CN113162375B - Modeling method for switch loss prediction in IGBT dynamic process - Google Patents

Abstract

The invention relates to a modeling method for switching loss prediction in the dynamic process of an IGBT, which is an IGBT switching loss prediction method optimized by an extreme learning machine based on the Jiadian krill swarm algorithm. After completing the experimental data processing, the basic parameter setting of the extreme learning machine and the krill swarm algorithm, the optimal point set algorithm is used to optimize the initial krill swarm, which is used as the weight threshold of the extreme learning machine, and the fitness of the good point krill is calculated. . In the process of optimization, Jiadian krill continuously updates its position with Levi flight and cosine control factor as wings, and calculates the fitness of Jiadian krill until the end; finally, according to the extreme learning obtained by Jiadian krill The optimal weight threshold of the machine is used to predict and output the IGBT on and off loss values. The invention implements dynamic adjustment for algorithm optimization, so that the prediction model has high prediction accuracy and fast prediction speed, and the prediction result has good guiding significance for engineers to improve the heat dissipation system of the IGBT module.

Description

A Modeling Method for Switching Loss Prediction in IGBT Dynamic Process

technical field

The technical scheme of the invention belongs to the technical field of IGBT reliability of power electronic devices, in particular to a modeling method for switching loss prediction in the dynamic process of IGBT.

Background technique

With the continuous aggravation of the energy crisis, the continuous development of power electronic technology has effectively promoted the progress and development of society. IGBTs, as modern power electronic switches, are widely used in power systems, electric vehicles, and high-speed traction. However, the failures of PV inverters with IGBT as the core account for about 37% of the total failures. Among the failure failures of power electronic systems, the failures caused by temperature account for about 55% of the total failures, and there is a negative correlation between the temperature of the device, its safety margin and thermal cycle life. The occurrence of faults seriously affects the normal operation of the system and reduces the reliability of the system operation. Therefore, it is particularly important to study the overheating fatigue and overheating loss in IGBT operation.

The research on overheating fatigue and overheating loss of IGBT by domestic and foreign scholars mainly focuses on the calculation of IGBT switching loss. The switching loss of an IGBT refers to the power dissipated on the IGBT in one cycle. The switching loss is closely related to the IGBT collector-emitter voltage, collector current, switching frequency and the resistance and voltage on the driving circuit. And the reliability of the IGBT has a great relationship with the temperature rise fluctuation caused by the switching loss, and the temperature rise to a certain threshold will cause thermal damage to the device. Therefore, if the switching loss of IGBT is modeled and predicted, it is beneficial to the package heat dissipation design and drive circuit design of the device, so as to avoid the failure of the device caused by the excessive temperature rise fluctuation caused by the heat generated by the IGBT switching loss during the operation of the device. Therefore, it is meaningful to predict the switching loss of IGBT. In recent years, some scholars have predicted the switching loss of IGBT and achieved certain results.

At present, there are three methods for calculating IGBT switching loss: switching loss calculation based on physical model, calculation method based on mathematical model and switching loss prediction based on intelligent model. First of all, the switching loss calculation based on the physical model uses the simulation software to simulate the dynamic characteristics of the IGBT, and then calculates the switching loss of the IGBT. The calculation result has high accuracy, but the process of building the model is more complicated and the simulation speed is slow; secondly, For the calculation of switching loss based on the mathematical model, the common method is to refer to the IGBT technical manual to calculate the switching loss, but the calculated value is very different from the actual value. Although the prediction accuracy of the polynomial model is improved, the prediction speed is relatively slow; The switching loss prediction of the intelligent model has improved prediction accuracy and prediction speed compared with the former two, but in the aspect of parameter selection of the intelligent model, if the parameter selection is improper, the intelligent model will fall into the local optimal solution, which is not conducive to the model to find the global optimal solution. optimal solution.

Therefore, the prediction accuracy of the existing measurement methods is not high, and the optimal intelligent model parameters are selected to predict the IGBT switching loss. The invention takes the extreme learning machine as the theoretical main body, fills in the optimal point set, krill swarm algorithm and cosine control factor, takes the collector current, DC bus voltage, switching frequency and gate voltage as the input, and uses the turn-on loss, turn-off Loss for the output establishes a mathematical model for the prediction of switching losses during IGBT dynamics. The invention takes multi-variable input, fully considers various factors affecting the switching loss of the IGBT, improves the prediction accuracy of the switching loss, and is not easy to fall into the local optimal solution, so the invention has certain practical value.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a modeling method for switching loss prediction in the dynamic process of IGBT, and obtain the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module based on the dynamic characteristic test of the IGBT power module. Data, and its turn-on and turn-off loss data, take DC bus voltage, collector current, gate voltage and switching frequency data as input, and use turn-on and turn-off loss data as output to establish the optimal limit of Jiadian krill swarm algorithm The learning machine (GKH-ELM) is a multi-parameter IGBT switching loss prediction model supported by theory; compared with the existing methods, after the optimization of the good point set algorithm, the krill swarm algorithm becomes the good point krill swarm algorithm. When the algorithm is initialized, the individual distribution of Jiadian krill is more uniform, which is conducive to finding the global optimal solution; the introduction of cosine control factor and Levy flight makes the parameters of the Jiadian krill swarm algorithm gradually speed up in the early and middle stages, and the It gradually decelerates in the later stage, and gradually tends to global convergence, so as to realize the dynamic search of krill swarms at the best point. The proposal of GKH-ELM makes up for the lack of prediction accuracy of the existing model, and has important practical significance for improving the reliability of IGBT operation.

