Method and system for rapidly predicting temperature of electrical equipment

Abstract

The invention provides an electrical equipment temperature rapid prediction method, which comprises the steps of establishing a steady-state thermal field simulation model, and combining input values of all simulation parameters under a plurality of simulation working conditions to obtain a hot-spot temperature of electrical equipment in a specific degradation state under each simulation working condition; on the basis of a BP neural network, a temperature characteristic prediction model is constructed, the input characteristic quantity of the model comprises all simulation parameters with a certain weight proportion, and the output characteristic quantity is the hot spot temperature; setting a certain weighted loss function to update the temperature characteristic prediction model and perform training and testing to obtain a trained temperature characteristic prediction model, and performing optimization adjustment on hyper-parameters of the model by using a Bayesian optimization algorithm in each training process; and obtaining actual measurement values of all simulation parameters of the electrical equipment, and importing the actual measurement values into the trained temperature characteristic prediction model to obtain a hot-spot temperature prediction value. According to the invention, the data accuracy and timeliness in the degradation prediction process of the electrical equipment can be ensured at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment detection, and particularly to a method and system for rapidly predicting the temperature of electrical equipment. Background Art

[0002] During the operation of various electrical equipment, the continuous action of temperature stress is one of the main factors causing the degradation and even failure of electrical equipment. Once the electrical equipment fails, it will cause significant economic losses. Therefore, real-time prediction of the hot spot temperature of various electrical equipment is the premise for ensuring the safe and stable operation of the power system and the basis for predicting the degradation of various electrical equipment.

[0003] The current main methods for obtaining the hot spot temperature are: sensor temperature measurement method, software simulation method, empirical thermal model method, and artificial intelligence algorithm. Among them, the sensor temperature measurement method mainly measures the temperature by arranging sensors inside the electrical equipment. However, limited by the number of sensors and the installation location, the complete temperature field distribution of the electrical equipment cannot be obtained, and the slow degradation process under the working state of the electrical equipment cannot meet the timeliness requirements of degradation prediction. The software simulation method mainly calculates the internal thermal field distribution through a finite element simulation model to accurately predict the temperature of the electrical equipment. However, there are many components inside the electrical equipment and the structure is often very complex, resulting in a long finite element calculation time (for example, a single calculation usually takes several hours or even days). Especially for mass-produced products, for a service life cycle of several years, it cannot meet the timeliness requirements of degradation prediction. The empirical thermal model method is a commonly used hot spot temperature calculation method. They describe the temperature rise value of the hot spot temperature for the electrical equipment through one or two differential equations. However, accurate models cannot be trained from the data obtained through experience, resulting in many deviations in the prediction results. The artificial intelligence algorithm has been a research hotspot in recent years. Using a typical regression prediction model with multiple variables and multiple objectives, taking the operating environment temperature of the electrical equipment, the degradation degree of the degraded components, etc. as variables, and taking the temperature of each degraded component of the electrical equipment as the target of the prediction model. However, limited by the slow degradation process under the working state of the electrical equipment, sufficient degradation data to support model training cannot be obtained through actual tests, and it completely relies on simulation calculations to obtain test data, resulting in the need to improve the accuracy of the prediction results.

[0004] Therefore, in order to solve the problem that the above methods cannot meet the requirements of electrical equipment degradation prediction, it is necessary to combine the software simulation method with the artificial intelligence algorithm to detect the temperature of electrical equipment, which can ensure both the data accuracy and timeliness during the electrical equipment degradation prediction process. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of the present invention is to provide a method and system for quickly predicting the temperature of electrical equipment, which can ensure both the data accuracy and timeliness in the degradation prediction process of electrical equipment.

[0006] To solve the above technical problem, the embodiments of the present invention provide a method for quickly predicting the temperature of electrical equipment, and the method includes the following steps:

[0007] According to the actual structure of the electrical equipment, establish a steady-state thermal field simulation model, and determine the simulation conditions of the steady-state thermal field simulation model and the input values of all simulation parameters corresponding to each simulation condition, so as to further obtain the hot spot temperature of the electrical equipment in a specific degradation state under each simulation condition;

[0008] Based on the BP neural network, construct a temperature characteristic prediction model; wherein, the input feature quantities of the temperature characteristic prediction model include all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature;

[0009] Construct a data set to divide it into a training set and a test set; wherein, the data set is composed of the input values of all simulation parameters under each simulation condition and the corresponding obtained hot spot temperatures;

[0010] Set a certain weighted loss function to update the temperature characteristic prediction model, and train and test the updated temperature characteristic prediction model according to the training set and the test set, so as to obtain a trained temperature characteristic prediction model; wherein, the hyperparameters of the temperature characteristic prediction model are optimized and adjusted using a preset Bayesian optimization algorithm in each training process;

[0011] Obtain the measured values of all simulation parameters corresponding to the electrical equipment, and import them into the trained temperature characteristic prediction model to obtain the predicted value of the hot spot temperature corresponding to the electrical equipment.

