Modeling method and device for IGBT switching loss prediction model

Abstract

The invention discloses a modeling method and device for an IGBT switching loss prediction model. The modeling method comprises the steps that a neural network structure used for constructing the prediction model is determined; obtaining dynamic characteristic data, wherein the dynamic characteristic data is divided into training data and test data; determining an optimal weight and an optimal threshold of the neural network structure based on a particle swarm algorithm; according to the training data, the optimal weight and the optimal threshold, training a neural network adopting a neural network structure to obtain a prediction model; evaluating the prediction performance of the prediction model according to the test data; and outputting an optimal prediction model according to the prediction performance. According to the method, the initial weight and the threshold of the neural network are optimized by adopting the particle swarm algorithm in a mode of combining global search and local search, the training process is prevented from falling into local optimum, the prediction precision of the prediction model is improved, the number of iterations in the neural network training process can be reduced through the optimal weight and the optimal threshold, and the prediction precision of the prediction model is improved. And the convergence speed is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of IGBT reliability online monitoring, and in particular to a modeling method and device for an IGBT switch loss prediction model. Background Art

[0002] High-voltage direct-mounted energy storage systems are a hot topic in current research and engineering applications due to their high safety and efficiency. However, for a 100-megawatt-level high-voltage direct-mounted energy storage system, which contains thousands of high-power high-voltage IGBTs, each of which has a high ground potential, it is difficult to measure the switching loss of the IGBTs online at the same time.

[0003] In recent years, with the successful application of big data technology and artificial intelligence technology in many fields, a large number of research results have shown that the artificial intelligence BP neural network can accurately, conveniently and reliably predict the IGBT switching loss without building an IGBT model or understanding its internal structure.

[0004] However, when using neural network training to obtain the IGBT prediction model, weights and thresholds, as important parameters of the neural network, are mostly obtained through random initialization, which will cause the training process to fall into the local optimum, resulting in slow convergence of the training process and low prediction accuracy of the obtained prediction model. Summary of the invention

[0005] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a modeling method and device for an IGBT switch loss prediction model, which can improve the convergence speed of the training process and the prediction accuracy of the prediction model.

[0006] In order to solve the above problems, the present invention is implemented according to the following scheme: A modeling method for an IGBT switching loss prediction model is provided, including: Determine the neural network architecture used to build the predictive model; Acquire dynamic characteristic data, and divide the dynamic characteristic data into training data and test data; Determine the optimal weights and thresholds of the neural network structure based on the particle swarm algorithm; According to the training data, the optimal weight and the optimal threshold, the neural network using the neural network structure is trained to obtain a prediction model; Evaluate the prediction performance of the prediction model based on the test data; According to the prediction performance, the optimal prediction model is output.

[0007] Compared with the prior art, the beneficial effects of the modeling method of an IGBT switch loss prediction model of the present invention are as follows: a particle swarm algorithm is used to optimize the initial weights and thresholds of the neural network by combining global search and local search, thereby avoiding the training process from falling into the local optimum, improving the prediction accuracy of the prediction model, and through the optimal weights and optimal thresholds, the number of iterations in the neural network training process can be reduced, thereby improving the convergence speed.

[0008] Optionally, the neural network structure includes: input variables, output variables, the number of hidden layer nodes, the activation function of the hidden layer, and the activation function of the output layer.

[0009] Optionally, it is characterized in that the input variables include DC voltage, AC current and IGBT junction temperature; Output variables include turn-on loss, turn-off loss, and diode reverse turn-off loss; Determine the number of hidden layer nodes according to the number of input variables and the number of output variables; The activation function of the hidden layer is the tansig function; The activation function of the output layer is purelin function.

[0010] Optionally, the method is characterized in that obtaining dynamic characteristic data comprises: Use high-voltage direct-mounted energy storage power modules to build an IGBT switch loss dynamic test system; Based on the actual working conditions of high-voltage direct-mounted energy storage, multiple test points are determined; The test point is used as the input condition of the IGBT switching loss dynamic test system, and the IGBT switching loss dynamic test system outputs the actual value of the switching loss corresponding to the test point; The test points and their corresponding actual switching loss values ​​are cleaned and normalized to obtain dynamic characteristic data.