The technical scheme adopted by the present invention to solve the technical problem is: a modeling method for switching loss prediction in the dynamic process of IGBT, which is a method for optimizing the IGBT switching loss prediction method of extreme learning machine based on Jiadian krill swarm algorithm, and the steps are as follows :

Step 1: Obtain IGBT dynamic characteristic test data

(1.1) Obtain m groups of test data through the IGBT dynamic characteristic test, each group of data includes the DC bus voltage, collector current, gate voltage and switching frequency data of the IGBT module, and its turn-on and turn-off loss data;

Step 2: Normalize and distribute the test data of IGBT dynamic characteristics

(2.1) Use the normalization formula (1) to normalize the test data of IGBT dynamic characteristics:

(2.2) Divide the normalized test data into learning data and test data, and the distribution ratio is:

The number of learning data: the number of test data = A: B;

Therefore, the m groups of dynamic characteristic test data of IGBT are divided into m×A/(A+B) groups of learning data and m×B/(A+B) groups of test data; each group includes six variables, the turn-on of the IGBT module , turn-off losses, and the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module;

Step 3: Set and initialize extreme learning machine parameters

The parameters of the extreme learning machine need to set the number of nodes in each layer and initialize the weights and thresholds of the extreme learning machine, including:

Set the number of input layer nodes of the extreme learning machine to num_in;

Set the number of hidden layer nodes of the extreme learning machine to num_hid;

Set the number of output layer nodes of the extreme learning machine to num_out, although the prediction is the turn-on or turn-off loss of the IGBT, num_out=1 due to the prediction respectively;

Step 4: Set the basic parameters of the krill swarm algorithm, and complete the assignment of weights and thresholds from the input layer of the extreme learning machine to the hidden layer nodes

(4.1) The parameters of the krill swarm algorithm need to set the number of krill in the initial krill swarm, the maximum number of iterations for krill swarm optimization, the dimension of the data contained in each krill and the initial data of the krill swarm, including:

Set the number of krill in the initial krill group to n;

Set the maximum number of iterations for krill swarm optimization to Iteration_max;

Set the dimension of the data contained in each krill to dim; dim is determined by the extreme learning machine input layer node and hidden layer node, dim=num_in×num_hid+num_hid;

Initialize the population data of the krill swarm and use this as the weights and thresholds of the extreme learning machine input layer to the hidden layer nodes. The method is: use MATLAB software to generate a set X _n (i) containing n points in the dim-dimensional unit space by using the method of dividing the circle domain. The construction formula of X _n (i) is shown in formula (2):

X _n (i)={[(r ₁ ×i),(r ₂ ×i),...,(r _j ×i),...,(r _dim ×i)], 1≤i≤n , 1≤j≤dim} (2)

并且(r_j×i)为r_j×i的小数部分，q是满足2×dim+3≤q的最小素数。通过公式(2)得到佳点矩阵X：In the formula, dim represents the data dimension of each good point;

And (r _j ×i) is the fractional part of r _j ×i, and q is the smallest prime number satisfying 2×dim+3≤q. The optimal point matrix X is obtained by formula (2):

In the matrix X, x _ij ∈ [L _b , U _b ] (1≤i≤n, 1≤j≤dim), and U _b and L _b are the upper and lower boundaries of the variable x _ij , respectively. Take the good point matrix X as the first generation krill population data matrix in the krill swarm algorithm, and record it as the good point krill group. Therefore, in formula (3), X ₁ , X ₂ , ..., X _i , ..., X _n represent the first, second, third, ..., i, ... , n good-point krill, x _ij represents the data on the jth dimension on the i-th good-point krill. Therefore, the weights and thresholds of the extreme learning machine input layer to the hidden layer nodes are initialized.

