Prediction Method of Driving Strategy of High-Power IGBT Module Based on MEA-BP Neural Network

An insulated gate bipolar transistor (IGBT) driver is crucial for improving the reliability of a power electronics system. This paper proposes a method for predicting the optimal driving strategy of high-power IGBT module based on backpropagation neural network optimized by the mind evolutionary algorithm in order to solve the problem of compromise among switching loss, switching time and overshoot and achieve a good driving effect. The three regions of switching transitions are analyzed based on the switching characteristics of the IGBT module. Neural networks are established to predict turn-on and turn-off driving strategies for variable gate resistance active gate driver of the IGBT module. The mind evolutionary algorithm is used to optimize the weights and biases of the neural networks so that the optimal weights and biases can be obtained. In order to verify the effectiveness of the driving strategy prediction method proposed in this paper, experiments are carried out for 4500V/900A the IGBT module. Compared to the conventional gate driver, the predicted driving strategies reduce the turn-on energy loss, turn-on time, over-current, comprehensive evaluation method, turn-on delay time and tail voltage duration by 59.31%, 46.38%, 36.99%, 65.65%, $1.9~\mu \text{s}$ , $2.9~\mu \text{s}$ , respectively. It was also found that the Planar-IGBT turn-off process was rarely affected by the gate resistance. The proposed method in this paper can be used not only for the guidance of the driving strategy determination of high-power the IGBT module driver, but also for the driver circuit improvement in the design process.


I. INTRODUCTION
The semiconductor power device insulated gate bipolar transistor (IGBT) was invented by Dr. Baliga in 1979, and since then, the IGBT have made great contributions in different fields, such as transportation, industry, lighting, consumer electronics, medical treatments, national defense, renewable The associate editor coordinating the review of this manuscript and approving it for publication was Zhilei Yao . energy, and power transmission [1]. The IGBT and its driver have been constantly studied and researched. Under the premise of ensuring the safety and reliability of a chip, reducing the switching loss of the device to improve the energy conversion efficiency is one of the research topics.
During design of a driver, not only the switching characteristics, switching loss, turn-off over-voltage, short-circuit safe operating area (SCSOA), reverse biased safe operating area (RBSOA), and electromagnetic interference (EMI) need to VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ be considered, but also an optimal compromise among the switching overshoot, the switching time and the switching loss need to be achieved [2]. If the compromise problem is not handled well, there will be long switching time, large switching loss, or the module damage. In [2]- [5], active gate driver (AGD) reduced switching loss and overshoot by variable gate current. In [6]- [9], AGD reduced switching loss and overshoot by variable gate resistance. In [10]- [15], AGD reduced switching loss and overshoot by variable gate voltage. However, it was not explained how the parameters of the driving strategy should be selected to achieve the minimum switching energy loss, minimum switching time, minimum peak voltage, and minimum peak current, or comprehensive consider all aspects of the driving effect. Neither did it explain how to achieve the optimal compromise among the switching overshoot, the switching time and the switching loss. Moreover, when designing an AGD, it is important to understand IGBT device characteristics. The IGBT module model can be used to understand the characteristics of the IGBT module. Common physical IGBT module models include the Sheng model [16], Palmer model [17], Hefner model [18]- [21], and Kraus model [22]. These models can describe the steady-state or dynamic characteristics of IGBT chips accurately. Besides, many behavioral models have been proposed [23]- [27]. When the physical model is used, it is necessary to know the internal structure of an IGBT. However, the IGBT model parameters are numerous, so the determination of parameters is very complex [28], which puts higher demands on driver design. Besides, physical models include a large number of complex differential equations, which increases both the computation burden and the simulation time [23]. Moreover, there is a problem of computational convergence, which is particularly prominent in complex multi-IGBT module circuits. In addition, in practical applications, the IGBT module integrates a fast recovery diode (FRD), and there are many parasitic parameters in the module [29]- [33], which makes the IGBT model more complex and costly. Therefore, this paper proposes a simple method to solve the compromise problem and formulate the optimal driving strategy, while exerting the AGD optimal driving effect. With the development of big data in the information technology field, the computer-aided solutions to various problems have been promoted in many fields. Some of the machine learning-assisted methods are the computer-aided prediction of highly-selective catalysts [34], a fault prediction for vehicular networks [35], the discovery of new peptide substrates for enzymes [36], and an assistant somatic mutation detection [37]. In particular, Zeng conducted the virtual measurement of the IGBT module current using machine learning [38]. In [39], the lifetime estimation of the IGBT devices was conducted using machine learning. Also, Oukaour detected the aging defect of the IGBT power module using machine learning [40]. Machine learning-based methods have the ability to find the relationships in data needed to conduct scientific research tasks efficiently. Therefore, this paper takes the three-stage variable gate resistance method as the research object and presents a simple prediction method of driving strategy based on backpropagation (BP) neural network optimized by the mind evolutionary algorithm (MEA). Taking switching time as a constraint condition, the compromise problem between switching loss and overshoot is solved, while achieving the minimum switching energy loss, minimum switching time, and minimum peak current, or comprehensive consider all aspects of the driving effect. The proposed method is simple and does not require knowing the internal structure and model parameters of an IGBT, and the professional requirements for designers are low, thus achieving the following: • reduce delay time at turn-on and total turn-on time; • reduce turn-on di c /dt and associated reverse-recovery effects; • lower tail voltage and turn-on energy loss; • realize more simple turn-on and turn-off variable gate resistance driving circuit.
The rest of the paper is organized as follows. In Section II, the three-stage IGBT module switching process, conventional gate driver, and variable gate resistance method are presented in detail. Subsequently, the Prediction method of the driving strategy is introduced in Section III, and the verification test is presented in Section IV. In Section V, the analysis and discussion are provided. Finally, conclusions are given in Section VI.

