of the mixer, and use the optimized prediction model for the outlet temperature of the mixer to predict the test set samples;

[0011] Step 5: Evaluate the prediction accuracy and training effect of the prediction model for the outlet temperature of the mixer by comparing the error between the predicted value of the prediction model for the outlet temperature of the mixer and the true outlet temperature value.

[0012] Furthermore, the data preprocessing in Step 2 includes outlier processing, data smoothing processing, and normalization processing.

[0013] Furthermore, the method used for screening the input features is the improved ElasticNet method, and by analyzing the influence degree of different parameters in the data on the outlet temperature of the mixer, the variables with high correlation with the outlet temperature of the mixer in the data are retained.

[0014] Furthermore, the expression of the ElasticNet method is:

[0015] ;

[0016] where is the optimal solution of the regression coefficient, represents the target output vector, which is the outlet temperature of the mixer in the present invention, Denote the input feature variables, which are the various features affecting the temperature at the outlet of the mixer. Denote the regression coefficients of the respective feature vectors. Denote the L1 norm of Denote the L2 norm of Denote the penalty coefficient, which is used for the compressive estimation of the compressed regression coefficients. is a special mixing ratio parameter that determines the penalty coefficient for the degree of compression of the feature vectors. Denote the parameter value corresponding to the minimum value of the objective function within the value range of the parameter .

[0017] Furthermore, in the ElasticNet method, the optimal and are determined by using the 10-fold cross-validation method to solve, and the value corresponding to the minimum mean square error of the model is the optimal value.

[0018] Furthermore, the expression for solving the parameter is:

[0019] ;

[0020] where is the mean square error value of the model at the current , is the optimal value corresponding to the minimum mean square error, denotes the parameter value corresponding to the minimum value of the objective function within the value range of the parameter .

[0021] Furthermore, the said step three includes:

[0022] Step (1) Divide the data samples into a training set and a test set. The test set is intercepted and spliced from multiple samples. Use the training set to train the mixer outlet temperature prediction model, and use the test set to test the training effect of the mixer outlet temperature prediction model;

[0023] Step (2) Select multiple base learners to form a candidate set;

[0024] Step (3) Evaluate different combinations of base learners according to the GA algorithm. The fitness function is set to minimize the root mean square error RMSE, and the best combination of base learners selected by GA is used as the first-layer prediction model of Stacking;

[0025] Step (4) uses the prediction results of the first-layer base learners to train the meta-learner, adjusts the parameters of the meta-learner, and uses the optimized meta-learner as the second-layer prediction model of Stacking.

[0026] Further, the outlet temperature prediction in Step 4 includes the following steps:

[0027] Step (1) trains the training set according to the base learners determined by GA optimization. After training is completed, the predicted outlet temperature vectors are concatenated into a new feature matrix.

[0028] Step (2) constructs a temperature mechanism model based on the fluid mixing mechanism and the laws of conservation of energy and mass conservation. The calculated temperature value obtained from the temperature mechanism model and the actual temperature value are used as the inputs of the error compensation model, and the GDBT algorithm is used to train the error model to obtain the temperature error compensation output.

[0029] Step (3) uses the new feature matrix in Step (1) and the error compensation output in Step (2) as the inputs of the second-layer prediction model of Stacking, and uses the BP algorithm of the second layer for training. After training is completed, the test set data is predicted.

[0030] Further, the error evaluation indexes in Step 5 are the mean square error RMSE and the coefficient of determination , and the calculation formulas are as follows:

[0031] ;

[0032] ;

[0033] Among them, represents the number of samples, represents the sample index value, represents the temperature value of the th sample predicted by the model, represents the th actual temperature value of the sample.

[0034] Further, when the mean square error RMSE value is smaller and is closer to 1, the prediction is more accurate.

[0035] Beneficial effects:

[0036] The present invention fully explores the correlation between time-dependent temperature data and multi-dimensional input features, optimizes the base learners of the Stacking framework through the GA algorithm, improves the generalization ability of the model, and at the same time introduces a physical constraint loss function in combination with the flow characteristics of the mixer to make up for the error deviation caused by data training. The time series characteristics of the temperature change of the mixer are integrated in the error compensation, effectively improving the prediction accuracy and achieving accurate prediction of the outlet temperature of the mixer. At the same time, this method considers the influence of different air inlet pressures on the outlet temperature of the mixer. Through input data such as inlet air flow, inlet air temperature, and inlet air pressure, the model can capture the dynamic influence of the change in inlet air pressure on the outlet temperature of the mixer, thereby achieving more accurate temperature prediction. Description of the Drawings

[0037] Figure 1 It is a flowchart of a method for predicting the outlet temperature of a mixer based on the GA-Stacking framework and error compensation fusion according to an embodiment of the present invention;

[0038] Figure 2 It is a flowchart of sample training and prediction of a prediction model;