[0011] Optionally, it is characterized in that, based on the actual working conditions of the high-voltage direct-mounted energy storage, multiple test points are determined, including: Based on the actual working conditions of high-voltage direct-mounted energy storage, determine the DC voltage range, AC current range and IGBT junction temperature range; Multiple test points are determined based on the DC voltage range, AC current range and IGBT junction temperature range.

[0012] Optionally, it is characterized in that determining the optimal weight and optimal threshold of the neural network structure based on a particle swarm algorithm comprises: Initialize the particle swarm randomly and encode each particle of the particle swarm based on the input variable; Calculate the fitness value of each particle at the current position based on the code of each particle in the particle swarm; Obtain the fitness value of each particle at the historical optimal position and the fitness value of the group at the historical optimal position; Compare the fitness value of the particle at the current position and its fitness value at the historical optimal position, and update the optimal position of the particle to the corresponding position with a smaller fitness value; Compare the fitness value of the particle at the current position with the fitness value of the group's historical optimal position, and update the group's optimal position to the corresponding position with a smaller fitness value; Update the particle's velocity, and update the particle's position based on the particle's velocity; When the termination condition is reached, the optimal weight and optimal threshold are output; when the termination condition is not reached, the above steps are repeated until the termination condition is reached, and the optimal weight and optimal threshold are output; the optimal weight and optimal threshold are the particle inertia coefficients of the particle swarm algorithm.

[0013] Optionally, the feature is that the training data includes the test points and the actual values ​​of the switching losses corresponding thereto; the optimal weights are the weights corresponding to the test points respectively, and the test points and the weights corresponding thereto respectively are weighted summed to obtain the target data; According to the training data, the optimal weight and the optimal threshold, the neural network using the neural network structure is trained to obtain a prediction model, including: Delete the test points and their actual switching loss values ​​corresponding to the target data being less than the optimal threshold; According to the remaining test points in the training data and their corresponding actual switching loss values, the neural network is trained to obtain a training model; The test points in the training data are used as input data of the training model in sequence, and the training model outputs the first prediction values ​​corresponding to the test points respectively; Obtaining a first prediction error of the training model according to actual values ​​of switching losses and first prediction values ​​corresponding to the test points respectively; When the first prediction error is less than the first preset prediction error, the training model is used as the prediction model; when the first prediction error is greater than or equal to the first preset prediction error, return to normalizing the test point and its corresponding actual value of the switching loss and execute backward until the first prediction error is less than the first preset prediction error, and the training model is used as the prediction model.

[0014] Optionally, it is characterized in that the test data includes the test points and their corresponding actual values ​​of switching losses; The prediction performance of the prediction model is evaluated based on the test data, including: The test points in the test data are used as input data of the prediction model in sequence, and the prediction model outputs the second prediction values ​​corresponding to the test points respectively; According to the actual values ​​of the switching losses corresponding to the test points and the second predicted values, a second prediction error of the prediction model is obtained; the prediction performance of the prediction model includes the second prediction error.

[0015] Optionally, the feature is that, according to the prediction performance, an optimal prediction model is output, including: When the second prediction error is less than the second preset prediction error, outputting the prediction model as the optimal prediction model; When the second prediction error is greater than or equal to the second preset prediction error, return to normalizing the test point and its corresponding actual switching loss value and execute it backward until the second prediction error is less than the second preset prediction error, and output the prediction model as the optimal prediction model.