II. ANALYSIS OF THE IGBT MODULE SWITCHING CHARACTERISTICS A. THREE-STAGE IGBT MODULE SWITCHING PROCESS
In this paper, the turn-on and turn-off transitions of the IGBT module are both divided into three regions in order to obtain different control purposes based on the successive stages of the switching transient, as shown in Fig. 1. Hence, a threestage AGD technique is improved. The three-stage turn-on transition is as follows. The turn-on stage I lasts t 1 , that is, from the moment gate voltage v ge rises from v goff to the moment of the collector current i c starts to rise. The turn-on stage II lasts t 2 , and in this stage, the current rises from i c to its maximum turn-on current peak I peak . The turn-on stage III lasts t 3 , and in this stage, the current decreases from I peak to the end of the turn-on transition. The three-stage turn-off transition is as follows. The turn-off stage I lasts t 4 ; this stage starts from the gate voltage v ge decreasing from v gon till the collector current i c starts declining. The turn-off stage II lasts t 5 , from the moment i c fails from I L to the moment of the maximum turn-off voltage peak V peak decreasing to bus voltage V dc . Lastly, the turn-off stage III lasts t 6 , and this stage lasts from the end of the turn-off stage II to the end of the turn-off transition.
Gate resistance R g has an effect on the turn-on time t on , the delay time at turn-on t don , overshoot collector current I c , di c /dt, turn-on energy loss E on , turn-off time t off , the delay time at turn-off t doff, overshoot collector-emitter voltage V ce , dv ce /dt, and turn-off energy loss E off [9]. The greater the gate resistance is, the more switching time and switching loss increase, while overshoot decreases. The turn-on time t on and I c are respectively expressed as: (1) where I L denotes the rated current of an IGBT module. Further, the turn-on energy loss E on is given by: where v ce (t) denotes the collector-emitter voltage, and i c (t) represents the collector current. The turn-off time t off , V ce , and turn-off energy loss E off are respectively expressed as: The operation principles of turn-on and turn-off circuits of conventional gate driver (CGD) are shown in Figs. 2(a) and 2(b), respectively [41]. In Fig. 2, Q 1 and Q 2 represent MOSFETs, and R gon and R goff represents gate resistances.
In Fig 2(a), in turn-on transition, Q 1 and R on are used. In Fig 2(b), in turn-off transition, Q 2 and R goff are used.