[0012] Among them, the steady-state thermal field simulation model is constructed through a finite element model; among them,

[0013] The electrical equipment includes a heat-generating component and a non-heat-generating component;

[0014] If the non-heat-generating component contains a symmetric structure, each symmetric structure adopts a structured network, and a single-surface free triangular mesh division and a vertical sweeping meshing method are used to construct the corresponding finite element model; if the non-heat-generating component contains an asymmetric structure, each asymmetric structure adopts an unstructured mesh division, and a free tetrahedral mesh division and a mesh calibration method according to hydrodynamics are directly used to construct the corresponding finite element model;

[0015] The heat - generating component constructs the corresponding finite - element model by using a fine grid division for the area where the temperature is higher than a certain threshold and a coarse grid division for the area where the temperature is lower than the certain threshold.

[0016] Among them, the weighted loss function is Loss = 0.7RMSE high calorie +0.3RMSE low calorie ; where RMSE is the root - mean - square error; RMSE high calorie is the root - mean - square error of the high - calorific - value set; RMSE low calorie is the root - mean - square error of the low - calorific - value set.

[0017] Among them, the specific steps for optimizing and adjusting the hyperparameters of the temperature characteristic prediction model using a preset Bayesian optimization algorithm in each training process are as follows:

[0018] Determine the hyperparameters of the temperature characteristic prediction model, including the learning rate, training target error, and number of iterative trainings, and set the value range of each hyperparameter;

[0019] In each training process, take the hyperparameters in the temperature characteristic prediction model as the input of the Bayesian optimization algorithm, and the root - mean - square error RMSE high calorie of the high - calorific - value set and the root - mean - square error RMSE low calorie of the low - calorific - value set corresponding to each group of hyperparameters as the output of the Bayesian optimization algorithm; among them, the Bayesian optimization algorithm selects the Gaussian process as the surrogate model for iteration and selects the next group of hyperparameters through a preset expected improvement maximization function until the iterative calculation ends.

[0020] Among them, the electrical equipment is a space - borne power distributor, and its corresponding simulation parameters include the convective heat - transfer coefficient, surface radiation coefficient, coil thermal power, contact thermal power, and ambient temperature.

[0021] The embodiment of the present invention also provides an electrical equipment temperature rapid prediction system, including:

[0022] A historical data simulation unit, which is used to establish a steady - state thermal - field simulation model according to the actual structure of the electrical equipment, determine the simulation conditions of the steady - state thermal - field simulation model and the input values of all simulation parameters corresponding to each simulation condition, so as to further obtain the hot - spot temperature of the electrical equipment in a specific degradation state under each simulation condition;

[0023] A prediction model construction unit for constructing a temperature characteristic prediction model based on a BP neural network; wherein, the input feature quantities of the temperature characteristic prediction model include all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature;

[0024] A historical data set construction unit for constructing a data set to divide out a training set and a test set; wherein, the data set is composed of the input values of all simulation parameters under each simulation condition and the corresponding obtained hot spot temperatures;

[0025] A prediction model training unit for setting a certain weighted loss function to update the temperature characteristic prediction model, and training and testing the updated temperature characteristic prediction model according to the training set and the test set to obtain a trained temperature characteristic prediction model; wherein, the hyperparameters of the temperature characteristic prediction model are optimized and adjusted using a preset Bayesian optimization algorithm in each training process;

[0026] A predicted temperature unit for obtaining the measured values of all simulation parameters corresponding to the electrical equipment and importing them into the trained temperature characteristic prediction model to obtain the predicted value of the hot spot temperature corresponding to the electrical equipment.

[0027] Wherein, the steady-state thermal field simulation model is constructed through a finite element model; wherein,

[0028] The electrical equipment includes a heat-generating component and a non-heat-generating component;

[0029] If the non-heat-generating component contains a symmetric structure, each symmetric structure adopts a structured network, and a corresponding finite element model is constructed by using a single-surface free triangular mesh division followed by a vertical sweep meshing method; if the non-heat-generating component contains an asymmetric structure, each asymmetric structure adopts an unstructured mesh division, and a corresponding finite element model is directly constructed by using a free tetrahedral mesh division and a mesh calibration method according to hydrodynamics;

[0030] The heat-generating component constructs a corresponding finite element model by using a method of fine meshing in the area where the temperature is higher than a certain threshold and coarse meshing in the area where the temperature is lower than the certain threshold.

[0031] Wherein, the weighted loss function is Loss = 0.7RMSE high calorie +0.3RMSE low calorie ; wherein, RMSE is the root mean square error; RMSE high calorie is the root mean square error of the high heat generation set; RMSE low calorie is the root mean square error of the low heat generation set.