[0039] Figure 3 It is a temperature prediction curve graph using a GBDT error compensation model;

[0040] Figure 4 It is a curve graph of the prediction of the test set by the outlet temperature prediction model. Detailed Embodiments

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments, and are intended to explain the present invention and should not be construed as a limitation to the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] As Figure 1 shown, a method for predicting the outlet temperature of a mixer based on the GA-Stacking framework and error compensation fusion according to an embodiment of the present invention includes the following steps:

[0043] Step 1: Obtain the parameters of the two-way inlet airflows of the mixer (that is, Figure 1 obtain the parameter data of the inlet airflows of the mixer and the outlet temperature data in , including the inlet air temperature of the normal-temperature air, the inlet air pressure , the inlet air flow , the inlet air temperature of the low-temperature air, the inlet air pressure , intake air flow ; Obtain the parameters of the air flow at the mixer outlet, including the mixer outlet pressure ; Obtain the parameters that may affect the mixing process, including the secondary flow temperature , pressure , flow rate , the intake pressure of the front cabin and the intake temperature ;

[0044] Step 2: Perform data preprocessing on the obtained samples, use the ElasticNet algorithm to perform dimensionality reduction on the data, and retain the input features with non-zero compression coefficients;

[0045] Step 3: Build a mixer outlet temperature prediction model based on the GA-Stacking framework and error compensation (that is, Figure 1 the temperature prediction model of building the GA-Stacking framework in , introduce the error compensation model as the compensation input);

[0046] Step 4: Train the mixer outlet temperature prediction model, and use the optimized mixer outlet temperature prediction model to predict the test set samples (that is, Figure 1 the optimized temperature prediction model in , use the optimized model to predict the prediction set);

[0047] Step 5: Evaluate the prediction accuracy and training effect of the mixer outlet temperature prediction model by comparing the error between the predicted value of the mixer outlet temperature prediction model and the true outlet temperature value (that is, Figure 1 evaluate the training effect according to the error between the model predicted value and the data true value in ).

[0048] Specifically, the data preprocessing in Step 2 includes the following steps:

[0049] First, perform outlier processing and normalization on the original data obtained in Step 1. The outlier processing includes data smoothing processing and invalid data elimination. The normalization processing limits the data to the interval (0,1) to eliminate the scale difference between the data.

[0050] Then use the improved ElasticNet method to analyze the influence degree of different parameters in the data on the mixer outlet temperature, and retain the variables with high correlation with the mixer outlet temperature in the data. The expression of the improved ElasticNet method is:

[0051] ;

[0052] Among them, represents the target output vector, which is the mixer outlet temperature in the present invention, Denote the input feature variables, which are the various features affecting the temperature at the outlet of the mixer in the present invention. Denote the regression coefficients of the respective feature vectors. Is the optimal solution of the regression coefficient. Denote The L1 norm of Denote The L2 norm of Denote the penalty coefficient, which is used for the compressive estimation of the compressed regression coefficient. Is a special mixing ratio parameter that determines The degree of compression of the feature vector. Denote that within the value range of the parameter The corresponding parameter value when the objective function achieves the minimum value;

[0053] The optimal mixing ratio parameter in the dimensionality reduction model And the penalty coefficient Are determined by using the method of 10-fold cross-validation. The data set is divided into 10 subsets. Each time, 9 subsets are selected as the training set for training, and the remaining 1 subset is used as the test set for testing. This is repeated 10 times, with each subset being used as a test set once. The average value of the 10 validation results is taken as the overall performance evaluation of the dimensionality reduction model. The Value corresponding to the minimum mean square error is the optimal value. The expression for solving the parameter Is:

[0054] ;

[0055] Wherein, Is the mean square error value of the model at the current , Is the optimal Value corresponding to the minimum mean square error, Denote that within the value range of the parameter The corresponding parameter value when the objective function achieves the minimum value;

[0056] After dimensionality reduction processing, the penalty coefficients of the respective variables Are shown in Table 1:

[0057] Table 1 Penalty coefficients of each variable

[0058] Specifically, the said step three includes:

[0059] Step (1) Divide the data samples into a training set and a test set. The test set is intercepted and spliced from multiple samples. Use the training set to train the mixer outlet temperature prediction model, and use the test set to test the training effect of the mixer outlet temperature prediction model.

[0060] Step (2) Select multiple base learners to form a candidate set;

[0061] Step (3) Evaluate different combinations of base learners according to the GA algorithm. The fitness function is set to minimize the root mean square error (RMSE). The best combination of base learners selected by GA is used as the first-layer prediction model of Stacking;

[0062] Step (4) Use the prediction results of the first-layer base learners to train the meta-learner, adjust the parameters of the meta-learner, and use the optimized meta-learner as the second-layer prediction model of Stacking.

