Transformer temperature rise evaluation method and system based on Stacking integrated learning framework

Abstract

The invention discloses a transformer temperature rise evaluation method and system based on a Stacking integrated learning framework, and the method comprises the steps: constructing a sample data set based on input variables including the size of an oil baffle plate, the width of an oil duct, the number of oil baffle plates and boundary temperature and response variables including the temperature rise of winding wire oil and the temperature rise of a hot spot through employing a Latin hypercube sampling principle and CFD simulation; constructing a Stacking integrated temperature rise model fusing the multi-model information in a layered manner, independently training a first layer by using the sample data set to generate a plurality of prediction results, and training a second layer after splicing the prediction results to obtain the Stacking integrated temperature rise model fusing the multi-model information; and after hyper-parameter joint optimization is carried out to determine hyper-parameters, a machine learning library is called to automatically complete training of the Stacking integrated temperature rise model fused with multi-model information, a trained temperature rise prediction model is obtained, temperature rise prediction is carried out on a to-be-predicted transformer configured by a new oil baffle plate and an oil duct structure, and predicted values of output line oil temperature rise and hot spot temperature rise values are obtained.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power transformer thermal performance analysis and artificial intelligence agent modeling, and particularly relates to a transformer temperature rise evaluation method and system based on a Stacking integrated learning framework. BACKGROUND

[0002] During the operation of a power transformer, the windings and the core thereof will generate heat due to the load and short-time impact. If the temperature rise is not properly controlled, it will accelerate the insulation aging, shorten the service life, and even cause serious faults such as thermal breakdown. Therefore, accurately predicting the hotspot temperature and the line oil temperature rise under different structural configurations of the transformer in the design stage is of great significance for ensuring the thermal stability of the equipment.

[0003] At present, the CFD (Computational Fluid Dynamics) simulation method is generally used in engineering, such as establishing an oil flow-thermal coupling model based on the ANSYS FLUENT platform to simulate the influence of parameter changes such as different oil channel sizes and oil baffle arrangements on thermal distribution. This method has complete physical mechanisms and high prediction accuracy, and is suitable for high-fidelity temperature rise analysis of a single scheme. However, it has the following significant shortcomings in the design space of multiple schemes and multiple structural combinations: (1) high time cost: each set of parameter combinations needs to build a complete geometric model and perform iterative solution, and a single simulation often takes several to tens of hours; (2) high labor cost: complex fluid domain modeling and boundary condition setting are highly dependent on artificial experience, and the simulation process cannot be automated; (3) no parallel evaluation capability: when facing a large number of structural configuration combinations (such as dozens of oil baffle + oil channel schemes), it is impossible to realize rapid batch analysis, which seriously restricts the design optimization efficiency.

[0004] The existing technology cannot realize efficient batch temperature rise prediction and design parameter sensitivity analysis when facing a large number of oil channel and oil baffle structural configuration combinations, which limits the rapid evaluation of thermal performance and the promotion of the optimization design process in engineering. Therefore, there is an urgent need for a temperature rise prediction method combining artificial intelligence and physical modeling to replace part of the high-cost simulation and realize rapid, multi-batch, high-dimensional structural combination temperature rise evaluation under the premise of controllable precision. SUMMARY

[0005] The purpose of the present application is to overcome the problems of long modeling time, high artificial dependence, and difficulty in supporting parallel evaluation of multiple structural combinations when relying on high-fidelity CFD simulation for transformer temperature rise evaluation, and to propose a transformer temperature rise evaluation method and system based on a Stacking integrated learning framework.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a transformer temperature rise evaluation method based on a Stacking integrated learning framework, comprising the following steps: For a transformer with a known oil baffle and oil duct structure configuration, a Latin hypercube sampling principle and CFD simulation are used to construct a sample data set based on input variables and response variables; the input variables include the size of the oil baffle, the width of the oil duct, the number of the oil baffles and the boundary temperature, and the response variables include the winding wire oil temperature rise and the hot spot temperature rise; A Stacking integrated temperature rise model fusing multi-model information is constructed in layers, including a first layer of basic learners and a second layer of meta-learners; the sample data set is used to independently train the first layer of basic learners to generate a plurality of prediction results, the prediction results are spliced, and the second layer of meta-learners is trained to obtain the Stacking integrated temperature rise model fusing multi-model information; After the Stacking integrated temperature rise model fusing multi-model information is subjected to joint optimization of hyperparameters to determine the hyperparameters, a machine learning library is called to automatically complete the training of the Stacking integrated temperature rise model fusing multi-model information to obtain a trained temperature rise prediction model; The trained temperature rise prediction model is used to predict the temperature rise of a new oil baffle and oil duct structure configuration of a transformer to be predicted to obtain output wire oil temperature rise and hot spot temperature rise prediction values.

[0007] Further, a Latin hypercube sampling is used to generate a plurality of representative parameter combinations in the input variable space to obtain an input matrix; Each group of parameters is subjected to thermodynamic calculation temperature rise output through CFD simulation to obtain an output matrix; The input matrix and the output matrix are combined to obtain a sample data set.

[0008] Further, the first layer of basic learners includes a linear regression model, a random forest regressor and a K nearest neighbor regressor; the sample data set is input into the linear regression model, the random forest regressor and the K nearest neighbor regressor for independent training, respectively, to obtain linear regression prediction values, random forest regression prediction values and K nearest neighbor regression prediction values; the linear regression model introduces a regularization coefficient; The linear regression prediction values, the random forest regression prediction values and the K nearest neighbor regression prediction values are spliced into a new feature matrix; The new feature matrix and the original true temperature rise label are input into the second layer of meta-learners for training to obtain the Stacking integrated temperature rise model fusing multi-model information; the second layer of meta-learners adopts a support vector regressor, and the kernel function of the support vector regressor adopts a radial basis function.