[0032] Among them, the electrical equipment is a space power distributor, and its corresponding simulation parameters include convective heat transfer coefficient, surface radiation coefficient, coil thermal power, contact thermal power, and ambient temperature.

[0033] Implementing the embodiments of the present invention has the following beneficial effects:

[0034] The present invention first simulates a data set through a steady-state thermal field simulation model to train a temperature characteristic prediction model, and then imports the measured data into the trained temperature characteristic prediction model to quickly predict the predicted value of the hot spot temperature. Thus, it can realize the combination of software simulation methods and artificial intelligence algorithms to detect the temperature of electrical equipment, and can simultaneously ensure the data accuracy and timeliness in the process of predicting the degradation of electrical equipment. Description of the Drawings

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, obtaining other drawings based on these drawings still belongs to the scope of the present invention.

[0036] Figure 1 It is a flowchart of a method for quickly predicting the temperature of electrical equipment provided by an embodiment of the present invention;

[0037] Figure 2 It is a heat transfer process diagram inside the space power distributor in the application scenario of a method for quickly predicting the temperature of electrical equipment provided by an embodiment of the present invention;

[0038] Figure 3 It is a comparison diagram of the prediction accuracy of the low heat generation set of the temperature characteristic prediction model in the application scenario of a method for quickly predicting the temperature of electrical equipment provided by an embodiment of the present invention;

[0039] Figure 4 It is a comparison diagram of the prediction accuracy of the high heat generation set of the temperature characteristic prediction model in the application scenario of a method for quickly predicting the temperature of electrical equipment provided by an embodiment of the present invention;

[0040] Figure 5 It is a schematic structural diagram of a system for quickly predicting the temperature of electrical equipment provided by an embodiment of the present invention. Detailed Embodiments

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the present invention in detail with reference to the drawings.

[0042] As Figure 1As shown in the figure, in an embodiment of the present invention, a method for rapidly predicting the temperature of an electrical device is provided. The method includes the following steps:

[0043] Step S1: According to the actual structure of the electrical device, establish a steady-state thermal field simulation model, and determine the simulation conditions of the steady-state thermal field simulation model and the input values of all simulation parameters corresponding to each simulation condition, so as to further obtain the hot spot temperature of the electrical device in a specific degradation state under each simulation condition.

[0044] The specific process is as follows. First, according to the actual structure of the electrical device (such as an aerospace power distributor), establish a steady-state thermal field simulation model to simulate the hot spot temperature data. Among them, the steady-state thermal field simulation model is constructed by a finite element model; among them, the electrical device includes heat-generating components and non-heat-generating components; if the non-heat-generating components contain symmetric structures, each symmetric structure adopts a structured network, and after using a single-surface free triangular mesh division and adding a vertical sweep meshing method to construct the corresponding finite element model; if the non-heat-generating components contain asymmetric structures, each asymmetric structure adopts an unstructured mesh division, and directly uses a free tetrahedral mesh division and a method of calibrating the mesh according to hydrodynamics to construct the corresponding finite element model; the heat-generating components adopt a method of fine meshing in the area where the temperature is higher than a certain threshold and coarse meshing in the area where the temperature is lower than a certain threshold to construct the corresponding finite element model.

[0045] Secondly, in the steady-state thermal field simulation model, select different simulation conditions (for example, use different ambient temperatures as the corresponding simulation conditions), and determine the corresponding simulation parameters (that is, the input variables of the steady-state thermal field simulation model).

[0046] Finally, input the values of all simulation parameters under each simulation condition, and perform simulation through the steady-state thermal field simulation model, so as to obtain the hot spot temperature of the electrical device in a specific degradation state under each simulation condition.

[0047] In an example, taking an aerospace power distributor as an example, the heat generation of the internal relay is the main heat source inside the single machine, and the heat is transferred to each component inside the single machine through radiation, conduction, etc. The heat transfer process inside the single machine is as follows Figure 2 shown:

[0048] A large amount of continuous heat is generated by the coil at the heat source position, and is transferred to the surface adjacent to the coil inside the relay by heat conduction, so that the temperature of the lead foot and the shell gradually increases. The heat flux Φ transferred by heat conduction inside this part of the solid is:

[0049]

[0050] Where: λ is the thermal conductivity; δ is the surface spacing; ΔT is the temperature difference between the surfaces, S is the surface area; R is the thermal resistance.

[0051] Subsequently, the lead-out pin and the housing transfer heat by means of thermal radiation and air convection. Among them, the expression of the heat flow rate of natural air convection and the temperature difference is:

[0052] φ = Ah(T s ―T a )

[0053] Where: A is the surface area; h is the convective heat transfer coefficient; T s and T a are the outer surface temperature of the relay and the ambient temperature, respectively.