C. VARIABLE GATE RESISTANCE METHOD
The schematic of a three-stage variable gate resistance driver circuit is shown in Fig. 3. In Fig. 3, Q 1 -Q 6 represent MOS-  FETs, and R g1 -R g6 represents gate resistances. Gate resistances R g1 , R g2 , and R g3 are used in the three stages of the turn-on transition, respectively; and gate resistances R g4 , R g5 , and R g6 are used in the three stages of the turn-off transition, respectively. As previously mentioned, the driving strategy is determined by the MEA-BP neural network prediction model. Then, according to the driving strategy, the controller complex programmable logic device (CPLD) controls Q 1 -Q 6 to realize the driving of the IGBT module. The operation principles of the three stages of the turnon circuit are shown in Figs. 4(a)-4(c), respectively, and the operation principles of the three stages of the turn-off circuit are shown in Figs. 4(d)-4(f), respectively. On the one hand, in turn-on stage I, Q 1 and R g1 are used, in turn-on stage II, Q 2 and R g2 are used, and in turn-on stage III, Q 3 and R g3 are used. On the other hand, in turn-off stage I, Q 4 and R g4 are used, in turn-off stage II, Q 5 and R g5 are used, and in turn-off stage III, Q 6 and R g6 are used. The variable gate resistance method is simply to improve t on , t don , I c , di c /dt, E on , t off , t doff , V ce , dv ce /dt, and E off using resistors with different resistance values in the three regions of turn-on and turn-off transitions. There is a complex nonlinear relationship between variable gate resistance and driving effect of the IGBT. Therefore, the variable gate resistance and driving effect of the IGBT were predicted by the neural network prediction model because it is suitable for nonlinear modeling. In this paper, the variable gate resistance method uses t on as a constraint condition to make both E on and I c smaller without increasing the value of t on . The improvement in I c by the variable gate resistance method denoted the improvements in I peak and di c /dt. In the same way, this method uses t off as a constraint condition to decrease both E off and V ce without increasing the value of t off . The improvement in V ce by the variable gate resistance method is improvements in V peak and dv ce /dt.

III. PREDICTION METHOD
The proposed method is data-driven, and it is intended to help designers improve the variable gate resistance driving method of the IGBT module. This method operates in five steps. The first step is to load the IGBT module double pulse test data. The second step is to preprocess and normalize the data. The third step is to create the neural network first, and then the data is used to train the neural network, generating the network prediction model. In order to reduce the cost of program running time, in the fourth step, the network prediction model is used to traverse the whole region with a large step to find out the region that satisfies a certain condition better than conventional gate driver, thus reducing the solution region. Lastly, in the fifth step, the network prediction model is used to traverse the solution region with a small step to find the optimal driving strategy in the solution region.

A. DOUBLE PULSE TEST DATA
In order to capture the data of the power semiconductor devices related to the IGBT module switch processes, which is then used to create the prediction model, the double pulse test was conducted to generate the data samples of the IGBT module. The double pulse test circuit is shown in Fig. 5(a). The power devices used in the test were 4500V/900A highpower IGBT modules (CM900HG-90H) from Mitsubishi. A film capacitor acted as a DC bus support capacitor. The lower IGBT served as a device under test, and an air-core inductor L served as an inductive load. The photograph of the experimental setup is shown in Fig. 5(b). The parasitic  inductances were measured in combination with the existing method presented in [42]. The detailed parameters of the test bench and measurement equipment are presented in Table 1.
Six resistors of 3.3 , 10 , 12 , 16.5 , 33 , and 47 were used as switching gate resistances. The turn-on and turn-off transitions included six stages (three stages per each transition type), and six resistances were optional in each stage; thus, there were 216 (6 × 6 × 6 = 216) groups of gate resistance combinations. The double pulse test was conducted using these 216 groups of gate resistance combinations.