(4.2) Considering the number of nodes in the input layer and hidden layer of the extreme learning machine, the dim-dimensional data of each Jiadian krill is allocated, and the [1, num_in×num_hid] dimensional data of Jiadian krill is used as extreme learning. The weights from the input layer of the machine to the hidden layer, take the [num_in×num_hid+1, dim] dimension data of the krill from the extreme learning machine as the threshold from the input layer to the hidden layer of the extreme learning machine, and import the above-distributed data into extreme learning in the machine. So far, the parameter setting of the extreme learning machine is basically completed.

Step 5: Optimize the extreme learning machine using Jiadian krill swarm algorithm, and calculate the fitness value of each krill

(5.1) Establish a prediction model of Jiadian Krill Swarm Algorithm-Extreme Learning Machine, called GKH-ELM prediction model, to predict IGBT turn-on loss or turn-off loss. Take the data of each good point krill in the good point krill swarm algorithm as a set of weights and thresholds from the input layer node to the hidden layer node of the GKH-ELM prediction model; (A+B) group of learning data is imported into the extreme learning machine, and the learning data is divided into two parts: input data and output data. A set of input data includes 4 variables, namely the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module, and the output data is the turn-on loss or turn-off loss of the IGBT module. After importing the learning data into the extreme learning machine, n groups of m×A/(A+B) IGBT turn-on loss or turn-off loss prediction values for each group will be obtained, and the The fitness value, that is, the prediction performance evaluation value of the extreme learning machine corresponding to each good point krill,

In formula (4), F is the fitness value of each good point krill, and the larger the fitness value, the better; p _k is the data obtained from the IGBT dynamic characteristic test, which is regarded as the difference between the turn-on loss or turn-off loss of the IGBT. Actual value; p′ _k is the predicted value of IGBT turn-on loss or turn-off loss; N is the predicted output number of extreme learning machine turn-on loss or turn-off loss, that is, N is m×A/(A+B) during the learning period. Similarly, during testing, N=m×B/(A+B).

则记录佳点磷虾i第t次运动的位置数据；若

则记录当代最优佳点磷虾位置数据，并记为历史最优佳点磷虾。(5.2) Screen Jiadian krill with high fitness, and record and retain the position data of Jiadian krill with the largest fitness value in the Jiadian krill swarm algorithm. The selection of good-point krill with high fitness includes: screening the best-fit krill in the contemporary era, and comparing the fitness values of the contemporary best-point krill and the historical best-point krill, and selecting the optimal one. That is, if

Then record the position data of the t-th movement of Jiadian krill i; if

Then record the position data of the krill at the best point of the present time, and record it as the best point of krill in history.

Step 6: Improve the good point krill swarm algorithm to find the global optimal solution

(6.1) Update the good point krill group X obtained in step 4 or formula (13). The position update motion of Jiadian krill individual mainly includes induced motion I _i , foraging motion Fi and physical diffusion motion P _{i ,} the formulas are shown in (5), (6), (7) respectively:

表示第i只佳点磷虾的第t次诱导运动，同理

表示第i只佳点磷虾的第t-1次诱导运动；I^max表示最大诱导速度，取I^max＝0.01；

表示周围佳点磷虾第t次对第i只佳点磷虾产生的影响，而

表示历史最优佳点磷虾对佳点磷虾i产生的诱导作用；ω_I为诱导运动的惯性权重，范围是[0，1]。In formula (5),

Represents the t-th induced motion of the i-th best point krill, and the same is true

Indicates the t-1th induced motion of the i-th best point krill; ^Imax represents the maximum induction speed, taking ^Imax =0.01;

represents the t-th impact of the surrounding krill at the best spot on the i-th best spot krill, and

Indicates the induction effect of the historical optimal point krill on the good point krill i; ω _I is the inertia weight of the induced motion, the range is [0, 1].

表示当前食物对第i只佳点磷虾的诱导强度，

表示历史最优佳点磷虾对佳点磷虾i运动的觅食牵引力；v_F表示磷虾的觅食速度，取v_F＝0.02；ω_F为觅食运动的诱导权重，其范围是[0，1]。In formula (6), F _i ^t represents the t-th foraging movement of the i-th best spot krill, and similarly F _i ^t-1 represents the t-1th foraging movement of the i-th best spot krill;

represents the induction intensity of the current food on the i-th best spot krill,

is the foraging traction force of the krill at the best point in history to the motion of krill i at the best point in history; v _F is the foraging speed of krill, taking v _F = 0.02; ω _F is the induction weight of the foraging movement, and its range is [ 0, 1].

表示第i只佳点磷虾的第t次物理扩散运动；P_max表示佳点磷虾扩散的最大速度，取P_max＝0.005；t表示当前运动次数；δ_i为当前佳点磷虾的随机扩散的方向向量，且δ_i∈[0，1]。In formula (7),

Represents the t-th physical diffusion movement of the _i -th krill in the best spot; _Pmax represents the maximum speed of the krill diffusion in the best spot, taking _Pmax = 0.005; t represents the current number of movements; the direction vector of the diffusion, and δ _i ∈ [0, 1].