[0054] The expression of thermal radiation is:

[0055]

[0056] Where: σ is the Stefan–Boltzmann constant, and its value is σ = 5.67×10 ―8 W / (m 2 ×K 4 );E is the surface emissivity.

[0057] The above is the finite element temperature field simulation calculation method. The initial thermal simulation parameters of the main components are shown in Table 1 below:

[0058] Table 1

[0059]

[0060] At this time, the aerospace power distributor includes structures such as connectors, relays, mounting plates, struts, and housings. After drawing the complete single-unit structure model, mesh generation is carried out. Due to the excessive number of single-unit components and the huge number of meshes, the simulation cannot be carried out. Therefore, the drawn geometric model is simplified on the premise of not affecting the temperature calculation, which can not only ensure the accuracy of the simulation calculation but also ensure the smooth progress of the simulation.

[0061] Considering the internal structure characteristics, simulation speed, and accuracy, different finite element meshes are used for meshing. Since the structural parts such as the housing, mounting plate, and strut are symmetric structures, structured meshes are used for meshing, and the meshing method of adding vertical sweeping after single-surface free triangular meshing is adopted. The advantage of the structured mesh meshing method is to reduce the number of meshes, improve the mesh quality, and increase the calculation speed.

[0062] Considering irregular structural components such as relays, unstructured meshes are used for meshing. And due to its good adaptability, free tetrahedral meshes can be directly used for meshing and the meshes are calibrated according to hydrodynamics.

[0063] Considering that the internal coil of the relay is a heat source, a more detailed mesh division is carried out on the relay coil to improve the accuracy of temperature simulation. There are more meshes in the area with higher temperature, making the simulation calculation more accurate, and coarser meshes are divided in the area with lower temperature to save calculation time.

[0064] For the mesh divided according to the above method, the element quality, skewness and aspect ratio all reach the reference values, indicating good mesh quality and strong simulation accuracy. The mesh division parameters of the single machine are shown in Table 2 below, and the reference values and actual values of the mesh quality standard are shown in Table 3 below.

[0065] Table 2

[0066]

[0067] Table 3

[0068]

[0069] The steady-state thermal simulation of ANSYS workbench is used to simulate the single-machine model. Since each simulation takes more than one hour, it is impossible to exhaust all simulation tests. Therefore, the orthogonal test method is used to overall design the test conditions. This method can evenly sample within the range of factor changes, making each test have strong representativeness, and collecting relatively complete test data through a small number of test times to achieve the effect of saving test time.

[0070] The core of the orthogonal test is the orthogonal table, and L represents the orthogonal table. It consists of three elements, and the formula is:

[0071] L n (q s )

[0072] In the formula, s represents the number of input conditions, q represents the number of values of each experimental factor, and n represents the total number of test cases, that is, the number of test cases generated by this orthogonal table.

[0073] At this time, the simulation parameters of the aerospace power distributor include the convective heat transfer coefficient, surface radiation coefficient, coil thermal power, contact thermal power and ambient temperature. These five simulation parameters take 36 values within the change range as the q value of the orthogonal test, and using the total number of test cases calculation formula of the orthogonal table, n = 176 is obtained. These 176 cases not only ensure the typicality of the test cases, but also ensure a large reduction in scale, reducing a large amount of redundancy. The n calculation formula is as follows

[0074]

[0075] The orthogonal experimental table is shown in Table 4 below:

[0076] Table 4

[0077]

[0078] Step S2: Based on the BP neural network, construct a temperature characteristic prediction model; wherein, the input feature quantities of the temperature characteristic prediction model include all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature;

[0079] Specifically, through multi-layer non-linear transformation, the BP neural network can effectively learn the non-linear relationships in complex data, has good feature learning and abstraction capabilities, and can automatically extract high-order feature representations from the data. In addition, the BP neural network has strong adaptability, can handle various data types and noise environments, shows excellent generalization ability, and maintains stable prediction ability for unseen data samples. Its parallel computing characteristics and adaptive learning algorithm enable it to efficiently perform large-scale data training and inference, and continuously optimize the model to improve the prediction accuracy. Among them, the BP neural network uses the newff function to create a forward neural network with an input layer, a hidden layer, and an output layer. The input layer has 5 neurons, corresponding to the five input feature quantities of the training data. The hidden layer uses the logarithmic sigmoid activation function, and the output layer uses the linear activation function. The Levenberg-Marquardt algorithm is used in the training process.

[0080] Therefore, based on the BP neural network, construct a temperature characteristic prediction model; wherein, the input feature quantities of this temperature characteristic prediction model include all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature.

[0081] In one example, construct a temperature characteristic prediction model corresponding to the aerospace power distributor. At this time, the convective heat transfer coefficient, surface radiation coefficient, coil thermal power, contact thermal power, and ambient temperature are all used as the input feature quantities of this model, and the hot spot temperature is used as the output feature quantity of this model.