[0009] Further, the optimization method of joint optimization of hyperparameters includes grid search and K-fold cross-validation; The hyperparameters include the number of standard decision trees and the maximum depth of the random forest regressor, the number of neighbors and the weight function of the K- nearest neighbor regressor, the penalty parameter and the kernel function parameter of the support vector regressor, the depth of the decision tree of the second layer meta-learner, and the K value of K-fold cross-validation.

[0010] Further, the sample dataset is shown as follows:

[0011] wherein, represents the input vector of the i-th sample, represents the response variable of the i-th sample; The optimization expression of the linear regression model is shown as follows:

[0012] wherein, represents the linear regression weight coefficient matrix, represents the regularization coefficient; The optimization objective of the random forest regressor is shown as follows:

[0013]

[0014] wherein, and represent the decision tree leaf node weight, and represent the split factor coefficient; The decision tree output model of the random forest regressor is shown as follows:

[0015] wherein, represents the number of leaf splits, represents the split set; The optimization expression of the support vector machine model is shown as follows:

[0016] wherein, represents the number of training samples, represents the normal vector of the linear hyperplane, represents the bias item of the linear hyperplane, represents the regularization parameter, represents the slack variable, represents the kernel function; The types of the kernel function include a linear kernel, a polynomial kernel, or an RBF kernel.

[0017] Further, the optimization objective of the hyperparameter joint optimization is to minimize the root mean square error, and the root mean square error is as follows:

[0018] wherein, represents the output vector of the trained machine learning model under the equivalent sample, represents the true output value of the sample.

[0019] Further, the machine learning library adopts the Scikit-learn machine learning library in python3.11.4.

[0020] In a second aspect, the present application provides a transformer temperature rise evaluation system based on a Stacking ensemble learning framework, comprising: The data set construction module is used for constructing a sample data set based on input variables and response variables by using Latin hypercube sampling principle and CFD simulation for the transformer with known oil baffle and oil duct structure configuration; the input variables include oil baffle size, oil duct width, oil baffle number and boundary temperature, and the response variables include winding wire oil temperature rise and hot spot temperature rise; The Stacking ensemble modeling module is used for hierarchically constructing a Stacking ensemble temperature rise model fusing multiple model information, including a first layer of basic learners and a second layer of meta-learners; the sample data set is used to independently train the first layer of basic learners to generate a plurality of prediction results, the prediction results are spliced to train the second layer of meta-learners, and the Stacking ensemble temperature rise model fusing multiple model information is obtained; The trained temperature rise prediction model module is used for training the Stacking ensemble temperature rise model fusing multiple model information after determining the hyperparameters by hyperparameter joint optimization, calling the machine learning library to automatically complete the training of the Stacking ensemble temperature rise model fusing multiple model information, and obtaining the trained temperature rise prediction model; The temperature rise prediction module is used for predicting the temperature rise of the transformer to be predicted with new oil baffle and oil duct structure configuration by using the trained temperature rise prediction model, and obtaining the output wire oil temperature rise and hot spot temperature rise prediction value; The prediction visualization module is used for receiving the parameters of the transformer to be predicted with new oil baffle and oil duct structure configuration, outputting the winding hot spot temperature rise and wire oil temperature rise prediction value, and visualizing the temperature rise trend chart and error distribution chart results.

[0021] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a transformer temperature rise evaluation method based on a Stacking ensemble learning framework when executing the computer program.

[0022] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program, when executed by a processor, implements a transformer temperature rise evaluation method based on a stacking ensemble learning framework.

[0023] Compared with the prior art, the present application has the following beneficial technical effects: The transformer temperature rise evaluation method based on the stacking ensemble learning framework, As a multi-model fusion machine learning strategy, the stacking ensemble learning has good nonlinear modeling capability and prediction stability, and is particularly suitable for proxy modeling of complex thermal field problems. By introducing simulation results as training samples and using multiple base models and meta-learners to construct a stacking framework, the winding temperature rise and hot spot temperature rise under different structural parameters of the transformer can be accurately predicted while greatly reducing the modeling cost. The present application reduces the dependence on high-cost CFD simulation while maintaining a certain prediction accuracy, realizes rapid prediction and auxiliary evaluation of winding temperature rise under various structural parameter configurations of the transformer, and thus improves the efficiency and applicability of thermal performance analysis in the design stage. The present application solves the problems of long modeling time, high dependence on manual work, and difficulty in supporting parallel evaluation of multiple structural combinations in the prior art of transformer temperature rise evaluation relying on high-fidelity CFD simulation. The transformer oil-line temperature rise evaluation of the present application combining FLUENT simulation results and the stacking ensemble learning algorithm is particularly suitable for rapid thermal field prediction and design auxiliary optimization under different oil duct and oil baffle arrangements. BRIEF DESCRIPTION OF DRAWINGS

[0024] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the present application in any way. In addition, the shapes and scale of the components in the drawings are only illustrative and are used to help understand the present application, and are not specific limitations on the shapes and scale of the components. In the drawings: Figure 1 A flowchart of the transformer temperature rise evaluation method based on the stacking ensemble learning framework of the present application.

[0025] Figure 2 A structural diagram of the transformer temperature rise evaluation system based on the stacking ensemble learning framework of the present application.

[0026] Figure 3 An electronic device diagram of the transformer temperature rise evaluation method based on the stacking ensemble learning framework of the present application.

[0027] Figure 4 A physical model under a certain data sample in an embodiment of the present application.

[0028] Figure 5 The calculation result under a certain data sample in an embodiment of the present application.

[0029] Figure 6 A 5-fold cross-validation schematic diagram in an embodiment of the present application.

[0030] Figure 7 A simple schematic diagram of an ensemble learning framework in an embodiment of the present application.