B. TEST DATA PREPROCESSING AND NORMALIZATION
Because 216 groups of test data were very large, so we need to preprocess them to get the data we need. The 216 groups of test data were preprocessed to obtain t on , E on , I c , t don , di c /dt, t off , E off, V ce , t doff, dv ce /dt, and then the data were normalized. The data were normalized by Equation 7 where x min denotes the minimum value in the data sequence, and the x max denotes the maximum value in the sequence.
C. NEURAL NETWORK DEVELOPMENT AND TRAINING 1) BP NEURAL NETWORK The neural networks were created for the prediction of the turn-on and turn-off effects, and their structures are shown in Figs. 6(a) and 6(b), respectively. In Fig. 6, R g1 , R g2 , and R g3 denote the input parameters used to predict E on , I c , and t on ; and R g4 , R g5 , and R g6 denote the input parameters used to predict E off , V ce , and t off . As shown in Fig. 6, the turn-on and turn-off prediction models were both three-layer neural networks, consisting of an input layer, one hidden layer, and output layer. In the input layer, there were three neurons. In the hidden layer, there were q neurons. Lastly, in the output layer, there were three neurons. As for the connection between the input and hidden layers, w ij denotes the connection weight between neuron i in the input layer and neuron j in the hidden layer, and b j denotes the bias of neuron i. As for the connection between the hidden and output layers, v j denotes the connection weight between two neurons from these layers, and θ denotes the neuron bias. The input of the hidden layer is given by Equation (8). As shown in Equation (9), the sigmoid function was selected as an activation function of neurons in the hidden layer.
The input of the output layer is given by Equation (10). As shown in equations (11), the identity function was selected as an activation function of neurons in the output layer.
The mean squared error (MSE) and absolute error (AE) were used to evaluate the performance of the models, and they are given by Equations (12) and (13), respectively. In Equations (12) and (13), y i denotes the actual value of the i th data, y i denotes the value of the i th data predicted by the neural network, and n is the number of samples.
During the training of the BP neural network, the initial weights and biases were randomly set and adjusted according to the output error value. Namely, when the actual output cannot reach the expected value, the weights and biases are adjusted according to the error between the actual output and the expected value, until the actual output meets the expectation. Besides, the number of neurons in the hidden layer determines network nonlinearity. If the number of neurons is small, the network is under-fitting; on the contrary, a large number of neurons leads to the overfitting when the network nonlinearity is higher than that of the model itself. The number of neurons in the hidden layer is given by Equation (14), where q denotes the number of neurons in the hidden layer, N denotes the number of neurons in the input layer, M denotes the number of neurons in the output layer, m is an integer between 1 and 8. Therefore, the selection of weights and biases will be the focus of BP neural network parameters adjustment.
2) MEA-BP NEURAL NETWORK The mind evolutionary algorithm (MEA) was employed to optimize the BP neural network. The MEA represents the machine learning-based iterative optimization method. In the MEA, all individuals in each iteration of the evolution process are integrated into a population. A population is then divided into several subgroups. Information is exchanged between individuals and subgroups through the billboard. During the convergence process, the individuals within a subgroup compete to become winners. If there is no new winner (i.e., there is no individual with the score higher than those of the other individuals in the group), it is considered that the subgroup is mature. For a subgroup, the period from its creation to its maturity is called the life cycle. In the whole solution space, each subgroup competes to become a winner by constantly detecting new points in the solution space so as to help BP neural network obtain the optimal weights and biases [43]. The flowchart of the BP neural network where the MEA is employed for parameter optimization is shown in Fig. 7. In Fig. 7, in the MEA part, the meaning of the symbols is as follows: S Gi denotes the i th subgroup size, N G denotes the number of subgroups that exist simultaneously in the algorithm; r denotes the proportion of choice in the dissimilation operation, N R denotes the number of subgroups to be released or the number of subgroups to be created; S R denotes the number of individuals released, t i denotes the iteration time of the i th subgroup; lastly, c t i i implies the i th subgroup is at the center of the t th iteration. The pseudocode of the MEA-BP algorithm is given in Algorithm 1. In the pseudocode, Steps 9-13 represent the similartaxis operation, and Steps 3-7 and Steps 14-21 represent dissimilation operation.