[0082] At this time, the input feature vector is expressed as: X = [X 1 , X 2 ,..., X i ,..., X C ;

[0083] In the formula: C represents the number of working conditions, so that the input feature quantities under the i-th group of working conditions can be expressed as: X i = [h i , E i , P i , W i , T i .

[0084] Among them, h i , Ei , P i , W i , T i are respectively the convective heat transfer coefficient, surface radiation coefficient, coil thermal power, contact thermal power, and ambient temperature under the i-th working condition.

[0085]

[0086] In the formula: t represents the characteristic quantity under the i-th working condition.

[0087] Therefore, the hot spot temperature under the i-th working condition can be expressed as:

[0088]

[0089] Meanwhile, to improve the prediction accuracy of the model and effectively quantify the influence of each input and output characteristic quantity on the temperature field distribution, weight allocation is carried out mainly based on the influence of the change of each input characteristic quantity on the hot spot temperature.

[0090] Under the condition of keeping other parameters unchanged, use ANSYS workbench to simulate the hot spot temperature, and compare it with the initial hot spot temperature to obtain the temperature change. Its parameter adjustment and temperature change are shown in Table 5 below:

[0091] Table 5

[0092]

[0093] It can be seen from the above table that the changes in ambient temperature, convective heat transfer coefficient, and coil thermal power have a greater impact on the hot spot temperature and their parameter weights are higher, while the changes in contact thermal power and surface radiation coefficient have a smaller impact on the temperature and their parameter weights are lower. Based on the above and combined with previous experience, the initial weights of the five input characteristic quantities are set as [1 29 341 42] in sequence.

[0094] Step S3, construct a data set to divide out a training set and a test set; among them, the data set is composed of the input values of all simulation parameters under each simulation working condition and the corresponding obtained hot spot temperature;

[0095] The specific process is to construct a data set based on the simulation data in Step S1, including the input values of all simulation parameters under each simulation working condition and the corresponding obtained hot spot temperature, that is, to collect the data of the input characteristic quantity and the output characteristic quantity in the corresponding temperature characteristic prediction model into the data set.

[0096] The dataset is divided into a training set and a test set using the Hold-Out method, and they are stored in training(cv) and test(cv) respectively. As the indices for obtaining the training set and the test set, the data is split into training set feature variables and target variables, as well as test set feature variables and target variables according to the indices. The rng function is used to fix the random seed to facilitate result reproduction, where the test set accounts for 20% of the total dataset.

[0097] X = [X train + X test

[0098] Y = [Y train + Y test

[0099] In the formula, X train is the training set feature variable, X test is the training set target variable, Y train is the test set feature variable, and Y test is the test set target variable.

[0100] In one example, the aerospace power distributor is simulated and calculated according to the simulation conditions selected by the orthogonal experiment method, and the hot spot temperature data of the aerospace power distributor is obtained. A dataset under different environmental temperatures and specific degradation states is established, as shown in Table 6 below:

[0101] Table 6

[0102]

[0103]

[0104] Due to the large numerical differences and different dimensional units between variables, the mapminmax function is used for data normalization to ensure that the numerical ranges of the input feature quantities and the target data are consistent, improving the training effect and generalization ability of the model. The normalized data X i The calculation formula is as follows:

[0105]

[0106] The input_weights function is used to define the weights of the input feature quantities, which are set to [1 29 34 1 42] in sequence.

[0107] Step S4: Set a certain weighted loss function to update the temperature characteristic prediction model, and train and test the updated temperature characteristic prediction model according to the training set and the test set to obtain a trained temperature characteristic prediction model; among them, the hyperparameters of the temperature characteristic prediction model are optimized and adjusted using a preset Bayesian optimization algorithm in each training process;​​

[0108] Specifically, in machine learning, hyperparameters and parameters are two different concepts and play completely different roles in the model training process. Parameters are the learnable parameters inside the model, which directly affect the output of the model. During the training process, the model adjusts these parameters according to the input data to minimize the loss function. These parameters are learned by the model through training and can reflect the characteristics and patterns of the data. For example, in a linear regression model, the parameters are the weights, which determine the slope and intercept of the linear relationship. Hyperparameters are the parameters set before the model training. They control the training process of the model. These parameters cannot be directly learned from the training data and usually need to be adjusted manually. They affect the complexity and learning ability of the model, including the learning rate, regularization parameter, depth of the tree, and so on. The selection of hyperparameters usually needs to be determined based on experience and experiments. Different hyperparameter settings may lead to different model performances.

[0109] For different model requirements, the setting of hyperparameters should also have its tendency. It will be divided into a high calorific value set and a low calorific value set according to the different calorific values, and hyperparameter tuning will be carried out based on Bayesian optimization to reduce the RMSE of the high calorific value and improve its prediction accuracy.