[0031] Figure 8 A flowchart of a transformer temperature rise evaluation method based on a Stacking ensemble learning framework in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a 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 skilled in the art without creative labor should fall within the scope of protection of the present application.

[0033] Embodiment one Referring to Figure 1 A transformer temperature rise evaluation method based on a Stacking ensemble learning framework includes the following steps: For a transformer with known oil baffle and oil duct structure configuration, Latin hypercube sampling principle and CFD simulation are used to construct a sample data set based on input variables and response variables. The input variables include oil baffle size, oil duct width, number of oil baffles, and boundary temperature. The response variables include winding wire oil temperature rise and hot spot temperature rise. The Latin hypercube sampling method can make the samples uniformly distributed in the input variable space. Compared with simple random sampling, it can more comprehensively cover the value range of the input variables with fewer samples, improving the representativeness of the samples. For example, when studying the influence of input variables such as oil baffle size and oil duct width on transformer temperature rise, Latin hypercube sampling can ensure that different combinations of size and width are reasonably sampled, avoiding excessive concentration or absence of samples in certain areas. CFD simulation can accurately simulate the distribution of oil flow and temperature field inside the transformer, providing accurate response variable (winding wire oil temperature rise and hot spot temperature rise) data for constructing the sample data set. Through simulation, detailed temperature rise information of the transformer under different oil baffle and oil duct structure configurations can be obtained, avoiding measurement errors and high costs that may exist in actual experiments.

[0034] The Stacking integrated temperature rise model fusing multi-model information is constructed in layers, including a first layer of basic learners and a second layer of meta-learners. The first layer of basic learners is independently trained using a sample data set to generate a plurality of prediction results. The second layer of meta-learners is trained after the prediction results are spliced to obtain the Stacking integrated temperature rise model fusing multi-model information. The Stacking integrated learning framework fuses the prediction results of multiple basic learners by constructing the basic learners and the meta-learners in layers. Different basic learners have different characteristics and advantages, and their advantages can be combined through the Stacking framework to improve the generalization ability and prediction accuracy of the model. The layered training method can fully utilize the information of different learners and avoid the limitations of a single learner.

[0035] After the Stacking integrated temperature rise model fusing multi-model information is optimized for hyperparameters, the machine learning library is called to automatically train the Stacking integrated temperature rise model fusing multi-model information to obtain the trained temperature rise prediction model. Hyperparameters have an important impact on the performance of the model. By jointly optimizing the hyperparameters, the most suitable combination of hyperparameters for the transformer temperature rise evaluation problem can be found. Calling the machine learning library can save a lot of time and labor costs. Machine learning libraries usually provide rich optimization algorithms and tools, which can efficiently complete the hyperparameter optimization and model training process.

[0036] The trained temperature rise prediction model is used to predict the temperature rise of a transformer with a new baffle plate and oil duct structure configuration to obtain output line oil temperature rise and hot spot temperature rise prediction values. Since the model is constructed based on a large amount of sample data and multiple learners, it has strong generalization ability and can adapt to transformers with different structural configurations. The prediction values can help engineers evaluate the thermal performance of the transformer and promptly identify potential thermal problems to take appropriate measures for optimization and improvement.

[0037] The transformer temperature rise evaluation method based on the Stacking integrated learning framework has significant advantages in data construction, model construction, model optimization, and prediction application, and can provide more accurate and reliable temperature rise evaluation and prediction for the design and operation of transformers.

[0038] Latin hypercube sampling is used to generate a plurality of representative parameter combinations in the input variable space to obtain an input matrix. Latin hypercube sampling ensures uniform distribution of input variables (such as baffle plate size, oil duct width, etc.) within the value range through stratified sampling, avoiding the local area sample missing or over-concentration problem that may be caused by traditional random sampling. Uniformly distributed samples can more comprehensively reflect the complex relationship between input variables and output variables (such as temperature rise), providing a more reliable data basis for subsequent model training.

[0039] The output matrix is obtained by CFD simulation and thermodynamic calculation of temperature rise output for each set of parameters. CFD simulation accurately calculates the temperature rise output under each set of parameters, including winding wire oil temperature rise and hot spot temperature rise, avoiding data deviation caused by measurement error or environmental interference in actual experiments. CFD can handle complex geometric structures and boundary conditions, providing detailed and accurate physical information for temperature rise evaluation. Compared with actual experiments, CFD simulation has lower cost and shorter cycle, and is especially suitable for scenarios requiring a large number of parameter combinations.

[0040] The sample data set is obtained by combining the input matrix and the output matrix. The direct correspondence between the input matrix (parameter combination) and the output matrix (temperature rise result) ensures the integrity and logical consistency of the data set. The complete data set can provide rich training samples for the Stacking integrated model, improving the model's ability to capture nonlinear relationships. The combination of Latin hypercube sampling and CFD simulation ensures that the data set covers a variety of combinations of input variables, making the model have stronger generalization ability when facing new parameters.

[0041] By generating a large number of representative samples, engineers can quickly evaluate the impact of different parameter combinations on temperature rise, optimize design parameters (baffle size, oil duct width), and improve transformer performance. The data set can reveal potential thermal problems (high hot spot temperature rise) and provide a basis for design improvement, reducing product failure risk. Uniformly distributed input samples and accurate CFD output ensure the quality of training data, avoiding model overfitting or underfitting due to data bias. Not only does it improve the coverage and representativeness of data in the input space, but it also ensures the reliability of the output results through high-precision simulation, providing a solid foundation for subsequent model training and engineering optimization. In complex engineering problems such as transformer temperature rise evaluation, this method can significantly improve design efficiency, reduce cost, and improve product reliability and performance.