见式(8)，以及其第t次运动结束后的位置见公式(9)：Therefore, the variation of the positional movement of Jiadian krill i during the t-th movement

See formula (8), and its position after the t-th motion is shown in formula (9):

作为第t+1次运动的初始位置。Note the position of Jiadian krill i after the t-th exercise

as the initial position of the t+1th movement.

(6.2) Introduce cosine control factor and Levy flight to improve the position update formula of krill swarm in good point. The cosine control factor is used to improve ω _I and ω _F in the krill induced movement formula (5) and foraging movement formula (6), and the improved formula is shown in formula (10).

Levi flight is used to improve formula (9) and set Levi flight. If the optimal fitness value of the krill swarm in Jiadian does not change after more than 20 iterations, the Levy flight is started until the optimal fitness value of the krill swarm changes. See (11) for the Levi flight formula:

Among them, both u and ε obey the standard normal distribution, and β=1.5, and the calculation formula of φ is shown in formula (12).

根据公式(13)进行位置更新，生成新的位置。Therefore, combining formulas (9), (10), (11), and (12), the improved Jiadian krill position update formula (13) is obtained, so the position of each individual in the Jiadian krill group X is

The position is updated according to formula (13) to generate a new position.

Step 7: It is judged whether the current number of motions t of the krill swarm has reached the maximum number of iterations Iteration_max set by the krill swarm algorithm. If t<Iteration_max, go back to step five; if t=Iteration_max, go to step eight.

Step 8: Output the prediction results of IGBT turn-on loss and turn-off loss

The historical optimal krill location data obtained in step 5 is used as the optimal weight and threshold of the GKH-ELM prediction model; Test data, each group of test data includes: the turn-on or turn-off loss of the IGBT module, and the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module; The pole voltage and switching frequency are used as the input data of the GKH-ELM prediction model, and the IGBT turn-on or turn-off loss value is predicted; the turn-on or turn-off loss data obtained from the IGBT dynamic characteristic test is used as the GKH-ELM prediction IGBT turn-on or turn-off loss value The reference value of , and its fitness value is calculated by formula (4), that is, the prediction performance evaluation value of the extreme learning machine under the krill position of the optimal point is calculated.

With the help of MATLAB software, the comparison chart of the turn-on and turn-off loss values of the IGBT predicted in step 8 and the actual turn-on and turn-off loss values, as well as the predicted rms of the turn-on and turn-off loss values of the IGBT test data, are displayed on the display screen of the computer. Error RMSE and coefficient of determination R ² .

The above-mentioned modeling method of switching loss prediction in the dynamic process of IGBT, the IGBT parameters, such as: DC bus voltage, collector current, gate voltage, switching frequency and its turn-on and turn-off losses of the IGBT module are: well known to those skilled in the art.

The above-mentioned modeling method for switching loss prediction in the dynamic process of IGBT, the good point set algorithm, the krill swarm algorithm, the extreme learning machine, the cosine control factor and the Levy flight formula are the prior art and are the present technology. well known to those skilled in the art.

The above-mentioned modeling method of switching loss prediction in the dynamic process of the IGBT, the input method of inputting the parameter data obtained by the IGBT dynamic characteristic test into the computer is a well-known method, and the computer, the display and the MATLAB computer software are all obtained. Commercially obtained.

The beneficial effects of the present invention are: compared with the prior art, the present invention has the following characteristics,

(1) The present invention adopts the GKH-ELM prediction model constructed by optimizing the extreme learning machine based on Jiadian krill swarm algorithm to predict the turn-on and turn-off losses of the IGBT, and realizes the high-precision and high-precision turn-on and turn-off losses of the IGBT. Reliable prediction is helpful for engineers to improve the cooling system of IGBT modules, etc., and improve the reliability of IGBT operation;

(2) The IGBT switching loss prediction model of the extreme learning machine based on the optimal point krill swarm algorithm is established in the present invention, and the krill swarm algorithm initialized by the good point set algorithm is used as the carrier to optimize the optimal weights and thresholds of the extreme learning machine. Perform global search, and introduce cosine control factor and Levy flight formula to realize dynamic search for optimal solution;

(3) The present invention is a modeling method for predicting switching loss in the dynamic process of IGBT. The method has strong scalability and compatibility, and can not only expand the factors affecting the switching loss of IGBT, but also can be extended to other fields , such as relay life prediction, photovoltaic power generation prediction, etc.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

FIG. 1 is a schematic flow chart of the method of the present invention.

Figure 2 is a block diagram of the model for the prediction of switching losses during IGBT dynamics.

FIG. 3 is a display diagram showing the reliability prediction of the turn-on loss of the IGBT according to the present invention.