[0016] A computer device is also provided, comprising a processor and a memory, wherein at least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the modeling method of the IGBT switch loss prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The process of the modeling method of the present invention is Figure 1 ; Figure 2 The process of the modeling method of the present invention is Figure 2 ; Figure 3 is an activation function curve diagram of the hidden layer of the present invention; Figure 4 is an activation function curve diagram of the output layer of the present invention; Figure 5 The basic principle of the IGBT switch loss prediction model of the present invention is Figure 1 ; Figure 6 The basic principle of the IGBT switch loss prediction model of the present invention is Figure 2 ; Figure 7 This is a diagram of the training process without the introduction of the particle swarm algorithm; Figure 8 A diagram of the training process of introducing a particle swarm algorithm into the present invention; Fig. 9 This is a curve chart of predicted value and actual value without introducing particle swarm algorithm; Fig.10 The predicted value and actual value curve graph of the particle swarm algorithm introduced in the present invention; Fig.11 This is the error ratio curve of the prediction model without the introduction of the particle swarm algorithm; Fig.12This is a graph showing the error ratio of the prediction model using the particle swarm algorithm introduced in the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0019] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0020] See also Figure 1-2 As shown, the present invention provides a modeling method for an IGBT switch loss prediction model, comprising: S1: Determine the neural network structure used to construct the prediction model; in the neural network used to construct the IGBT switch loss prediction model, it includes an input layer, a hidden layer and an output layer. When training the neural network to obtain the IGBT switch loss prediction model, the data is fed in through the input layer, processed by the hidden layer, and the output layer gives the prediction result; data processing is performed between the input layer and the hidden layer, and between the hidden layer and the output layer through the activation function, and the activation function mainly performs operations such as amplification, reduction, and amplitude limitation on the input data; therefore, the neural network structure includes: input variables, output variables, number of hidden layer nodes, activation function of the hidden layer, and activation function of the output layer.

[0021] For the IGBT of the high-voltage direct-mounted energy storage system, its driving voltage, driving resistance, switching frequency and parasitic parameters of the main circuit are determined and are almost unaffected by the external environment and working conditions; therefore, it is only necessary to consider the influence of the DC voltage, AC current and junction temperature of the IGBT device of the high-voltage direct-mounted energy storage system; in the high-voltage direct-mounted energy storage system, the DC voltage is determined by the battery voltage. During the charging process, the battery voltage continues to rise, and during the discharging process, the battery voltage continues to drop; the AC current is controlled by the energy storage energy management system. When the AC grid voltage remains unchanged, the energy storage energy management system realizes the energy management of battery charging and discharging by changing the AC current; the temperature of the IGBT device is directly affected by the ambient temperature and AC current. When the ambient temperature rises, the junction temperature of the IGBT device will also increase accordingly. When the AC side current increases, the temperature of the IGBT device will also increase; therefore, the three variables of DC voltage, AC current and IGBT junction temperature are used as the input variables of the IGBT switching loss prediction model, that is, the input variables include DC voltage, AC current and IGBT junction temperature.

[0022] The IGBT switching loss of the high-voltage direct-mounted energy storage system is affected by its main circuit, working environment, actual working conditions, etc.; in multi-level application fields such as high-voltage direct-mounted energy storage, during the turn-on process of the IGBT, the current rises from zero to the working current, and the voltage drops from a high level to a low level. During this turn-on process, the overlapping area of ​​the current and voltage will produce turn-on losses; during the turn-off process of the IGBT, the current drops from the working current to zero, and the voltage rises from a low level to a high level. During this turn-off process, the overlapping area of ​​the current and voltage will produce turn-off losses; there is stray inductance in the main circuit of the large-capacity IGBT, and the reverse-parallel diode of the IGBT will produce a voltage overshoot during the turn-off process, and the turn-off peak current is large, so the reverse turn-off loss of the diode on the thermal fatigue of the IGBT cannot be ignored; therefore, the three variables of turn-on loss, turn-off loss and diode reverse turn-off loss are used as the output variables of the IGBT switching loss prediction model, that is, the output variables include turn-on loss, turn-off loss and diode reverse turn-off loss.

[0023] Before determining the number of hidden layer nodes, that is, determining the number of neuron nodes in the hidden layer, it is necessary to determine the number of hidden layer layers. Usually, a BP neural network with one hidden layer can approximate any continuous function with arbitrary precision. Therefore, the number of hidden layers in the present invention is one layer; and the number of hidden layer nodes is constrained by the number of input variables and the number of output variables, and its calculation formula is as follows:

[0024] Among them, q is the number of hidden layer nodes, n is the number of input variables, and m is the number of output variables; in the present invention, the number of input variables is 3, the number of output variables is 3, that is, n=m=3, and substituting it into the calculation formula of the number of hidden nodes, it can be obtained that q=4.87, that is, the number of hidden nodes is 5.