[0042] The first layer base learner includes a linear regression model, a random forest regressor, and a K-nearest neighbor regressor; the sample data set is input into the linear regression model, the random forest regressor, and the K-nearest neighbor regressor for independent training, respectively, to obtain linear regression prediction values, random forest regression prediction values, and K-nearest neighbor regression prediction values; the linear regression model introduces a regularization coefficient; the linear regression model has a simple structure and strong interpretability, and can capture the linear relationship between input variables and output variables. The introduction of the regularization coefficient can prevent model overfitting and improve the generalization ability of the model. The random forest regressor is an ensemble learning method that improves the accuracy and stability of the model by constructing multiple decision trees and combining their prediction results. It can handle high-dimensional data and nonlinear relationships, and has strong robustness to outliers and noise. The K-nearest neighbor regressor is an instance-based learning method that finds the K nearest training samples to the predicted sample, and predicts the output value of the predicted sample by weighted average of the output values of these samples. The K-nearest neighbor regressor does not need explicit modeling of data, can handle complex nonlinear relationships, and has no strict requirements for data distribution.

[0043] The linear regression prediction values, random forest regression prediction values, and K-nearest neighbor regression prediction values are spliced into a new feature matrix; the information of different base learners can be fused. Each base learner has its unique advantages and limitations, and by fusing their prediction results, the information of each base learner can be fully utilized to improve the prediction accuracy and generalization ability of the model. Each feature in the new feature matrix is based on the understanding and prediction of the original data by different models, and this multi-perspective feature representation can enhance the model's ability to express data.

[0044] The new feature matrix and the original true temperature rise label are input into the second layer meta-learner for training to obtain a Stacking integrated temperature rise model that fuses multiple model information, and the second layer meta-learner uses a support vector regressor with a radial basis function kernel. The support vector regressor is a powerful machine learning algorithm that finds an optimal hyperplane to fit the training data and has strong robustness to outliers. The radial basis function kernel can map input variables to high-dimensional space, enabling the support vector regressor to handle complex nonlinear relationships. Compared with other meta-learners, the support vector regressor has better generalization ability and interpretability, and is less sensitive to hyperparameters. In addition, the selection of the radial basis function kernel enables the support vector regressor to have good performance in handling nonlinear problems.

[0045] Through the complementarity of the first layer of base learners and the optimization of the second layer of meta-learners, the Stacking ensemble temperature rise model can fully utilize the advantages of different models and improve the prediction accuracy. In practical applications, this model can more accurately predict the temperature rise of transformers, providing more reliable basis for the design and operation of transformers. The Stacking ensemble learning framework reduces the dependence of a single model on a specific data distribution by combining the prediction results of multiple models, improving the model's generalization ability. This means that the model can still maintain good prediction performance when faced with new oil baffle and oil duct structure configurations. Due to the different sensitivities of different base learners to outliers and noise, by combining their prediction results, the Stacking ensemble temperature rise model can reduce the impact of outliers and noise on the prediction results, improving the model's robustness. Based on the Stacking temperature rise prediction of multiple model integration, through the complementarity of the first layer of base learners, the construction of new feature matrices and information fusion, the selection and optimization of the second layer of meta-learners, and the performance improvement of the overall model, it has significant advantages in transformer temperature rise evaluation, and can provide more accurate and reliable temperature rise prediction results.

[0046] The optimization method of hyperparameter joint optimization includes grid search and K-fold cross-validation; grid search will traverse all possible cases of given parameter combinations to ensure that no possible optimal parameter combination is missed. Through comprehensive search, grid search can more accurately find the hyperparameter values that make the model perform best on the training set and validation set, thereby improving the model's prediction accuracy. K-fold cross-validation divides the dataset into K subsets, and each subset is used as the validation set in turn, with the remaining K-1 subsets as the training set. K times of training and validation are performed, and the average performance is taken as the evaluation index of the model. This method can make full use of data and reduce the evaluation bias caused by unreasonable data division, and more accurately evaluate the model's generalization ability. Through multiple training and validation, K-fold cross-validation can detect whether the model has overfitting phenomenon for specific data division.

[0047] The hyperparameters include the number of standard decision trees and the maximum depth of the random forest regressor, the number of neighbors and the weight function of the K-nearest neighbors regressor, the penalty parameter and the kernel function parameter of the support vector regressor, and the decision tree depth and the K value of K-fold cross-validation of the second layer meta-learner. Optimizing the number of standard decision trees and the maximum depth can balance the complexity and generalization ability of the model. Too many decision trees may lead to overfitting, while too few may underfit; too large a maximum depth will make the model too complex and prone to overfitting, while too small a maximum depth may not capture the complex relationships in the data. By jointly optimizing these two hyperparameters, the random forest regressor can have better generalization ability in transformer temperature rise evaluation. The choice of the number of neighbors and the weight function will affect the prediction performance of the model. Too many neighbors may lead to overly smooth prediction results, losing the ability to capture local features; too few neighbors may be sensitive to noise. The choice of the weight function (such as uniform weight and distance weight) will also affect the contribution of different neighbors to the prediction result. By optimizing these two hyperparameters, the generalization ability of the K-nearest neighbors regressor in transformer temperature rise evaluation can be improved. The optimization of the penalty parameter and the kernel function parameter is crucial for the performance of the support vector regressor. The penalty parameter is used to control the complexity of the model and the tolerance to errors, while the kernel function parameter affects the mapping of data in high-dimensional space. By jointly optimizing these two hyperparameters, the support vector regressor can better handle the nonlinear relationships in transformer temperature rise evaluation and improve its generalization ability. The optimization of the decision tree depth and the K value of K-fold cross-validation can further improve the generalization ability of the Stacking ensemble temperature rise model. Too large a decision tree depth will lead to overfitting, while too small a decision tree depth may underfit; the choice of the K value will affect the stability and reliability of K-fold cross-validation. By optimizing these two hyperparameters, the second layer meta-learner can better integrate the prediction results of the first layer base learners, improving the performance of the entire Stacking ensemble temperature rise model.