FIG. 4 is a display diagram showing the reliability prediction of the turn-off loss of the IGBT according to the present invention.

Detailed ways

Fig. 1 shows the process flow of a modeling method for switching loss prediction in the dynamic process of an IGBT according to the present invention: start→acquisition of test data of IGBT dynamic characteristics→distribution and normalization of test data→setup and initialization of extreme learning machine parameters→setup krill Basic parameters of the swarm algorithm → Initialize the initial value of the krill swarm by using the good point set algorithm → Calculate the fitness value of the krill group in the good point, and record the position data of the current optimal and historical optimal krill → Update using the improved position formula The position of Jiadian krill → determine whether the current number of iterations of Jiadian krill has reached the maximum number of iterations, if not, continue to calculate the fitness value of Jiadian krill group, and update the position of Jiadian krill; , then execute the next step→output the IGBT switching loss prediction result→end.

Figure 2 shows a block diagram of the model for the prediction of switching losses during IGBT dynamics. This model takes the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module as the input, and takes the IGBT turn-on and turn-off losses as the output, takes the extreme learning machine as the main body, and uses the krill swarm initialized by the good point set algorithm. Taking the algorithm as the carrier and the cosine control factor and the Levy flight formula as the wings, an IGBT switching loss prediction model based on the optimal extreme learning machine based on the Jiadian krill swarm algorithm was established.

Example

The invention uses a PC as a platform for model building, wherein the CPU is i5-3230M 2.60GHz, the installed memory is 4GB, the operating system is Windows 7-64 bits, and the MATLAB R2016a version is used. The IGBT module is MMG75S120B6HN from Macmic Company. The rated value of the module is 1200V/75A. The module includes two identical IGBT chips and freewheeling diodes, and the distance between the IGBT chip and the FWD chip is 6.4 mm.

Step 1: Obtain IGBT dynamic characteristic test data

(1.1) Obtain 240 sets of test data through IGBT dynamic characteristic test, each set of data includes DC bus voltage, collector current, gate voltage and switching frequency data, as well as turn-on and turn-off loss data of IGBT modules;

Step 2: Normalize and distribute the test data of IGBT dynamic characteristics

(2.1) Use formula (1) to normalize the IGBT characteristic test data;

(2.2) Divide the normalized test data into learning data and test data, and the distribution ratio is:

The number of learning data: the number of test data = 8:2;

Therefore, the 240 sets of dynamic characteristic test data of IGBT are divided into 192 sets of learning data and 48 sets of test data; each set of data includes six variables, the turn-on and turn-off losses of the IGBT module, and the DC bus voltage and collector of the IGBT module. current, gate voltage and switching frequency;

Step 3: Set and initialize extreme learning machine parameters

Set the number of input layer nodes of the extreme learning machine num_in=4;

Set the number of hidden layer nodes of the extreme learning machine num_hid=7;

Set the number of output layer nodes of the extreme learning machine num_out=1;

Step 4: Set the basic parameters of the krill swarm algorithm, and complete the weight and threshold assignment from the input layer of the extreme learning machine to the hidden layer nodes

set the number of krill in the krill group n=50;

Set the maximum number of iterations for krill group optimization Iteration_max=100;

Set the dimension of the data contained in each krill dim=num_in×num_hid+num_hid=4×7+7=35;

Initialize the krill swarm location data as the weights and thresholds of the extreme learning machine input layer to the hidden layer nodes. The method is as follows: use MATLAB software to generate a set X ₅₀ (i) containing 50 points in a 35-dimensional unit space by using the method of dividing the circle domain. Its construction formula is shown in formula (2). When using the formula (2) to construct the optimal point matrix X, the upper and lower boundaries U _b and L _b of q and x _ij need to be determined. Since q≥2×35+3=73, and q is the smallest prime number within this range , so q is set to 73; U _b and L _b are assigned as 1 and -1 respectively; and then the good point matrix X is obtained, which is used as the first-generation population position data matrix of the krill group, and is recorded as the good point krill group.

Based on the number of extreme learning machine input layer and hidden layer nodes, the 35-dimensional data of each Jiadian krill is allocated, and the [1, 28] dimensional data of Jiadian krill is used as the input layer of the extreme learning machine to the hidden layer. The weight of the layer, the [29, 35] dimensional data of Jiadian krill is used as the threshold from the input layer of the extreme learning machine to the hidden layer. So far, the parameter settings of the extreme learning machine are all completed.