[0025] Since the IGBT switching loss prediction model belongs to the function approximation type, the activation function of its hidden layer neurons selects the hyperbolic tangent S-type function tansig function. The antisymmetry of the tansig function makes the network learn faster, so the activation function of the hidden layer is ttansig function; and for the IGBT switching loss prediction model with only one hidden layer and using the tansig function as the activation function, the output layer uses the linear function purelin function as the activation function, which can achieve arbitrary precision approximation of multiple input functions, so the activation function of the output layer is purelin function.

[0026] See also Figure 3 As shown, it is a curve diagram of the activation function of the hidden layer, that is, a curve diagram of the tansig function. The expression of the tansig function is:

[0027] Among them, x is the input data of the hidden layer.

[0028] See also Figure 4 As shown, it is a curve diagram of the activation function of the output layer, that is, a curve diagram of the purelin function. The expression of the purelin function is:

[0029] Among them, x is the input data of the output layer.

[0030] See also Figure 5 As shown, the basic principle of the IGBT switch loss prediction model of the present invention is Figure 1 , Figure 5 Where X is the input variable of the input layer, Z is the neuron node of the hidden layer, and Y is the 1 is the actual output data, Y 2 is the output variable of the output layer, that is, Y 2 Output data for prediction.

[0031] The input variables of the present invention are DC voltage, AC current and IGBT junction temperature, namely Figure 5X1 corresponds to the DC voltage U, X2 corresponds to the AC current I, and X3 corresponds to the IGBT junction temperature T; the number of hidden layer nodes of the present invention is 5, that is, Z1, Z2, ..., Z5 are five neuron nodes of the hidden layer respectively; the output variables of the present invention are the three of turn-on loss, turn-off loss and diode reverse turn-off loss, that is, Y1 corresponds to the turn-on loss Eon, Y2 corresponds to the turn-off loss Eoff, and Y3 corresponds to the diode reverse turn-off loss Erec; e is the actual output data Y 1 And the predicted output data Y 2 The error value of .

[0032] S2: Acquire dynamic characteristic data, which is divided into training data and test data; wherein the training data is used to train the neural network to obtain the IGBT switch loss prediction model; the test data is used to evaluate the prediction performance of the obtained IGBT switch loss prediction model to obtain the optimal prediction model; acquire dynamic characteristic data, including: Firstly, a high-voltage direct-mounted energy storage power module is used to build an IGBT switching loss dynamic test system for obtaining dynamic characteristic data. In order to ensure the prediction accuracy of the IGBT switching loss prediction model, it is necessary to ensure the accuracy of the training data used to train the neural network to obtain the IGBT switching loss prediction model. Therefore, the input variable test points of the IGBT switching loss dynamic test system need to cover all operating conditions, that is, multiple test points are determined based on the actual operating conditions of the high-voltage direct-mounted energy storage system, and the test points are used as input conditions of the IGBT switching loss dynamic test system. The IGBT switching loss dynamic test system outputs the actual value of the switching loss corresponding to the test point. After data cleaning and normalization of the test points and their corresponding actual values ​​of switching losses, the dynamic characteristic data are obtained.

[0033] To determine multiple test points based on the actual operating conditions of the high-voltage direct-mounted energy storage system, it is first necessary to determine the input variable range of the IGBT switching loss prediction model, namely the DC voltage range, AC current range and IGBT junction temperature range. Then, multiple test points are determined based on the DC voltage range, AC current range and IGBT junction temperature range.

[0034] For example, the DC voltage range is 600V~1200V, the AC current range is 100A~1200A; the IGBT junction temperature range is 25℃~95℃, so that all test points are within these three ranges, that is, to ensure that the test points cover all actual working conditions of the high-voltage direct-mounted energy storage system; select multiple DC voltage points, multiple AC current points and multiple IGBT junction temperature points in the DC voltage range, AC current range and IGBT junction temperature range respectively, and combine multiple DC voltage points, multiple AC current points and multiple IGBT junction temperature points. For example, under the condition that the IGBT junction temperature is 25 degrees, the DC voltage is 600V and the AC current is 100A, and one combination can be obtained; under the condition that the IGBT junction temperature is 25 degrees, the DC voltage is 650V and the AC current is 100A, and the second combination can be obtained; and so on, all test points are listed; by changing the input conditions (i.e., test points) of the IGBT switching loss dynamic test system, the IGBT switching loss test data of all test points are obtained, that is, all test points and their corresponding actual values ​​of switching losses are obtained.