[0048] Grid search and K-fold cross-validation provide a systematic method for hyperparameter optimization, reducing the subjectivity of manual parameter selection. In traditional parameter selection methods, it is often necessary to rely on experience or trial-and-error to determine the value of the hyperparameters, which not only is inefficient, but also is easily affected by subjective factors. Grid search and K-fold cross-validation can automatically select the optimal combination of hyperparameters based on the performance indicators of the model on the validation set, improving the objectivity and accuracy of parameter selection. Through joint optimization of hyperparameters, the model can better adapt to different data sets.

[0049] The sample data set is as follows:

[0050] wherein, represents the input vector of the i-th sample, represents the response variable of the i-th sample; The optimized expression for the linear regression model is shown below:

[0051] in, This represents the weight coefficient matrix for linear regression. Represents the regularization coefficient; The optimization objective of the random forest regressor is shown in the following equation:

[0052]

[0053] in, and This represents the weight of the tree nodes in the decision tree. and Represents the splitting factor coefficient; The decision tree output model of the random forest regressor is shown in the following equation:

[0054] in, Indicates the number of times the leaf splits. Represents a split set; The optimization expression for the support vector machine model is shown below:

[0055] in, Indicates the number of training samples. Describes the normal vector of the linear hyperplane. The bias term represents the linear hyperplane. Represents the regularization parameter. Represents slack variables. Represents the kernel function; Kernel functions can be of various types, including linear kernels, polynomial kernels, or RBF kernels.

[0056] The optimization objective of the joint optimization of hyperparameters is to minimize the root mean square error, which is shown in the following equation:

[0057] in, This represents the output vector of the machine learning model trained on the reduced sample. This represents the actual output value of the sample.

[0058] The machine learning library used is the Scikit-learn machine learning library in Python 3.11.4.

[0059] Example 2 Referring to Figure 2 A transformer temperature rise evaluation system based on a Stacking integrated learning framework, comprising: A dataset construction module is configured to construct a sample dataset based on input variables and response variables by using Latin hypercube sampling principles and CFD simulation for a transformer with a known oil baffle and oil duct structure configuration; the input variables include oil baffle size, oil duct width, oil baffle quantity, and boundary temperature; and the response variables include winding wire oil temperature rise and hot spot temperature rise. A Stacking integrated modeling module is configured to hierarchically construct a Stacking integrated temperature rise model that fuses multiple model information, including a first layer of basic learners and a second layer of meta-learners; the sample dataset is used to independently train the first layer of basic learners to generate a plurality of prediction results; the prediction results are spliced to train the second layer of meta-learners to obtain the Stacking integrated temperature rise model that fuses multiple model information. A trained temperature rise prediction model module is configured to automatically train the Stacking integrated temperature rise model that fuses multiple model information after jointly optimizing the hyperparameters of the Stacking integrated temperature rise model that fuses multiple model information to determine the hyperparameters, to obtain a trained temperature rise prediction model. A temperature rise prediction module is configured to use the trained temperature rise prediction model to predict the temperature rise of a new oil baffle and oil duct structure configuration of a transformer to be predicted, to obtain output wire oil temperature rise and hot spot temperature rise prediction values. A prediction visualization module is configured to receive parameters of a new oil baffle and oil duct structure configuration of a transformer to be predicted, output winding hot spot temperature rise and wire oil temperature rise prediction values, and visualize temperature rise trend graphs and error distribution graph results.

[0060] Embodiment three Referring to Figure 3 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the method for evaluating the temperature rise of a transformer based on a Stacking integrated learning framework when executing the computer program.

[0061] Embodiment four A computer-readable storage medium stores a computer program, and the computer program implements the method for evaluating the temperature rise of a transformer based on a Stacking integrated learning framework when executed by a processor.

[0062] Embodiment five Referring to Figure 8 The embodiment provides a method for evaluating the temperature rise of a transformer based on a Stacking integrated learning framework, and the specific steps are as follows: Data construction: Based on CFD simulation software (optional ANSYS FLUENT), the thermal model of the power transformer under different oil baffle and oil duct structure configurations is constructed, and the winding wire oil and hot spot are both responses, that is, outputs Y， The size of the oil baffle, the width of the oil duct, the number of the oil baffle and the boundary temperature are inputs X， The winding wire oil temperature rise and hot spot temperature rise are extracted as response variables, and the size of the oil baffle, the width of the oil duct, the number of the oil baffle and the boundary temperature are extracted as input variables. A sample data set is constructed based on the Latin hypercube sampling principle (LHS). LHS constructs a representative input combination. Since there are infinite combinations for continuous variables, it is impossible to exhaust them one by one. Then, the sample results are solved using finite element modeling. When generating samples, if the number of design variables is d, the number of samples N satisfies N≥10×d to ensure high-dimensional space coverage. Each sample is distributed in the normalized design space [0,1]^d, and the sampling points are arranged according to the stratified uniform principle in each variable coordinate interval to avoid variable overlap or dense clustering. Based on the sampling principle, determine the combination scheme (input matrix X of all samples), then use finite element to obtain the output matrix (Y), form the sample data DATA=(X,Y), and the number of samples is N. Then use these discrete data to train the integrated learning model.

[0063] Basic learner modeling: Construct multiple basic regression models as the first layer learners of the Stacking framework, including but not limited to: (a) Linear Regression (LR): fitting the linear relationship between input structural parameters and temperature rise output; (b) Random Forest Regressor (RFR): using multiple decision trees to model nonlinear relationships; (c) K-Nearest Neighbors (KNN): local prediction by calculating Euclidean distance; each model is trained independently to obtain its own prediction output for the target temperature rise index. Train each sub-model based on the sample, then use the output Y* of the sub-model and the output Y of the final sample to form a new sample set (Y*,Y), and the second-level model is trained using the new sample set.