[0110] To balance the prediction requirements of the high / low calorific value sets, a weighted loss function is designed as Loss = 0.7RMSE high calorie + 0.3RMSE low calorie , which is used to update the temperature characteristic prediction model;

[0111] where RMSE is the root mean square error; RMSE high calorie is the root mean square error of the high calorific value set; RMSE low calorie is the root mean square error of the low calorific value set.

[0112] In the process of optimizing and adjusting the hyperparameters of the temperature characteristic prediction model using the preset Bayesian optimization algorithm, first, determine the hyperparameters of the temperature characteristic prediction model, including the learning rate, training target error, and number of iterative trainings, and set the value range of each hyperparameter; second, in each training process, take the hyperparameters in the temperature characteristic prediction model as the input of the Bayesian optimization algorithm, and the root mean square error RMSE of the high calorific value set corresponding to each group of hyperparameters high calorie and the root mean square error RMSE of the low calorific value set low calorieAs the output of the Bayesian optimization algorithm; wherein, the Gaussian process is selected as the surrogate model for iteration in the Bayesian optimization algorithm, and the next set of hyperparameters is selected through a preset expected improvement maximization function until the iterative calculation ends.

[0113] In one example, the learning rate range is set to 0.001 to 0.1, the training target error range is set to 1e-8 to 1e-4, and the number of iterations range is set to 100 to 5000.

[0114] The Gaussian process is selected as the surrogate model, the Matern kernel function is configured to capture the non-linear relationship between hyperparameters, and 10 sets of initial hyperparameter combinations are generated through Latin hypercube sampling. During the optimization process, the next set of hyperparameters is selected through the expected improvement (EI) maximization function, and the EI calculation formula is:

[0115] EI(x) = E[MAX(f min ―f(x), 0)]

[0116] Set the initial hyperparameter combinations and the corresponding high and low calorific value set RMSE. In each iteration, update the model according to the known hyperparameters and performance metrics, and select the next hyperparameter combination to be evaluated to maximize the expected improvement. The feasible point parameters estimated by the Bayesian optimization algorithm are shown in Table 7 below:

[0117] Table 7

[0118]

[0119] Train the neural network according to the optimized hyperparameters, and the performance parameters of the high calorific value set and the low calorific value set are shown in Table 8 below. Its accuracy is as follows Figure 3 、 4 shown. It is expected that the accuracy of the temperature characteristics will be greatly improved after the hyperparameters are adjusted by the Bayesian optimization algorithm, and the improvement of the high calorific value set is more obvious.

[0120] Table 8

[0121]

[0122] Wherein, R2 is the coefficient of determination, MAE is the mean absolute error, and MBE is the mean relative error.

[0123] Step S5, obtain the measured values of all simulation parameters corresponding to the electrical equipment, and import them into the trained temperature characteristic prediction model to obtain the predicted value of the hot spot temperature corresponding to the electrical equipment.

[0124] The specific process is as follows. First, obtain the measured values of all simulation parameters corresponding to the electrical equipment. Second, import the measured values of all simulation parameters into the pre-trained temperature characteristic prediction model to quickly obtain the predicted value of the hot spot temperature corresponding to the electrical equipment.

[0125] In summary, the steady-state thermal field simulation model established by the finite element algorithm can not only accurately simulate the internal temperature distribution of electrical equipment, but also solve the problem of insufficient actual test data during the training of the algorithm model. Further combining the temperature characteristic prediction model established by the Bayesian optimization neural network can quickly and accurately predict the hot spot temperature of electrical equipment under different operating temperatures and degradation conditions.

[0126] At this time, compared with the simulation calculation time of more than one hour of the traditional finite element calculation method, the prediction time of this temperature characteristic prediction model is only 5 seconds, reducing the calculation time by 99.99%. Aiming at improving the accuracy of the prediction results of high-temperature relays, based on the method of Bayesian optimization hyperparameter optimization, the prediction accuracy is improved, and the prediction accuracy of high-heat-generating relays is increased from 94.90% to 99.37%. Through the physical thermal standby test and the method of the temperature simulation verification database for example verification, the maximum error is only 2.31%, and the average error is only 0.63%, and the prediction results can meet the requirements of reliability prediction.