3) NEURAL NETWORK TRAINING
After many tests, in the MEA, there were 200 populations, five superior subgroups, and five temporary subgroups; the subgroup size was 20, and the maximal number of MEA iteration was set to ten. The hidden layer of the BP neural network contained eight neurons; the maximal number of iteration was set to 100. The MEA-BP neural network was developed using the Neural Network Toolbox of MATLAB software.
A random function was used to scramble 216 groups of test data, and then these data were divided into training and test datasets containing 151 and 65 groups, respectively. The training and test data were different in order to evaluate the generalization ability of the developed neural network prediction model, i.e., to examine whether the model can predict the results well for unknown gate driving resistance values. After the networks training, the root mean square error of turn-on neural network was found to be 1.21 e −12 , the root mean square error of turn-off neural network was 1.24 e −12 , the maximum absolute error of turn-on neural network was 4.4886 e −6 , and the maximum absolute error of turn-off neural network was 4.0897 e −6 . Figs. 8 and 9 show the best validation performance, indicating that the MSE generally improves as the number of epochs increases. In Fig. 8, it stopped after 15 epochs while the best validation performance occurred at epoch 9. In Fig. 9, it stopped after The relationship between the targets and outputs is measured by the correlation coefficient RR. The RR was found to be 0.98966 for the turn-on neural network, and 0.98232 for the turn-off neural network was, indicating high prediction accuracy of the network and that the network captured the relevant characteristics of the IGBT module switching process well.

D. LARGE-STEP SEGMENTATION OF SOLUTION REGION
The trained network prediction model traversed the whole region by a 1-step. The relationships between gate resistance R g and turn-on parameters are presented in Fig. 12, and the relationships between R g and turn-off parameters are VOLUME 8, 2020  presented in Fig. 13. As shown in Fig. 12, turn-on energy loss E on was in the range of 2-12 J, R g1 had little effect on E on , and E on increased with the increase in the values of R g2 and R g3 ; the value of turn-on time t on was in the range of 2-20 µs, and it increased with the increase in the three resistances. The value of I c was in the range of 200-800 A; the smaller R g2 was, the larger I c was. Further, R g1 and R g3 had little effect on I c , which was consistent with the operational theory of the IGBT module. As shown in Fig. 13, turn-off energy loss E off was in the range of 1.6-1.72 J, and R g4 , R g5 , and R g6 had little effect on E off ; turn-off time t off was in the range of 3-3.4 µs, and R g4 , R g5 , and R g6 also had little effect on t off . Lastly, V ce was in the range of 620-700 V, and R g4 , R g5 , and R g6 had little effect on V ce . The results clearly present the influence of variable gate resistance on the relevant characteristics of the IGBT module. In the turn-off process, with the changes in R g , the ranges of E off , t off , and V ce were small.
The trained network prediction model runs for a long time if the optimal driver strategy is found in the whole region. Therefore, we reduced the solution region based on the driving effect of the CGD. The driving effect of the CGD is shown in Fig. 14. In Fig. 14(a), the radar map of the CGD turn-on transition driving effect (E on , I c , di c /dt, t on , t don ) is presented, and in Fig. 14(b), the radar map of the CGD turnoff transition driving effect (E off , V ce , dv ce /dt, t off , t doff ) is presented. The trained prediction model traversed the whole region, and the region with the driving effect better than that of the CGD was kept, and the region with the driving inferior to the effect of the CGD was eliminated. Namely, each index of the AGD should be better than that of each index of the CGD; for instance, E on of the AGD should be smaller than or equal to E on of the CGD. The evaluation method based on the radar map of the driving effect presented in Fig. 14 is called the comprehensive evaluation method (CEM).
Then, the solution region reduced, as shown in Fig. 15 and Fig. 16. As shown in Fig. 15, E on was in the range of 2.3 -3.65 J, turn-on time t on was in the range of 4 -7 µs, and I c was in the range of 236-580 A. Further, as presented in Fig. 16, E off was in the range of 1.6-1.72 J, t off was in the range of 3-3.4 µs, and V ce was in the range of 620-700 V.