[0064] Stacking integrated learning modeling: The prediction results of the above multiple basic learners are input into the second layer meta-learner for relearning. The meta-learner uses a support vector regressor (Support Vector Regression, SVR) with a radial basis function (Radial Basis Function, RBF) kernel function to enhance the nonlinear mapping capability, thereby fusing the prediction information of different models and optimizing the overall prediction performance.

[0065] Model training and hyperparameter optimization step: Joint optimization of key hyperparameters of each base learner and meta-learner. For these machine learning algorithms, model training has a certain execution step and strategy. As long as the hyperparameters are confirmed, mature libraries can be called to achieve it. This embodiment uses the Scikit-learn machine learning library in python3.11.4 to train each machine learning module. The optimization algorithm uses grid search (Grid Search), and the evaluation standard is to minimize the root mean squared error (Root Mean Squared Error, RMSE), and combines five-fold cross-validation (k=5) for robustness test. Specific optimization parameters include but are not limited to: the number of trees (n_estimators) and the maximum depth (max_depth) of the RFR model; the number of neighbors (n_neighbors) and the weight function of the KNN model; the penalty parameter C and the kernel function parameter γ of the SVR model.

[0066] Temperature rise prediction and evaluation: The trained Stacking integrated model can be used for rapid temperature rise prediction of new structure parameter combinations, outputting line oil temperature rise and hot spot temperature rise values, and realizing efficient thermal performance analysis and design auxiliary decision-making under various structure conditions.

[0067] This embodiment uses a limited number of FLUENT simulation data for machine learning modeling, and realizes high-precision prediction of winding line oil temperature rise and hot spot temperature rise under different oil duct and baffle structure configurations through the Stacking integration strategy of fusing multiple base regressors and meta-learners, providing a fast, low-cost, and scalable thermal performance evaluation tool for engineering design. It overcomes the technical bottlenecks of relying on high-cost CFD simulation for transformer temperature rise evaluation, long modeling cycle, high degree of human involvement, and lack of batch scheme parallel prediction capability.

[0068] To cooperate with the above method, this embodiment also provides a transformer temperature rise evaluation system based on a Stacking integrated learning framework, comprising: Data interface module: used for importing sample data and structure parameters obtained by CFD simulation, and having LHS sampling and normalization preprocessing functions; Integrated modeling module: including submodules for constructing linear regression, random forest, and KNN three kinds of base learners, and a submodule for constructing a support vector machine meta-learner, completing a two-layer Stacking integrated modeling process; Parameter optimization module: performing hyperparameter grid search and cross-validation process, and outputting the optimal model structure; Prediction and visualization module: receiving new input parameters, outputting winding hot spot temperature rise and line oil temperature rise prediction values, and supporting temperature rise trend chart, error distribution chart, and other result visualization.

[0069] Taking a specific oil passage and baffle combination as an example, a transformer temperature rise assessment method based on the Stacking ensemble learning framework is used to assess the transformer's temperature rise. A simplified diagram of the ensemble learning framework is shown below. Figure 7 As shown, firstly, a transformer thermal model with different combinations of oil channels and baffles is constructed based on CFD simulation software to obtain data on winding oil temperature rise and hot spot temperature rise, thus building a sample set. Secondly, training samples are generated by uniformly sampling in the multi-dimensional structural parameter space using the Latin hypercube sampling method. Then, multiple basic learners such as linear regression, random forest, and K-nearest neighbors are established, and their outputs are used as feature inputs to train a support vector regressor to construct a two-layer stacking ensemble model. Further, the model hyperparameters are jointly optimized through grid search and cross-validation. Finally, the new structural parameters are input into the trained ensemble model to achieve rapid prediction of winding oil temperature rise and hot spot temperature rise. Correspondingly, the system consists of a data interface module, an ensemble modeling module, a parameter optimization module, and a prediction visualization module. This method can significantly reduce CFD simulation overhead while maintaining prediction accuracy, supports rapid batch temperature rise assessment under multiple structural conditions, and is suitable for thermal performance auxiliary verification and optimization analysis in the design stage of power transformers. The specific steps are as follows: (1) Data Construction Stage Two-dimensional heat flow models with different combinations of oil passages and baffles were established using CFD simulation software (ANSYS FLUENT preferred). The winding oil temperature rise and hot spot temperature rise were extracted as output variables, and the structural parameters (four design variables were selected in order: oil passage width, baffle size, number of winding discs and average oil temperature rise) were used as input variables. Based on the Latin hypercube sampling (LHS) method combined with the design variable interval, a data sample set with N representative values ​​was designed and constructed.

[0070] The final design variables can be represented by the following set structure:

[0071] in, Let represent the input vector for the i-th sample, preferably consisting of the oil passage width, baffle plate size, number of winding discs, and average oil temperature rise. This indicates the output of the corresponding winding hot spot temperature rise and wire oil temperature rise under the corresponding design variables; Figure 4 This demonstrates the physical model of a sample dataset. Figure 5 This demonstrates the finite element solution results for a specific data sample. (2) Basic learner modeling stage Constructing multiple basic regression models as the first layer learners in a stacking structure, including but not limited to: Linear Regression (LR): modeling the linear correlation between input and output; Random Forest Regressor (RFR): modeling the non-linear characteristics by integrating multiple decision trees; K-Nearest Neighbors (KNN): realizing local fitting based on the distance between samples. The output results of the above base learners are finally obtained. The following are the expressions of the regression optimization objectives of each model: The following is the optimization expression of the support vector machine model:

[0072] wherein, represents the number of training samples, represents the normal vector of the linear hyperplane, represents the bias item of the linear hyperplane, and the optimal solution can be obtained based on the convex optimization principle and the duality principle; represents the regularization parameter, represents the slack variable, represents the kernel function; the kernel function type can be selected from types such as linear kernel, polynomial kernel, and RBF kernel; hyperparameter optimization of the SVM support vector machine can not only improve the prediction accuracy and robustness, but also significantly reduce the computational overhead caused by traditional exhaustive search. and The selection of the kernel function will be based on the grid search principle to carry out hyperparameter optimization.