Step 5: Optimize the extreme learning machine using Jiadian krill swarm algorithm, and calculate the fitness value of each Jiadian krill

(5.1) Establish a prediction model based on Jiadian krill swarm algorithm-extreme learning machine, called GKH-ELM prediction model, to predict IGBT turn-on or turn-off loss. The prediction process is as follows: take each good point krill data in the good point krill swarm algorithm as a set of weights and thresholds from the input layer node to the hidden layer node of the GKH-ELM prediction model; import 192 sets of learning data into the limit In the learning machine, the learning data includes: input data and output data. A set of input data includes: DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module, and the output data is the turn-on loss or turn-off loss of the IGBT module. After importing the learning data into the extreme learning machine, the turn-on loss or turn-off loss value of 50 groups of 192 IGBTs in each group will be obtained, and the fitness value of each good point krill is calculated by formula (4), that is, each good point The performance evaluation value of extreme learning machine under krill.

则记录佳点磷虾i第t次运动的位置数据；若

则记录当代最优佳点磷虾的位置数据。(5.2) Screen Jiadian krill with high fitness, and record and retain the position data of Jiadian krill with the largest fitness value in the Jiadian krill swarm algorithm. The selection of good-point krill with high fitness includes: screening the best-point krill at the contemporary optimal fitness level, and comparing the fitness values of the contemporary optimal-point-point krill and the historical optimal-point krill, and selecting the optimal one. That is, if

Then record the position data of the t-th movement of Jiadian krill i; if

The position data of the contemporary best point krill is recorded.

In step 6, the position of Jiadian krill is updated using the improved Jiadian krill group position update formula (13).

Step 7: Determine whether the current number of movements t of the Jiadian krill swarm has reached the maximum number of iterations of 100 set by the Jiadian krill swarm algorithm. If t<100, go back to step five; if t=100, go to step eight.

Step 8: Output the prediction results of IGBT turn-on loss and turn-off loss

The krill location data at the historical optimum point obtained in step 5 is used as the optimal weight and threshold of the GKH-ELM prediction model; 48 sets of data in the 240 sets of dynamic test characteristic test data are used as test data, and each set of test data includes : Turn-on or turn-off loss of the IGBT module, and the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module; take the DC bus voltage, collector current, gate voltage and switching frequency of the IGBT module as GKH- The input data of the ELM prediction model, and predict the IGBT turn-on or turn-off loss value; the turn-on or turn-off loss data obtained from the IGBT dynamic characteristic test is used as the reference value for GKH-ELM to predict the IGBT turn-on or turn-off loss value, and use the formula ( 4) Calculate its fitness value, that is, calculate the prediction performance evaluation value of the extreme learning machine under the optimal point of krill.

With the help of MATLAB software, the IGBT turn-on and turn-off loss values predicted in step 8 and the actual turn-on and turn-off loss values are displayed on the display screen of the computer, as well as the root mean square error RMSE=0.3772, The coefficient of determination R ² =0.9984; the root mean square error of turn-off loss prediction RMSE=0.0368, and the coefficient of determination R ² =0.9991.

Table 1 Comparison table of turn-on loss prediction and prediction by the method of the present invention and the existing technical method

Table 2 Comparison table of turn-on loss prediction and prediction using the method of the present invention and the existing technical method

From Table 1, Table ² , it can be seen that the GKH-ELM model outperforms other algorithms in terms of RMSE and R2.

In the above implementation cases, the IGBT parameters, such as the DC bus voltage, collector current, gate voltage, switching frequency, and turn-on and turn-off losses of the IGBT module are well known to those skilled in the art; Point set algorithm, krill swarm algorithm, extreme learning machine, cosine control factor and Levy flight formula are existing technologies and are well known to those skilled in the art; the parameter data obtained by the IGBT dynamic characteristic test is input into the computer The input method in is a well-known method, and the computer, display and MATLAB computer software are all commercially available.

The above is only a better embodiment of the present invention, but the protection scope of the present invention is not limited to this. Equivalent replacements or changes should be included within the protection scope of the present invention.

Claims (2)

1. A modeling method for predicting switching loss in an IGBT dynamic process is characterized in that the method is a method for predicting the IGBT switching loss of an extreme learning machine based on a good krill swarm algorithm, and comprises the following steps:

step one, obtaining IGBT dynamic characteristic test data

(1.1) acquiring 240 groups of test data through an IGBT dynamic characteristic test, wherein each group of data comprises data of direct current bus voltage, collector current, gate voltage and switching frequency of an IGBT module, and data of turn-on and turn-off loss of the IGBT module;

step two, carrying out normalization processing and distribution on the IGBT dynamic characteristic test data

(2.1) carrying out normalization processing on the IGBT dynamic characteristic test data by using a normalization formula (1):