E. SMALL-STEP OPTIMIZING OF DRIVING STRATEGY
According to the results presented in Figs. 15 and 16, the network prediction model found out the driving strategies of the AGD using a small step of 0.1 , and the corresponding distribution maps of the driving strategies are drawn in Fig. 17. In Figs. 17(a) and 17(b), there are three parts, the first part represents the variable gate resistance results of stage I, the second part represents the variable gate resistance results of stage II, and the third part represents the variable gate resistance results of stage III.
In Fig. 17, the ordinate denotes the resistance value, and the abscissa denotes the time. The turn-on strategies are shown in Fig. 17(a), where it can be seen that the resistance ranges of stage I, stage II, and stage III were 3.3-13 , 3.3-39 , and 3.3-15 . Also, the larger the resistance was, the greater the corresponding time was. The turn-off strategies are shown in Fig. 17(b), where it can be seen that the resistance ranges      of all the three stages were the same and equal to 3.3-47 . Namely, with the increase in the resistance, the corresponding time was almost unchanged, and there was no trend similar to that in Fig. 17(a). The times corresponding to all the resistor ranges of the three stages were basically the same. In Fig. 17, we can identify four strategies that are based on: 1) minimum E on or E off , 2) minimum t on or t off , 3) minimum I c or V ce , and 4) CEM driving strategy. In order to balance the driving effect, an optimal compromise between the switching over-voltage, over-current, and the switching loss, should be achieved. The CEM driving strategy refers to the strategy of the minimum area formed by the driving effect in a radar map. The area formed by the driving effect in a radar map can be where α is 72 • , and a i is the coordinate value in the radar map. In the ideal situation, the IGBT switching process is completed in an instant, there is not switching energy loss, and there is no switching time, that is, the driving effect of the ideal IGBT is the center point of the radar map. So the smaller the area is, the closer the situation to the ideal situation will be, and all of the parameters will be more balanced. The neural network prediction model helped us identify these four strategies and provided variable gate resistance strategies, turn-on time t on and turn-off time t off , so that t 1 , t 2 , t 3 could be easily obtained by the double pulse test. Take the CEM driving strategy as an example. In Step 1, the resistance of 3.3 was used as a gate resistance to drive the IGBT module, and t 1 was obtained using the test data. In Step 2: the resistance of 3.3 and t 1 were used as the turn-on stage I driving strategy, and then a 37 resistance was used in Stage II and III, and t 2 was obtained using the test data.
The results of the turn-on and turn-off strategies are shown in Table 2 and Table 3, respectively. It should be noted that the strategies based on minimum E on and minimum t on provided the same results, as presented in Table 2. The turnoff strategies provided are all the same results, are presented in Table 3.

IV. EXPERIMENTAL VERIFICATION
In order to evaluate the performance of the proposed strategy obtained by the MEA-BP neural network-based prediction model, the experiments were conducted. As already mentioned, the experimental platform is shown in Fig. 5, and the detailed parameters of the test bench and measurement equipment used in the test are presented in Table 1.

A. TURN-ON STRATEGY VERIFICATION
In order to verify the feasibility of the proposed AGD, the DC bus voltage was 2800 V, and the switched current was 900 A. The experimental turn-on waveforms of the IGBT module obtained by the proposed AGD are presented in Fig. 18. The experimental turn-on radar map is presented in Fig. 19, where it can be seen that the driving effect of the proposed AGD was better than that of the CGD. The driving effects of different AGD strategies were different. The values of I c , di c /dt, t don , t on , and E on of three AGD methods were all lower than those of the CGD. Compared to the CGD, the AGD reduced the value of t don and tail voltage duration by 1.9 µs and 2.9 µs, respectively. On the one hand, E on and t on of the strategies based on minimum E on and minimum t on were the smallest. On the other hand, I c of the strategy based on minimum I c was the smallest. Lastly, the driving effect of the CEM was more balanced than other strategies.
More turn-on waveforms of the IGBT module under different operation temperatures are shown in Fig. 20. In the experiment of different operation temperatures, the DC bus voltage was 2800 V and the switched current was 900 A. All these parameters are consistent with 25 • . It can be seen from Fig. 20 that the impact of the operation temperature on the voltage and current slopes is limited since the performance of the IGBT module varies little with different operation temperature, except that the I peak changed obviously. The higher the operation temperature, the lower the I peak . In addition, the turn-on delay time varies little with temperature. Therefore, the performance of the AGD will not be affected under different operation temperatures.