[0063] Specifically, step four includes the following steps:

[0064] Step (1) Use the first-layer prediction model of the optimized Stacking framework to train the training set. After training, splice the predicted outlet temperature vectors into a new feature matrix;

[0065] Step (2) Based on the fluid mixing mechanism, construct a temperature mechanism model according to the laws of energy conservation and mass conservation. Use the temperature calculated value obtained from the temperature mechanism model and the actual temperature value as the input of the error compensation model, and use the GDBT algorithm to train the error model to obtain the temperature error compensation output;

[0066] Step (3) Use the new feature matrix in step (1) and the error compensation output in step (2) as the input of the second-layer prediction model of Stacking, and use the BP algorithm of the second layer for training. After training, predict the test set data.

[0067] Specifically, the error evaluation indexes in step five are the root mean square error (RMSE) and the coefficient of determination , and the calculation formulas are as follows:

[0068] ;

[0069] ;

[0070] Among them, represents the number of samples, represents the sample index value, represents the temperature value of the th sample predicted by the model, represents the th actual temperature value of the sample.

[0071] When the RMSE value is smaller and is closer to 1, the prediction is more accurate.

[0072] Example:

[0073] The mixer outlet temperature prediction method based on the GA-Stacking framework and error compensation fusion in the embodiments of the present invention includes:

[0074] In this embodiment, multiple data files collected after testing different models of aero-engines on the same high-altitude test bench are selected to obtain the parameter data of the inlet air flow of the mixer and the outlet temperature data; the original data is preprocessed, including data smoothing processing, outlier processing, etc., and the improved ElasticNet method is used to perform dimensionality reduction on the data to screen out the features with high correlation with the outlet temperature; a mixer outlet temperature prediction model based on the GA-Stacking framework is built, the GA algorithm is used to find the optimal combination of base learners, and an error compensation model is introduced as the compensation input; the training data is put into the model for training, and the optimized model is used to predict the prediction set; the training effect is evaluated according to the error between the model prediction value and the true value of the data.

[0075] As Figure 2 shown, the prediction model sample training and prediction in this embodiment include the following steps:

[0076] (1) Divide the data samples into a training set and a test set. Among them, 345,000 groups of data samples are selected from four data files for the training set, and 5,000 data are intercepted from each of the four samples for the test set, with a total of 20,000 data. The training set is used to train the temperature prediction model, and the test set is used to test the training effect of the temperature prediction model;

[0077] (2) Initialize the GA algorithm, randomly initialize the combination of base learners, set the fitness function, perform Stacking training using the combination of base learners selected by GA, calculate the fitness to evaluate the pros and cons of the model after training is completed, perform genetic operations when selecting the combination of base learners, judge whether the maximum number of iterations or error convergence is reached. If satisfied, output the optimal combination of base learners as the first-layer prediction model of the Stacking framework, otherwise return to the training step to continue optimization. According to the optimal combination of base learners, train the training set. After training is completed, splice the predicted outlet temperature vectors into a new feature matrix as the input of the second-layer prediction model. The number of input samples is 345,000, and the input feature is the number of base learners;

[0078] (3) Use the temperature mechanism model to calculate the mixer outlet temperature value, establish an error data set between the temperature simulation value and the actual temperature value of the mechanism model, and use the simulation value and the error data set as the training samples of the GDBT model to train the error model to obtain the temperature error compensation input;

[0079] (4) Use the new feature matrix in step (2) and the error compensation output in step (3) as the input of the second-layer prediction model of Stacking, and use the BP algorithm of the second layer for training. After the training is completed, predict the test set data. After the prediction is completed, calculate the obtained mean square error RMSE and coefficient of determination as shown in Table 2. The prediction results using the GBDT error compensation model are as Figure 3 shown, where the abscissa is the sample number n, the ordinate is the temperature T, the solid line is the actual outlet temperature value of the sample, the dashed line is the temperature value based on the fluid mixing mechanism, and the dotted line is the temperature prediction value based on the GBDT error compensation. The comparison curve between the prediction result and the actual result is as Figure 4 shown, where the abscissa is the sample number n, the ordinate is the temperature T, the dotted line is the outlet temperature value of the mixer predicted using a single RF model, and the dotted line is the outlet temperature value of the mixer predicted using the GA-Stacking framework.

[0080] The evaluation indicators used are the mean square error RMSE and the coefficient of determination , and the calculation formula is as shown above.