[0127] As Figure 5 shown, in the embodiment of the present invention, a rapid temperature prediction system for electrical equipment is provided, including:

[0128] A historical data simulation unit 110, configured to establish a steady-state thermal field simulation model according to the actual structure of the electrical equipment, and determine the simulation working conditions of the steady-state thermal field simulation model and the input values of all simulation parameters corresponding to each simulation working condition, so as to further obtain the hot spot temperature of the electrical equipment in a specific degradation state under each simulation working condition;

[0129] A prediction model construction unit 120, configured to construct a temperature characteristic prediction model based on a BP neural network; wherein, the input feature quantity of the temperature characteristic prediction model includes all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature;

[0130] A historical data set construction unit 130, configured to construct a data set to divide a training set and a test set; wherein, the data set is composed of the input values of all simulation parameters under each simulation working condition and the corresponding obtained hot spot temperature;

[0131] The prediction model training unit 140 is used to set a certain weighted loss function, update the temperature characteristic prediction model, and train and test the updated temperature characteristic prediction model according to the training set and the test set to obtain a trained temperature characteristic prediction model; wherein the hyperparameters of the temperature characteristic prediction model are optimized and adjusted using a preset Bayesian optimization algorithm during each training process;

[0132] The predicted temperature unit 150 is used to obtain the measured values of all simulation parameters corresponding to the electrical device, and import them into the trained temperature characteristic prediction model to obtain the predicted value of the hot spot temperature corresponding to the electrical device.

[0133] Wherein, the steady-state thermal field simulation model is constructed by a finite element model; wherein,

[0134] The electrical equipment includes heat-generating components and non-heat-generating components;

[0135] If the non-heat-generating component contains a symmetrical structure, each symmetrical structure adopts a structured network, and a single-surface free triangle mesh division followed by vertical sweeping is used to construct a corresponding finite element model; if the non-heat-generating component contains an asymmetrical structure, each asymmetrical structure adopts an unstructured mesh division, and a free tetrahedral mesh division and a fluid dynamics calibration mesh are directly used to construct a corresponding finite element model;

[0136] The heat generating component constructs a corresponding finite element model by fine meshing of regions where the temperature is higher than a certain threshold and coarse meshing of regions where the temperature is lower than the certain threshold.

[0137] Wherein, the weighted loss function is Loss = 0.7RMSE high calorie +0.3RMSE low calorie ; Among them, RMSE is the root mean square error; RMSE high calorie is the root mean square error of the high heat generation set; RMSE low calorie is the RMS error of the low calorific value set.

[0138] Wherein, the electrical equipment is an aerospace distributor, and its corresponding simulation parameters include convection heat transfer coefficient, surface radiation coefficient, coil thermal power, contact thermal power and ambient temperature.

[0139] Implementing the embodiments of the present invention has the following beneficial effects:

[0140] In the present invention, a steady-state thermal field simulation model is first used to simulate a data set for training a temperature characteristic prediction model. Then, the measured data is imported into the trained temperature characteristic prediction model to quickly predict the predicted value of the hot spot temperature. Thus, it is possible to combine the software simulation method with the artificial intelligence algorithm to detect the temperature of electrical equipment, and at the same time ensure the data accuracy and timeliness in the degradation prediction process of electrical equipment.

[0141] It should be noted that in the above system embodiments, the included system modules are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional modules are only for easy distinction from each other and do not limit the protection scope of the present invention.

[0142] Those of ordinary skill in the art can understand that all or part of the steps in implementing the above method embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, such as ROM / RAM, disk, optical disc, etc.

[0143] The above-disclosed are only the preferred embodiments of the present invention, and of course, the scope of the rights of the present invention cannot be limited thereby. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope covered by the present invention.

Claims

1. A method for rapidly predicting the temperature of an electrical device, characterized in that, The method includes the following steps: According to the actual structure of the electrical equipment, establish a steady-state thermal field simulation model, and determine the simulation conditions of the steady-state thermal field simulation model and the input values of all simulation parameters corresponding to each simulation condition, so as to further obtain the hot spot temperature of the electrical equipment in a specific degradation state under each simulation condition; Based on the BP neural network, construct a temperature characteristic prediction model; wherein, the input feature quantities of the temperature characteristic prediction model include all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature; Construct a data set to divide it into a training set and a test set; wherein, the data set is composed of the input values of all simulation parameters under each simulation condition and the corresponding obtained hot spot temperature; Set a certain weighted loss function to update the temperature characteristic prediction model, and train and test the updated temperature characteristic prediction model according to the training set and the test set to obtain a trained temperature characteristic prediction model; wherein, the hyperparameters of the temperature characteristic prediction model are optimized and adjusted using a preset Bayesian optimization algorithm during each training process; Obtain the measured values of all simulation parameters corresponding to the electrical equipment and import them into the trained temperature characteristic prediction model to obtain the predicted value of the hot spot temperature corresponding to the electrical equipment.