B. TURN-OFF STRATEGY VERIFICATION
As presented in Table 3, in the turn-off strategy obtained by the MEA-BP neural network-based prediction model, all resistances were 10 . In order to verify the accuracy of the  MEA-BP neural network, six driving strategies were chosen. In the first strategy (called AGD I), R g4 was 3.3 , R g5 was 13 , and R g6 was 3.3 . In the second strategy (called AGD II), R g4 was 3.3 , R g5 was 37 , and R g6 was 3.3 . In the third strategy (called AGD III), R g4 was 3.3 , R g5 was 39 , and R g6 was 3.3 . In the fourth strategy, the gate resistance was 3.3 . In the fifth strategy, the gate resistance was 16.5 . Lastly, in the sixth strategy, the gate resistance was 47 . The experimental verification was carried out at the DC bus voltage of 2800 V and the switched current of 900 A. The experimental turn-off waveforms of the IGBT module are presented in Fig. 21. The experimental turn-off radar map of the IGBT module is presented in Fig. 22, where it can be seen that when the gate resistance changed, the turn-off effects of six driving strategies were the same as that of the CGD; thus, the turn-off result predicted by the MEA-BP neural network prediction model was correct, and there was no need to use the variable gate resistance method during the turn-off transition.
The turn-off waveforms of the IGBT module under different operation temperatures is shown in Fig. 23. The DC bus voltage was 2800 V and the switched current was 900 A. According to the turn-off strategy obtained by the prediction model, the gate resistance was 10 . In the tests conducted at different temperatures, the turn-off delay time increases as the operation temperature rises, while the voltage and current slopes maintain almost constant. Similar to the turn-on transient, only the V peak changed obviously under different operation temperatures. The higher the operation temperature, the lower the V peak . However, the performance of the turn-off strategy will not be affected under different operation temperatures since the gate resistance did not change during turn-off.

V. ANALYSIS AND DISCUSSION
Adding a suitable gate capacitance C ge to the driver circuit can improve the driving effect, but this is not studies in this work. Namely, we compared the AGD and CGD without adding C ge . Since C ge was added to the AGD, the driving   effect of the driving strategy predicted by the MEA-BP neural network could not be clearly determined.

A. PREDICTION ACCURACY ANALYSIS
The prediction accuracy of the MEA-BP neural network is very important. The predicted and experimental turn-on results are presented in Fig. 24, where it can be seen that there were small differences between the predicted and experimental results. According to the obtained results, the improvement was significant, namely: • I c of the strategy based on minimum I c was 59.31% smaller than that of the CGD.
• t on of the strategy based on minimum t on was 46.38% less than that of the CGD.
• E on of the strategy based on minimum E on was 36.99% smaller than that of the CGD.
• the driving effect of the CEM driving strategy was 65.65% smaller than that of the CGD. In Fig. 21, the over-voltage (V peak ) is determined using the stray inductance L s and di c /dt as follows: As shown in Fig. 21, di c /dt values of seven driving strategies were almost the same as that of the CGD, so V peak values were also almost the same. In Fig. 22, it can be seen that the turn-off results predicted by MEA-BP neural network were accurate. The driving effect at a different gate resistance R g was mostly the same.

B. ELECTROMAGNETIC INTERFERENCE ANALYSIS
In order to have an understanding of the EMI induced by high di c /dt and dv ce /dt, EMI analysis of the voltage and current waveforms were carried out, respectively. The approximation of the spectrum for both V ce and I c are shown in Figs. 25 and 26. These data were obtained by the oscilloscope Tektronix MDO4104-3 in the experiment and the spectrum was obtained applying the FFT in MATLAB software after the data were processed. The results show that the AGD did not cause additional EMI generation.