(2.2) dividing the normalized test data into learning data and test data, wherein the distribution proportion is as follows:

the number of learning data and the number of testing data are 8: 2;

therefore, 240 sets of dynamic characteristic test data of the IGBT were divided into 192 sets of study data, and 48 sets of test data; each group comprises six variables, the turn-on and turn-off losses of the IGBT module, and the direct-current bus voltage, the collector current, the gate voltage and the switching frequency of the IGBT module;

step three, setting and initializing extreme learning machine parameters

The parameters of the extreme learning machine need to set the number of nodes of each layer and the weight and threshold of the initialized extreme learning machine, and the method comprises the following steps:

setting the number num _ in of input layer nodes of the extreme learning machine to be 4;

setting the number num _ hid of hidden layer nodes of the extreme learning machine to be 7;

setting the number num _ out of the output layer nodes of the extreme learning machine to be 1;

step four, basic parameters of the krill group algorithm are set, and distribution of weights and thresholds from the input layer of the extreme learning machine to the hidden layer nodes is completed

(4.1) the parameters of the krill group algorithm set:

setting the number n of krill in the krill group to be 50;

setting the maximum Iteration number Iteration _ max of the optimization of the phopenaeus ensis group as 100;

setting the dimension dim to num _ in × num _ hid + num _ hid to 4 × 7+7 to 35 of the data contained in each krill;

initializing the population data of the shrimp groups, and taking the population data as the weight and the threshold value from the input layer of the extreme learning machine to the node of the hidden layer, wherein the method comprises the following steps: using MATLAB software, using a method of rounding to generate a set X containing 50 points in a unit space of 35 dimensions₅₀(i)，X₅₀(i) The structural formula of (2):

X₅₀(i)＝{[(r₁×i)，(r₂×i)，...，(r_j×i)，...，(r₃₅×i)]，1≤i≤50，1≤j≤35} (2)

in the formula

And (r)_jX i) is r_jA decimal part of × i, q is the smallest prime number satisfying 2 × dim +3 ≦ q, which can be calculated as q 73, and the best point matrix X is obtained by equation (2):

in matrix X, X_ij∈[L_b，U_b](1≤i≤50，1≤j≤35)，U_b、L_bAre respectively a variable x_ijUpper part ofThe lower boundary is respectively assigned as 1 and-1; taking the optimal point matrix X as a first generation of the krill group data matrix of the krill group algorithm, and marking as an optimal point krill group; thus, in formula (3), X₁，X₂，...，X_i，...，X₅₀Represents the 1 st, 2 nd, 3 rd.., i., 50 th euphausia superba in the euphausia superba group, x_ijData representing the ith krill in the jth dimension; therefore, the initialization of the weight and the threshold from the input layer of the extreme learning machine to the node of the hidden layer is finished;

(4.2) in consideration of the number of nodes of an input layer and a hidden layer of the extreme learning machine, distributing 35-dimensional data of each superbus, taking [1, 28] dimensional data of the superbus as a weight value from the input layer to the hidden layer of the extreme learning machine, taking [28, 35] dimensional data of the superbus as a threshold value from the input layer to the hidden layer of the extreme learning machine, and introducing the distributed data into the extreme learning machine; at this point, the parameter setting of the extreme learning machine is completely finished;

fifthly, optimizing the extreme learning machine by using the optimal krill group algorithm, and calculating the fitness value of each krill

(5.1) establishing a prediction model of the krill goodpoint algorithm-extreme learning machine, namely a GKH-ELM prediction model, and predicting the turn-on loss or turn-off loss of the IGBT; taking data of each krill in the krill group algorithm as a set of weight and threshold from an input layer node to a hidden layer node of the GKH-ELM prediction model; and (3) importing 192 omics learning data obtained in the step (2.2) into an extreme learning machine, wherein the learning data is divided into two parts: input data and output data; the input data comprises 4 variables which are respectively the direct current bus voltage, the collector current, the gate voltage and the switching frequency of the IGBT module, and the output data is the turn-on loss or the turn-off loss of the IGBT module; after the learning data are imported into the extreme learning machine, 50 groups of predicted values of the turn-on loss or the turn-off loss of 192 IGBTs in each group are obtained, the fitness value of each krill is calculated through a formula (4), namely the predicted performance evaluation value of the extreme learning machine corresponding to each krill,

in the formula (4), F is the fitness value of each euphausia superba, and the larger the fitness value is, the better the fitness value is; p is a radical of_kTaking the data obtained by the dynamic characteristic test of the IGBT as the actual value of the turn-on loss or turn-off loss of the IGBT; p'_kThe predicted value of the turn-on loss or turn-off loss of the IGBT is obtained; n is the predicted output number of the turn-on loss or the turn-off loss of the extreme learning machine, namely N is 192 during the learning period; similarly, during the test, N-48;

(5.2) screening the euphausia superba with high fitness, and recording and retaining the position data of the euphausia superba with the maximum fitness value in an euphausia superba group algorithm; screening of krill with high fitness comprises: screening current generation of superb krill with optimal fitness, and comparing the fitness values of current generation of superb krill with historical optimal superb krill to select optimal, i.e., if