2. The method for rapidly predicting the temperature of an electrical device according to claim 1, wherein The steady-state thermal field simulation model is constructed through a finite element model; wherein, The electrical equipment includes a heat-generating component and a non-heat-generating component; If the non-heat-generating component contains a symmetric structure, each symmetric structure adopts a structured network, and a single-surface free triangular mesh division is used followed by a vertical sweep meshing method to construct the corresponding finite element model; if the non-heat-generating component contains an asymmetric structure, each asymmetric structure adopts an unstructured mesh division, and a free tetrahedral mesh division and a mesh calibration method according to hydrodynamics are directly used to construct the corresponding finite element model; The heat-generating component adopts a method of fine meshing in the area where the temperature is higher than a certain threshold and coarse meshing in the area where the temperature is lower than the certain threshold to construct the corresponding finite element model.

3. The method for rapidly predicting the temperature of an electrical device according to claim 2, wherein The weighted loss function is Loss = 0.7RMSE high calorie + 0.3RMSE low calorie ; where RMSE is the root mean square error; RMSE high calorie is the root mean square error of the high calorific value set; RMSE low calorie is the root mean square error of the low calorific value set.

4. The method for rapidly predicting the temperature of an electrical device according to claim 3, wherein, The specific steps of optimizing and adjusting the hyperparameters of the temperature characteristic prediction model using a preset Bayesian optimization algorithm during each training process include: Determine that the hyperparameters of the temperature characteristic prediction model include the learning rate, training target error, and iterative training times, and set the value range of each hyperparameter; In each training process, the hyperparameters in the temperature characteristic prediction model are used as the input of the Bayesian optimization algorithm, and the root mean square error (RMSE) of the high calorific value set corresponding to each group of hyperparameters high calorie and the root mean square error (RMSE) of the low calorific value set low calorie are used as the output of the Bayesian optimization algorithm; among them, the Gaussian process is selected as the surrogate model for iteration in the Bayesian optimization algorithm, and the next group of hyperparameters is selected through a preset maximum expected improvement function until the iterative calculation ends.

5. The method for quickly predicting the temperature of an electrical device according to claim 1, wherein The electrical equipment is a space power distributor, and its corresponding simulation parameters include the convective heat transfer coefficient, surface radiation coefficient, coil heat power, contact heat power, and ambient temperature.

6. An electrical equipment temperature rapid prediction system, characterized in that, Include: A historical data simulation unit for establishing a steady-state thermal field simulation model according to the actual structure of the electrical equipment, and determining the simulation conditions of the steady-state thermal field simulation model and the input values of all simulation parameters corresponding to each simulation condition, so as to further obtain the hot spot temperature of the electrical equipment in a specific degradation state under each simulation condition; A prediction model construction unit for constructing a temperature characteristic prediction model based on a BP neural network; wherein, the input feature quantities of the temperature characteristic prediction model include all simulation parameters pre-allocated with corresponding weight ratios, and its output feature quantity is the hot spot temperature; A historical data set construction unit for constructing a data set to divide it into a training set and a test set; wherein, the data set is composed of the input values of all simulation parameters under each simulation condition and the corresponding hot spot temperature obtained; A prediction model training unit for setting a certain weighted loss function to update the temperature characteristic prediction model, and training and testing the updated temperature characteristic prediction model according to the training set and the test set to obtain a trained temperature characteristic prediction model; wherein, the hyperparameters of the temperature characteristic prediction model are optimized and adjusted using a preset Bayesian optimization algorithm during each training process; A predicted temperature unit for obtaining the measured values of all simulation parameters corresponding to the electrical equipment and importing them into the trained temperature characteristic prediction model to obtain the predicted value of the hot spot temperature corresponding to the electrical equipment.

7. The electrical equipment temperature rapid prediction system according to claim 6, wherein, The steady-state thermal field simulation model is constructed by a finite element model; wherein, The electrical equipment includes a heat-generating component and a non-heat-generating component; If the non-heat-generating component contains a symmetric structure, each symmetric structure adopts a structured network, and a single-surface free triangular mesh division is used and then a vertical sweep meshing method is used to construct the corresponding finite element model; if the non-heat-generating component contains an asymmetric structure, each asymmetric structure adopts an unstructured mesh division, and a free tetrahedral mesh division and a mesh calibration method according to hydrodynamics are directly used to construct the corresponding finite element model; The heat-generating component adopts a method of fine meshing in the area where the temperature is higher than a certain threshold and coarse meshing in the area where the temperature is lower than the certain threshold to construct the corresponding finite element model.

8. The electrical equipment temperature rapid prediction system according to claim 7, characterized in that, The weighted loss function is Loss = 0.7RMSE highcalorie + 0.3RMSE lowcalorie ; where RMSE is the root mean square error; RMSE highcalorie is the root mean square error of the high calorific value set; RMSE lowcalorie is the root mean square error of the low calorific value set.

9. The electrical equipment temperature rapid prediction system according to claim 6, characterized in that, The electrical equipment is an aerospace power distributor, and its corresponding simulation parameters include the convective heat transfer coefficient, the surface radiation coefficient, the coil heat power, the contact heat power, and the ambient temperature.

Images