C. DRIVING STRATEGY ANALYSIS
In the turn-on transition of the IGBT module, the fast recovery diode was turning off, so the turn-on strategy was affecting VOLUME 8, 2020 the turn-off process of the fast recovery diode. By making the fast recovery diode turn-off softer, the loss can be reduced.
In Fig. 17(a), for the statistical turn-on strategies, it can be seen that a small resistance was used in stage I with the aim of reducing t don ; a large resistance was used in stage II with the aim of reducing di c /dt and I peak ; and lastly, a small resistance was used in stage III, in order to make the IGBT enter the saturation region quickly.
There are two types of IGBT chips: Trench-IGBT and Planar-IGBT. The turn-off process of a Trench-IGBT is not affected by gate resistance, but the turn-off process of a Planar-IGBT is affected by gate resistance [1], [6]- [9]. The IGBT module used in this paper is the Planar-IGBT, but the turn-off strategies that are presented in Fig. 17(b) show that in the experiments, the gate resistance had little effect on the module turn-off process. The experimental results presented in Figs. 21 and 22 prove the same. After the confirmation from the chip manufacturer, it has been concluded that the IGBT chip technology has constantly been developing and that the semiconductor manufacturer had made the best opti- mization of the device, so there is no longer a need to improve the IGBT turn-off characteristics through the gate resistance. In Table 2, the same resistances are used for stages I and III of the turn-on strategies. Such a variable gate resistance drive circuit can be optimized in the way presented in Fig. 27(a), and the corresponding driver optimization is illustrated in Fig. 27(b). The operation principle of the three-stage turn-on circuit is shown in Fig. 28. In turn-on stages I and III, Q 1 and R g1 were used. Thus, due to driver optimization, the hardware cost is reduced.

VI. CONCLUSIONS AND FUTURE WORK
A prediction method of driving strategy of high-power IGBT module based on MEA-BP neural network is proposed in this paper to improve the variable gate resistance method of active gate driver. The switching process of the IGBT module is analyzed, of which the turn-on and turn-off transitions of the IGBT module are both divided into three regions in order to obtain different control purposes based on the successive stages of the switching transient. The prediction networks for turn-on and turn-off driving strategies of variable gate resistance active gate driver are established, in which the gate resistances in three-stages are input, and switching loss, switching time, overshoot are output. In order to obtain the optimal weights and biases, the mind evolutionary algorithm is used to optimize the BP neural network. The prediction method uses switching time as a constraint condition to make both switching loss and overshoot smaller without increasing the switching time. In order to verify the effectiveness of the driving strategy prediction method proposed in this paper, experiments are carried out for a 4500V/900A IGBT module. The experimental results show that the driving effect of the proposed driving strategies is significant. Compared to the CGD, the AGD reduces the value of t don and tail voltage duration by 1.9 µs and 2.9 µs, respectively. The over-current I c of the strategy based on minimum I c is 59.31% smaller than that of the CGD. The turn-on time t on of the strategy based on minimum t on is 46.48% less than that of the CGD. The turnon energy loss E on of the strategy based on minimum E on is 36.99% smaller than that of the CGD. Lastly, the driving effect of the CEM driving strategy is 65.65% smaller than that of the CGD.
A comprehensive evaluation method for driving effect is proposed in this paper. This method not only assists the neural network to find the optimal strategy but also determines the driving effect of the driver intuitively. It is found that the Planar-IGBT turn-off process is almost not affected by the gate resistance, which shows that due to the development of the IGBT chip technology, there is no need to use the gate resistance to improve the turn-off effect. Thus, a simpler turn-on and turn-off variable gate resistance driving circuit is realized.
A prediction method proposed in this paper is simple and versatile and can be used as a reference for further driver improvements. It can be used not only for the guidance of the driving strategy determination of high-power IGBT module driver, but also for the driver circuit improvement in the design process.
In our future work, we will use machine learning algorithms to improve the other driving methods of the IGBT and SiC MOSFET modules.