[0035] In the prior art, dynamic characteristic data are mostly obtained through the laboratory offline double-pulse method, and the IGBT switching loss is affected by its main circuit, working environment, actual working conditions, etc. The actual IGBT switching loss is greatly different from the double-pulse test data, which will cause the reliability of the predicted data actually output by the obtained prediction model to be affected; the present invention obtains dynamic characteristic data based on the actual working conditions of the high-voltage direct-mounted energy storage system, which can ensure the data accuracy of the prediction model obtained by training the neural network, and then ensure the accuracy of the predicted data output by the prediction model.

[0036] S3: Determine the optimal weights and thresholds of the neural network structure based on the particle swarm algorithm, including: The particle swarm is randomly initialized, and each particle of the particle swarm is encoded based on the input variables. Assuming that the particle swarm size is Q, the three input variables of DC voltage, AC current and IGBT junction temperature are encoded. The i-th particle is encoded as .

[0037] The fitness value of each particle at the current position is calculated based on the code of each particle in the particle swarm. The fitness function for calculating the fitness value of each particle at the current position is a prior art and will not be described in detail here.

[0038] Obtain the fitness value of each particle at the historical optimal position and the fitness value of the historical optimal position of the group; compare the fitness value of the particle at the current position and its fitness value at the historical optimal position, and update the optimal position of the particle to the corresponding position with a smaller fitness value; that is, if the fitness value of the current position is less than the fitness value of the historical optimal position, the optimal position of the particle is updated to the current position; if the fitness value of the current position is greater than or equal to the fitness value of the historical optimal position, the optimal position of the particle remains unchanged.

[0039] Compare the fitness values ​​of all particles at the current position with the fitness values ​​of the historical optimal position of the group, and update the optimal position of the group to the corresponding position with a smaller fitness value; that is, if there is a particle among all the particles whose fitness value at the current position is less than the fitness value of the historical optimal position of the group, the optimal position of the group is updated to the current position of the particle; if the fitness values ​​of the current positions of all particles are greater than or equal to the fitness value of the historical optimal position of the group, the optimal position of the group remains unchanged.

[0040] Update the particle's velocity, and update the particle's position based on the particle's velocity; the particle velocity update formula is as follows:

[0041] Among them, v represents the particle velocity, x represents the particle position, k represents the current iteration number, i represents the ith particle, ω represents the particle inertia coefficient, m represents the historical optimal position of the ith particle, j represents the total number of particle swarms, n represents the historical optimal position of the swarm, and C1 and C2 are the particle non-negative acceleration coefficients.

[0042] The update formula of the particle position is as follows:

[0043] Among them, x represents the particle position, v represents the particle velocity, k represents the current iteration number, i represents the i-th particle, and D represents the total number of independent variables, which represents the dimension of the particle.

[0044] In the particle swarm algorithm, the particle inertia coefficient ω is an important parameter; when the particle inertia coefficient ω is large, the particle swarm algorithm has a strong global search capability, and the particles are more inclined to explore new areas, which helps the algorithm to jump out of the local optimal solution, but may cause the algorithm to converge slowly; when the particle inertia coefficient ω is small, the particle swarm algorithm has a strong local search capability, and the particles are more inclined to search near their historical best positions, which is conducive to the rapid convergence of the algorithm, but it is very easy to fall into the risk of local optimal solutions; therefore, by linearly reducing the value of the particle inertia coefficient ω, the particle swarm algorithm can find the particle inertia coefficient ω that takes into account both global search and local search capabilities, and determine the final particle inertia coefficient as the optimal weight and optimal threshold; the hidden layer and the output layer correspond to the optimal weight and optimal threshold, respectively.

[0045] Whether the maximum number of iterations is reached is used as the termination condition, and the maximum number of iterations is not limited here; when the termination condition is reached, the optimal weight and the optimal threshold are output; when the termination condition is not reached, the above steps are repeated until the termination condition is reached, and the optimal weight and the optimal threshold are output; the optimal weight and the optimal threshold are the particle inertia coefficient ω of the particle swarm algorithm.