[0081] Table 2 Evaluation indicators

[0082] Figure 3 For the prediction results using GBDT error compensation, it can be seen that introducing the time series characteristics of temperature changes effectively predicts the temperature error information and effectively reduces the error. Figure 4 For the temperature prediction curve of the test set by the prediction model, the prediction effects of the hybrid model based on the GA-Stacking framework and the single RF model are compared. The data simulated by the model comes from four different sample files. Compared with the single RF algorithm, the hybrid model based on the GA-Stacking framework can accurately predict the temperature change trend under different intake temperatures and different intake flows, better fit the temperature change trend, and there is no overfitting phenomenon. The simulation error of most data does not exceed 0.6K, and the calculated RMSE is 0.26%, and is 0.9812. Compared with the single RF algorithm, the RMSE is reduced by 0.25%,

Claims

1. A mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion, characterized in that: The following steps are involved: Step 1: Obtain the parameters of the two inlet air flows of the mixer, including the room temperature air inlet temperature , Intake pressure , intake air flow , low temperature air intake temperature , Intake pressure , intake air flow ; Get the parameters of the mixer outlet air flow, including the mixer outlet pressure ; Obtain parameters that may affect the blending process, including secondary stream temperature ,pressure ,flow , front cabin intake pressure and intake air temperature ; Step 2: Perform data preprocessing on the acquired samples, use the ElasticNet algorithm to reduce the data dimension, and retain the input features whose compression coefficient is not 0; Step 3: Build a mixer outlet temperature prediction model based on GA-Stacking framework and error compensation; Step 4: Train the mixer outlet temperature prediction model, and use the optimized mixer outlet temperature prediction model to predict the test set samples; Step 5: By comparing the error between the predicted value of the mixer outlet temperature prediction model and the actual outlet temperature value, the prediction accuracy and training effect of the mixer outlet temperature prediction model are evaluated.

2. A mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 1, characterized in that: The data preprocessing in step 2 includes outlier processing, data smoothing processing and normalization processing.

3. A mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 2, characterized in that: The method used for screening input features is an improved ElasticNet method, which analyzes the influence of different parameters in the data on the mixer outlet temperature and retains variables with a high correlation with the mixer outlet temperature in the data.

4. A mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 3, characterized in that: The expression of the ElasticNet method is: ； in, is the optimal solution of the regression coefficient, represents the target output vector, which is the mixer outlet temperature in this invention, Represents the input characteristic variables, which are the various characteristics that affect the mixer outlet temperature. represents the regression coefficient of each eigenvector, express The L1 norm of express The L2 norm of represents the penalty coefficient, which is used to estimate the compressed regression coefficient. is a special mixing ratio parameter that determines the penalty coefficient The degree of compression of the feature vector, Indicated in the parameter The parameter value that corresponds to the minimum value of the objective function is within the value range of .

5. The mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 2 is characterized in that: The best of the ElasticNet methods described and The determination is solved using the 10-fold cross-validation method. The model with the minimum mean square error corresponds to The value is the optimal value.

6. A mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 5, characterized in that: parameter The expression to be solved is: ； in, For the current The mean square error value of the model is: is the optimal value corresponding to the minimum mean square error value, Indicated in the parameter The parameter value that corresponds to the minimum value of the objective function is within the value range of .

7. The mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 5 is characterized in that: The step three comprises: Step (1) dividing the data samples into a training set and a test set, wherein the test set is formed by cutting a segment from multiple samples and splicing them together, using the training set to train the mixer outlet temperature prediction model, and the test set to test the training effect of the mixer outlet temperature prediction model; Step (2) select multiple base learners to form a candidate set; Step (3) Evaluate different base learner combinations according to the GA algorithm, set the fitness function to minimize the mean square error RMSE, and use the best base learner combination selected by GA as the first-layer prediction model of Stacking; Step (4) uses the prediction results of the first-layer base learner to train the meta-learner, adjusts the meta-learner parameters, and uses the optimized meta-learner as the second-layer prediction model of Stacking.

8. The mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 1 is characterized in that: The outlet temperature prediction of step 4 comprises the following steps: Step (1) training the training set according to the base learner determined by GA optimization, and after the training is completed, concatenating the predicted outlet temperature vector into a new feature matrix; Step (2) Based on the fluid mixing mechanism, a temperature mechanism model is constructed according to the laws of conservation of energy and conservation of mass. The temperature calculation value and the actual temperature value obtained by the temperature mechanism model are used as inputs of the error compensation model. The error model is trained using the GDBT algorithm to obtain a temperature error compensation output. Step (3) uses the new feature matrix of step (1) and the error compensation output of step (2) as the input of the second-layer prediction model of Stacking, uses the second-layer BP algorithm for training, and predicts the test set data after the training is completed.

9. The mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 1, characterized in that: The error evaluation indicators of step 5 are the mean square error RMSE and the determination coefficient , the calculation formula is as follows: ； ； in, represents the number of samples, Represents the sample index value, The model predicts the The temperature value of each sample, Indicates The actual temperature value of each sample.

10. The mixer outlet temperature prediction method based on GA-Stacking framework and error compensation fusion according to claim 5, characterized in that: When the mean square error RMSE value is smaller, The closer it is to 1, the more accurate the prediction.

Images