Recording the position data of the t movement of the euphausia superba i; if it is

Recording the position data of the current optimal sweet krill and recording as historical optimal sweet krill;

step six, improving the krill goodpoint algorithm and searching a global optimal solution

(6.1) updating the superbus krill group X obtained in the step four or the formula (13); location updating exercise of individual euphausia superba, mainly comprising induced exercise I_iForaging movement F_iAnd physical diffusion movement P_iThe formulas are respectively shown in (5), (6) and (7):

in the formula (5), the first and second groups,

represents the t-th induced movement of the ith euphausia superba, and the same principle is adopted

Represents the t-1 induced movement of the ith euphausia superba; i is^maxExpressing the maximum induction rate, take I^max＝0.01；

Shows the effect of the t-th time of the peripheral euphausia superba on the ith euphausia superba

Representing the induction effect of the historical optimal krill on the krill i; omega_IInertial weight to induce motion, range is [0, 1 ]]；

In the formula (6), F_i ^tRepresents the t-th foraging movement of the ith euphausia superba, and has the same principle F_i ^t-1Represents the t-1 foraging movement of the ith euphausia superba;

indicating the induction intensity of the current food to the ith krill sweet spot,

representing the foraging traction of the historical optimal krill on the movement of the krill i; v. of_FExpressing the foraging speed of krill, v_F＝0.02；ω_FThe weight for induction of foraging motion is in the range of [0, 1%]；

In the formula (7), P_i ^tRepresents the t-th physical diffusion movement of the ith euphausia superba; p_maxRepresenting the maximum rate of diffusion of the krill at the sweet spot, taking P_max0.005; t represents the current number of movements; delta_iIs the direction vector of the random diffusion of the current sweet krill, and is delta_i∈[0，1]；

Therefore, the amount of change in the positional movement of the euphausia superba i at the t-th movement

See formula (8) and the position after the end of the tth movement see formula (9):

recording the position of the krill i after the end of the tth movement

As the initial position of the t +1 th motion;

(6.2) introducing a cosine control factor and a position updating formula of the Levy flight improvement Kyoho lobster colony, wherein the cosine control factor is used for improving omega in the Kyoho krill induced motion formula (5) and the foraging motion formula (6)_IAnd ω_FThe formula is improved as shown in formula (10);

the Lei-dimensional flight is used for improving the formula (9), and is set, if the optimal fitness value of the krill group exceeds 20 iterations and is not changed, the Lei-dimensional flight is started until the optimal fitness value of the krill group is changed, and the Lei-dimensional flight formula is shown in the formula (11):

wherein u and epsilon both follow a standard normal distribution, and β is 1.5, and the calculation formula of phi is shown in formula (12);

therefore, the modified superba location obtained by combining equations (9), (10), (11), and (12) is updated to equation (13), and thus the location of each individual in the superba group X is updated

Updating the position according to a formula (13) to generate a new position;

step seven, judging whether the current movement times t of the krill groups reach the maximum iteration times set by the krill group algorithm for 100 times; if t is less than 100, returning to the step five; if t is 100, executing step eight;

step eight, outputting the prediction results of the turn-on loss and the turn-off loss of the IGBT

Taking the historical optimal krill position data obtained in the fifth step as the optimal weight and threshold of the GKH-ELM prediction model; using 48 data of 240 sets of dynamic test characteristic test data as test data, wherein each set of test data comprises: the turn-on or turn-off loss of the IGBT module, and the direct current bus voltage, the collector current, the gate voltage and the switching frequency of the IGBT module; taking the direct-current bus voltage, the collector current, the gate voltage and the switching frequency of the IGBT module as input data of a GKH-ELM prediction model, and predicting the turn-on or turn-off loss value of the IGBT; taking the turn-on or turn-off loss data obtained by the IGBT dynamic characteristic test as a reference value of the GKH-ELM prediction IGBT turn-on or turn-off loss value, and calculating the fitness value of the IGBT turn-on or turn-off loss value by using a formula (4), namely calculating the prediction performance evaluation value of the extreme learning machine at the optimal sweet spot krill position;

displaying a comparison graph of the predicted turn-on and turn-off loss values and actual turn-on and turn-off loss values of the IGBT in the step eight, a root mean square error RMSE and a decision coefficient R of the prediction of the turn-on and turn-off loss values of the IGBT test data on a display screen of a computer by using MATLAB software²。

2. The modeling method for predicting switching loss in an IGBT dynamic process according to claim 1, characterized in that: variables affecting the switching loss of the IGBT are fully considered, namely the direct-current bus voltage, the collector current, the gate voltage and the switching frequency of the IGBT module.

Images