[0046] See also Figure 6 As shown, the basic principle of the IGBT switch loss prediction model of the present invention is Figure 2 ,in, Figure 6 V and b in 1 Respectively represent the optimal weight and optimal threshold of the hidden layer; W and b 2 They represent the optimal weight and optimal threshold of the output layer respectively.

[0047] S4: According to the training data, the optimal weights and the optimal thresholds, the neural network using the neural network structure is trained to obtain a prediction model; wherein the training data includes the test points and their corresponding actual values ​​of switching losses; the optimal weights are the weights corresponding to the test points respectively, that is, the optimal weights include the DC voltage weight, the AC current weight and the IGBT junction temperature weight; the test points and their corresponding weights are weighted and summed to obtain the target data.

[0048] According to the training data, the optimal weights and the optimal thresholds, the neural network using the neural network structure is trained, that is, the three-input-three-output neural network of the present invention is trained to obtain a prediction model, including: First, the test points and their actual switching loss values ​​corresponding to the target data being less than the optimal threshold are deleted; then, the neural network is trained based on the remaining test points in the training data and their corresponding actual switching loss values, that is, the neural network is trained based on the remaining training data after the deletion to obtain a training model; The test points in the training data are used as the input data of the training model in sequence, and the training model outputs the first prediction values ​​corresponding to the test points respectively; according to the actual values ​​of the switching losses corresponding to the test points and the first prediction values ​​respectively, the first prediction error of the training model (corresponding to Figure 5 (e).

[0049] When the first prediction error is less than the first preset prediction error, the training model is used as the prediction model; when the first prediction error is greater than or equal to the first preset prediction error, return to step S2 to normalize the test point and its corresponding actual value of the switching loss and execute backward until the first prediction error is less than the first preset prediction error, and the training model is used as the prediction model; the first prediction error here is the mean square error value, which is usually 0.0002.

[0050] S5: evaluating the prediction performance of the prediction model according to the test data; wherein the test data includes the test points and the actual values ​​of the switching losses corresponding thereto; evaluating the prediction performance of the prediction model according to the test data includes: The test points in the test data are used as input data of the prediction model in sequence, and the prediction model outputs second prediction values ​​corresponding to the test points respectively; the second prediction error of the prediction model is obtained according to the actual values ​​of the switching losses corresponding to the test points and the second prediction values; the prediction performance of the prediction model includes the second prediction error.

[0051] S6: Output the optimal prediction model based on the prediction performance, including: When the second prediction error is less than the second preset prediction error, the prediction model is output as the optimal prediction model; when the second prediction error is greater than or equal to the second preset prediction error, return to step S2 to normalize the test point and its corresponding actual value of the switching loss and execute backward until the second prediction error is less than the second preset prediction error, and the prediction model is output as the optimal prediction model.

[0052] The optimal prediction model finally obtained by the present invention is the IGBT switch loss prediction model, which is a prediction model of a three-input-three-output neural network structure. Compared with the prediction model of a multi-input-single-output neural network structure in the prior art, the present invention can simultaneously predict the turn-on loss, turn-off loss and diode reverse turn-off loss, thereby effectively improving the prediction efficiency.

[0053] In order to verify the improvement effect of introducing particle swarm optimization on IGBT switching loss prediction model, Figure 7 The training process diagram without introducing the particle swarm algorithm is shown. Figure 8 The training process diagram of the particle swarm algorithm is shown. By comparing Figure 7-8 That is, by comparing the training convergence process without introducing the particle swarm algorithm and with introducing the particle swarm algorithm, it can be found that the IGBT switch loss prediction model obtained by introducing the particle swarm algorithm has a faster convergence speed and a smaller mean square error value.

[0054] exist Figure 7 In the conventional BP neural network without the particle swarm algorithm, that is, the initial weights and thresholds of the neural network are randomly given, 92 rounds of iterations are required, and the mean square error is 0.00015003, which is less than the set 0.0002. Figure 8 In the , the neural network with particle swarm algorithm is introduced, that is, the initial weights and thresholds of the neural network are optimized, and only 53 rounds of iterations are needed, and the mean square error value can reach 0.00013416. That is, after the particle swarm algorithm is introduced, the number of iterations is less, the convergence speed is faster, and the mean square error value is smaller.