[0073] The improved linear regression model optimization expression is as follows:

[0074] wherein, is the number of training set samples, is the linear regression weight coefficient matrix; is the regularization coefficient; by introducing the regularization coefficient to each input weight, the problem of overfitting that may exist in ensemble learning can be alleviated by avoiding the excessive proportion of a single input variable; it is easy to know that based on the least squares method, for a linear regression equation with d input variables, if the sample number is 10d, when λ is known, based on the least squares method, the coefficient can obtain the optimal solution; For the random forest model, its essence is to integrate K decision trees, so the hyperparameter optimization involves the underlying model decision tree. For a standard decision tree, the optimization objective can be expressed as:

[0075] wherein, and denotes the weight of the decision tree leaf node, and denotes the split factor coefficient; Based on the least square idea, to meet the minimum split variance, c1, c2 need to meet:

[0076] So far, the final output model of the decision tree model can be obtained:

[0077] In the formula denotes the number of leaf splits, denotes the split set, which needs to be determined based on the split criterion and the split rule, and is also a hyperparameter that needs to be adjusted by machine learning; (3) Stacking integrated modeling stage: the prediction results of the above basic learners are taken as input features, which are input into the second layer meta-learner, support vector regression (SVR), to complete the fusion learning of multi-model prediction information.

[0078] (4) Hyperparameter optimization stage: grid search algorithm (Grid Search) combined with k-fold cross-validation, Figure 6 is the corresponding K-fold cross-validation diagram, where k=5, the hyperparameters (such as decision tree depth, K value, SVR kernel function parameters) of each basic model and meta-learner are optimized in combination to minimize the root mean square error (RMSE) as the target, to ensure the optimal performance of the model, and the optimization target is as follows:

[0079] In the formula, is the output vector of the machine learning model trained under the corresponding converted sample, is the true output value corresponding to the sample; in addition, because the cross-validation factor is considered, the output of the integrated learning model needs to meet the minimum error under k-fold combination, that is, based on sample division, the integrated learning framework is trained and performance evaluated under k sample combinations; (5) Prediction and evaluation stage: for any new input structure parameter combination, the trained Stacking integrated model is used to quickly output the winding wire oil temperature rise and hot spot temperature rise prediction value, realizing batch temperature rise evaluation under large-scale structure configuration scheme.

[0080] The embodiment is verified by experiments to significantly improve the evaluation efficiency: compared with the traditional CFD thermal flow field simulation method based on ANSYS FLUENT, the embodiment only needs to use a small amount of high-fidelity simulation data for training, and can realize temperature rise prediction under new structure parameters. In a typical application scenario, when predicting the temperature rise of 100 groups of oil duct and oil baffle combined structures, the traditional CFD method takes about 6-8 hours for simulation per group, while after the training of the embodiment method, the prediction time of each group is less than 1 second, and the overall evaluation efficiency is improved by more than 99%. The prediction accuracy is controllable and reliable: on 40 test samples not involved in training, the average absolute error (MAE) of the prediction results of the method and the CFD simulation results is controlled within ±2.5°C, and the maximum hotspot temperature rise error is not more than 4°C. The prediction results are verified by comparing with the high-fidelity simulation output, and the root mean square error (RMSE) is kept below 2.1°C, and the determination coefficient R2 can reach 0.96, meeting the accuracy requirements of engineering checking applications. Support high-dimensional multivariate modeling and strong structure generalization ability: by fusing multiple basic learners and introducing SVR as a meta-learner Stacking framework, the method has strong non-linear modeling and local adaptation ability, suitable for temperature rise response modeling of various structure parameters (such as oil duct width, oil baffle size, winding arrangement, etc.), and has good parameter expansion and model migration ability. Reduce modeling manpower and computing resource investment: the method can significantly reduce the dependence on simulation engineers, and the sample construction and model training process can be realized semi-automatically, greatly reducing the modeling manpower cost. At the same time, the model reasoning process has very low demand for computing resources, and is suitable for deployment on ordinary PCs, cloud servers or embedded platforms, and has wide engineering landing ability. It is suitable for parameter sensitivity analysis and rapid screening in the design stage: using the trained integrated model, the temperature rise response evaluation under parameter perturbation can be completed in seconds, supporting sensitivity analysis, pre-screening of optimization design, structure parameter optimization, etc. Engineering applications provide an efficient decision-making tool for transformer thermal field design.

[0081] The embodiment uses a Stacking integrated learning framework to establish a temperature rise prediction model for transformer CFD thermal simulation data, replaces the high computational burden simulation process, realizes rapid prediction of winding wire oil temperature rise and hotspot temperature rise, and is applied to transformer thermal performance evaluation. A two-layer Stacking framework is constructed, which includes three basic regression models (LR, RFR, KNN) as the first layer and a support vector machine regressor (SVR) as the second layer meta-learner, to realize model generalization ability enhancement and non-linear relationship modeling. The proxy model is coupled with the structure parameters to have fast prediction ability, and after inputting the structure parameter combination, the system can directly output the hotspot temperature rise and wire oil temperature rise prediction values, support rapid evaluation and batch optimization in the transformer structure design process, and are especially suitable for high-dimensional multivariate design space.

[0082] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0083] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0084] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0085] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0086] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing the technical solutions of the present application, but not for limiting it. Although the present application is described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the present application.