[0055] See also Figure 9-10 , Fig. 9 The graphs of predicted values ​​and actual values ​​without introducing the particle swarm algorithm are shown. Fig.10 The predicted value and actual value curve of the particle swarm algorithm are shown for comparison. Figure 9-10 It can be found that Fig.10 The fit between the predicted value and the actual value will be better than Fig. 9 The fit between the predicted value and the actual value is high, which proves that the prediction model obtained by introducing the particle swarm algorithm in the present invention has a higher accuracy of output prediction data than the prediction model obtained without introducing the particle swarm algorithm.

[0056] See also Figure 11-12 , Fig.11 The error ratio curve of the prediction model without the introduction of particle swarm algorithm is shown. Fig.12 The error ratio curve of the prediction model introduced with the particle swarm algorithm is shown; Fig.11 For the 14th test sample, the maximum error value of the diode reverse turn-off loss is 8.6%; for the 42nd test sample, the maximum error value of the IGBT turn-on loss is 9.2%; for the 40th test sample, the maximum error value of the IGBT turn-off loss is 8.6%; Fig.12 For the 19th test sample, the maximum error value of the diode reverse turn-off loss is 3.3%; for the 47th test sample, the maximum error value of the IGBT turn-on loss is 1.6%; for the 30th test sample, the maximum error value of the IGBT turn-off loss is 2.0%; it can be seen that the prediction model obtained by introducing the particle swarm algorithm in the present invention has an output prediction data accuracy improved by at least 5% compared with the prediction model obtained without introducing the particle swarm algorithm.

[0057] The present invention not only improves the convergence speed of the training process of the IGBT switch loss prediction model, but also improves the accuracy of the prediction model; the initial weights and thresholds of the conventional BP neural network are based on randomly given values, which are prone to fall into local optimal values, resulting in slow convergence of the model training process and low accuracy; the present invention optimizes the initial weights and thresholds, accelerates the convergence of the IGBT switch loss prediction model training process, and improves the prediction accuracy of the prediction model.

[0058] The IGBT switch loss prediction model of the present invention has high reliability; the conventional prediction model training data is obtained based on the laboratory double pulse method, which is very different from the actual working conditions, so the reliability of the training data is low. The training data (dynamic characteristic data) of the present invention is based on high-voltage direct-mounted energy storage power modules, simulating actual working conditions to test IGBT switch losses, thereby improving the reliability of the prediction model.

[0059] The present invention also provides a computer device, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the above-mentioned modeling method.

[0060] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0061] The memory can be used to store the computer program or module, and the processor realizes various functions of the modeling method by running or executing the computer program or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0062] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A modeling method for an IGBT switching loss prediction model, characterized in that: include: determining a neural network structure for a neural network used to construct a predictive model; Acquire dynamic characteristic data, and divide the dynamic characteristic data into training data and test data; Determine the optimal weights and thresholds of the neural network structure based on the particle swarm algorithm; According to the training data, the optimal weight and the optimal threshold, the neural network using the neural network structure is trained to obtain a prediction model; Evaluate the prediction performance of the prediction model based on the test data; According to the prediction performance, the optimal prediction model is output.

2. The modeling method of an IGBT switching loss prediction model according to claim 1, characterized in that: The neural network structure includes: input variables, output variables, number of hidden layer nodes, activation function of the hidden layer, and activation function of the output layer.

3. The modeling method of an IGBT switching loss prediction model according to claim 2, characterized in that: Input variables include DC voltage, AC current and IGBT junction temperature; Output variables include turn-on loss, turn-off loss, and diode reverse turn-off loss; Determine the number of hidden layer nodes according to the number of input variables and the number of output variables; The activation function of the hidden layer is the tansig function; The activation function of the output layer is purelin function.

4. The modeling method of an IGBT switch loss prediction model according to claim 3, characterized in that: Get dynamic characteristic data, including: Use high-voltage direct-mounted energy storage power modules to build an IGBT switch loss dynamic test system; Based on the actual working conditions of high-voltage direct-mounted energy storage, multiple test points are determined; The test point is used as the input condition of the IGBT switching loss dynamic test system, and the IGBT switching loss dynamic test system outputs the actual value of the switching loss corresponding to the test point; The test points and their corresponding actual switching loss values ​​are cleaned and normalized to obtain dynamic characteristic data.

5. The modeling method of an IGBT switch loss prediction model according to claim 4, characterized in that: Based on the actual working conditions of high-voltage direct-mounted energy storage, multiple test points are determined, including: Based on the actual working conditions of high-voltage direct-mounted energy storage, determine the DC voltage range, AC current range and IGBT junction temperature range; Multiple test points are determined based on the DC voltage range, AC current range and IGBT junction temperature range.

6. The modeling method of an IGBT switch loss prediction model according to claim 3, characterized in that: Determine the optimal weights and thresholds of the neural network structure based on the particle swarm algorithm, including: Initialize the particle swarm randomly and encode each particle of the particle swarm based on the input variable; Calculate the fitness value of each particle at the current position based on the code of each particle in the particle swarm; Obtain the fitness value of each particle at the historical optimal position and the fitness value of the group at the historical optimal position; Compare the fitness value of the particle at the current position and its fitness value at the historical optimal position, and update the optimal position of the particle to the corresponding position with a smaller fitness value; Compare the fitness value of the particle at the current position with the fitness value of the group's historical optimal position, and update the group's optimal position to the corresponding position with a smaller fitness value; Update the particle's velocity, and update the particle's position based on the particle's velocity; When the termination condition is reached, the optimal weight and optimal threshold are output; when the termination condition is not reached, the above steps are repeated until the termination condition is reached, and the optimal weight and optimal threshold are output; the optimal weight and optimal threshold are the particle inertia coefficients of the particle swarm algorithm.

7. The modeling method of an IGBT switch loss prediction model according to claim 5, characterized in that: The training data includes test points and their corresponding actual switching loss values; The optimal weight is the weight corresponding to each test point. The test points and their corresponding weights are weighted summed to obtain the target data. According to the training data, the optimal weight and the optimal threshold, the neural network using the neural network structure is trained to obtain a prediction model, including: Delete the test points and their actual switching loss values ​​corresponding to the target data being less than the optimal threshold; According to the remaining test points in the training data and their corresponding actual switching loss values, the neural network is trained to obtain a training model; The test points in the training data are used as input data of the training model in sequence, and the training model outputs the first prediction values ​​corresponding to the test points respectively; Obtaining a first prediction error of the training model according to actual values ​​of switching losses and first prediction values ​​corresponding to the test points respectively; When the first prediction error is less than the first preset prediction error, the training model is used as the prediction model; when the first prediction error is greater than or equal to the first preset prediction error, return to normalizing the test point and its corresponding actual value of the switching loss and execute backward until the first prediction error is less than the first preset prediction error, and the training model is used as the prediction model.

8. The modeling method of an IGBT switch loss prediction model according to claim 5, characterized in that: The test data includes the test points and their corresponding actual switching loss values; The prediction performance of the prediction model is evaluated based on the test data, including: The test points in the test data are used as input data of the prediction model in sequence, and the prediction model outputs the second prediction values ​​corresponding to the test points respectively; According to the actual values ​​of the switching losses corresponding to the test points and the second predicted values, a second prediction error of the prediction model is obtained; the prediction performance of the prediction model includes the second prediction error.

9. The modeling method of an IGBT switching loss prediction model according to claim 8, characterized in that: According to the prediction performance, the optimal prediction model is output, including: When the second prediction error is less than the second preset prediction error, outputting the prediction model as the optimal prediction model; When the second prediction error is greater than or equal to the second preset prediction error, return to normalizing the test point and its corresponding actual switching loss value and execute it backward until the second prediction error is less than the second preset prediction error, and output the prediction model as the optimal prediction model.

10. A computer device, characterized in that: The computer device includes a processor and a memory, wherein at least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the modeling method of the IGBT switch loss prediction model as described in any one of claims 1 to 9.

Images