Claims

1. A transformer temperature rise evaluation method based on a Stacking integrated learning framework, characterized in that, The method comprises the following steps: For a transformer with a known oil dam and oil duct structure configuration, Latin hypercube sampling principle and CFD simulation are used to construct a sample data set based on input variables and response variables; the input variables include oil dam size, oil duct width, number of oil dams, and boundary temperature; the response variables include winding wire oil temperature rise and hot spot temperature rise; A Stacking integrated temperature rise model fusing multi-model information is constructed in layers, including a first layer of basic learners and a second layer of meta-learners; the first layer of basic learners is independently trained using the sample data set to generate a plurality of prediction results; the prediction results are spliced to train the second layer of meta-learners, thereby obtaining the Stacking integrated temperature rise model fusing multi-model information; After the super parameter joint optimization of the Stacking integrated temperature rise model fusing multi-model information, the machine learning library is called to automatically complete the training of the Stacking integrated temperature rise model fusing multi-model information to obtain a trained temperature rise prediction model; The trained temperature rise prediction model is used to predict the temperature rise of a new oil dam and oil duct structure configuration of a transformer to be predicted, thereby obtaining output wire temperature rise and hot spot temperature rise value predictions.

2. The transformer temperature rise evaluation method based on a stacking ensemble learning framework according to claim 1, characterized in that, A plurality of representative parameter combinations are generated in the input variable space using Latin hypercube sampling to obtain an input matrix; The temperature rise output of each parameter is calculated through thermodynamic calculation by CFD simulation to obtain an output matrix; The sample data set is obtained by combining the input matrix and the output matrix.

3. The transformer temperature rise evaluation method based on a stacking ensemble learning framework according to claim 1, characterized in that, The first layer of basic learners includes a linear regression model, a random forest regressor, and a K nearest neighbor regressor; the sample data set is input into the linear regression model, the random forest regressor, and the K nearest neighbor regressor for independent training, thereby obtaining linear regression prediction values, random forest regression prediction values, and K nearest neighbor regression prediction values; the linear regression model introduces a regularization coefficient; The linear regression prediction values, the random forest regression prediction values, and the K nearest neighbor regression prediction values are spliced into a new feature matrix; The new feature matrix and the original true temperature rise label are input into the second layer of meta-learners for training to obtain the Stacking integrated temperature rise model fusing multi-model information; the second layer of meta-learners uses a support vector regressor, and the kernel function of the support vector regressor uses a radial basis function.

4. The transformer temperature rise evaluation method based on a stacking ensemble learning framework according to claim 3, characterized in that, The optimization method of the super parameter joint optimization includes grid search and K-fold cross-validation; The super parameters include the number of standard decision trees and the maximum depth of the random forest regressor, the number of neighbors and the weight function of the K nearest neighbor regressor, the penalty parameter and the kernel function parameter of the support vector regressor, and the decision tree depth of the second layer of meta-learners and the K value of K-fold cross-validation.

5. The transformer temperature rise evaluation method based on a stacking ensemble learning framework according to claim 4, characterized in that, The sample data set is as follows: wherein, represents an input vector for the i-th sample, represents a response variable for the i-th sample; The optimization expression of the linear regression model is as follows: wherein, denotes a linear regression weight coefficient matrix, denotes a regularization coefficient; The optimization objective of the random forest regressor is as follows: wherein, and denotes the decision tree leaf node weight, and denotes the split factor coefficient; The decision tree output model of the random forest regressor is as follows: wherein, represents the number of leaf divisions, represents a division set; The optimization expression of the support vector machine model is as follows: wherein, denotes the number of training samples, denotes a normal vector of a linear hyperplane, denotes a bias term of the linear hyperplane, denotes a regularization parameter, denotes a slack variable, denotes a kernel function; The types of the kernel function include linear kernel, polynomial kernel, or RBF kernel.

6. The transformer temperature rise evaluation method based on a stacking ensemble learning framework according to claim 4, characterized in that, The optimization objective of the super parameter joint optimization is to minimize the root mean square error, which is as follows: wherein, represents a vector of outputs of the trained machine learning model in terms of the sample, represents a true output value of the sample.

7. The transformer temperature rise evaluation method based on a stacking ensemble learning framework according to claim 1, characterized in that, The machine learning library adopts the Scikit-learn machine learning library in python3.11.

4.

8. A transformer temperature rise evaluation system based on a Stacking ensemble learning framework, characterized in that, Comprise: The construction dataset module is used for adopting the Latin hypercube sampling principle and CFD simulation to construct a sample dataset based on input variables and response variables for the transformer with a known oil baffle and oil duct structure configuration; the input variables include the size of the oil baffle, the width of the oil duct, the number of the oil baffles and the boundary temperature, and the response variables include the winding wire oil temperature rise and the hot spot temperature rise; The stacking integrated modeling module is used for hierarchically constructing a stacking integrated temperature rise model fusing multiple model information, including a first layer basic learner and a second layer meta-learner, independently training the first layer basic learner by using the sample dataset, generating a plurality of prediction results, training the second layer meta-learner after splicing the prediction results, and obtaining the stacking integrated temperature rise model fusing multiple model information; The training temperature rise prediction model module is used for automatically training the stacking integrated temperature rise model fusing multiple model information after determining the hyperparameters by jointly optimizing the hyperparameters of the stacking integrated temperature rise model fusing multiple model information, and obtaining the trained temperature rise prediction model; The temperature rise prediction module is used for predicting the temperature rise of the transformer to be predicted with a new oil baffle and oil duct structure configuration by using the trained temperature rise prediction model, and obtaining output wire oil temperature rise and hot spot temperature rise prediction values; The prediction visualization module is used for receiving the parameters of the transformer to be predicted with a new oil baffle and oil duct structure configuration, outputting the winding hot spot temperature rise and wire oil temperature rise prediction values, and visualizing the temperature rise trend chart and error distribution chart results.

9. An electronic device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the transformer temperature rise evaluation method based on the stacking integrated learning framework in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the transformer temperature rise evaluation method based on the stacking integrated learning framework in any one of claims 1-7.

Images