Power semiconductor transient simulation accuracy optimization method and system based on machine learning

Abstract

The present invention belongs to the field of power electronics simulation technology, and discloses a method and system for optimizing the transient simulation accuracy of power semiconductors based on machine learning. The method uses a lightweight convolutional neural network to establish a mapping between a low-precision waveform and a high-precision waveform. The low-precision waveform is processed by the neural network and then linearly interpolated to obtain a high-precision waveform with higher accuracy than direct iterative calculation at a lower cost of computing resources and time. Taking the IGBT device-level simulation model as an example, the present invention successfully applies the transient simulation accuracy optimization method to significantly improve the accuracy of its switching transient waveform. The present invention effectively solves the contradiction between simulation accuracy and computing resource consumption, simulation time, and the contradiction between precision and accuracy, and the problem that real-time simulation accuracy is difficult to further improve. It provides reliable and efficient technical support for improving the accuracy of power electronics simulation and opens up new ideas for improving the accuracy of device-level real-time simulation.

Description

Technical Field

[0001] The present invention belongs to, but is not limited to, the field of power electronics simulation technology, and in particular relates to a method and system for optimizing power semiconductor transient simulation accuracy based on machine learning. Background Art

[0002] With the rapid development of power electronics technology, the operating frequency of power semiconductor devices continues to increase. High-frequency switching devices, represented by SicMOSFET, have extremely short switching transient processes, even with rapid voltage and current changes at the nanosecond level. This makes it an urgent core technical requirement in power electronics simulation to accurately describe the transient characteristics of high-frequency switching devices (such as oscillations, overshoots, spikes and other non-ideal phenomena) and accurately predict switching losses.

[0003] Simulation models for power semiconductor devices can be categorized into two main categories based on their sophistication and complexity: system-level and device-level. System-level models, exemplified by average models, ideal switch models, binary resistance models, and constant admittance L / C models, focus solely on the effects of steady-state switching on power semiconductor devices, while ignoring transient switching processes. In contrast, device-level models describe the transient characteristics of voltage and current during switching and can be further categorized into physical models and behavioral models. Physical models primarily describe carrier distribution and electrical characteristics based on semiconductor physics equations. These models typically involve solving partial differential equations or nonlinear differential equations, resulting in high computational complexity. Parameter extraction is complex, as they are dependent on device manufacturing processes. Behavioral models, on the other hand, do not involve complex physical mechanisms. Instead, they employ macroscopic models and empirical formulas to describe device switching characteristics, making parameter extraction easier. They are widely used in simulation software such as SaberRD, PSIM, and PLECS. However, due to the nonlinear voltage-current relationship of power semiconductors, the numerical solution of behavioral models requires iterative calculations, resulting in limited convergence.

[0004] Both physical and behavioral models are based on numerical solutions and therefore face common challenges in simulation. Due to the high switching frequency and low on-state resistance of high-frequency switching devices, the non-ideal characteristics of their transient processes are more pronounced. Accurately solving the transient waveforms of high-frequency devices requires the establishment of highly accurate device-level simulation models and the solution to be performed with a sufficiently low simulation step size (an order of magnitude smaller than the transient timescale). However, this inevitably leads to increased computational resource consumption and solution time, creating a conflict between accuracy, computational resources, and simulation time.

[0005] Based on the development of power electronics device-level simulation over the past five years, the following major improvement strategies have been adopted to reduce resource consumption and solution time. One approach involves simplifying and approximating power semiconductor models to create simulation models with lower complexity and higher accuracy. However, this often comes at the expense of accuracy. This is particularly true when describing the fast transient characteristics of high-frequency switching devices (such as oscillations, overshoot, and spikes), as simplified models may not accurately capture these non-ideal phenomena. Another approach involves improving simulation solution strategies, such as using piecewise linearization to avoid iterative solutions, saving computational time and resources. However, while this approach is ideal, its accuracy depends on the selection of segmentation points and may result in the loss of key transient characteristics. For real-time simulation, multi-level parallelization can significantly shorten the critical path within a single computational step of converter simulation, accelerating simulation speed. However, parallelization comes at the expense of solution accuracy and numerical stability, and consumes significant computational resources. With the increasing complexity and higher switching frequencies of power electronics systems, parallelization approaches are difficult to maintain and improve solution accuracy. Consequently, device-level simulation also faces a trade-off between accuracy and precision.

[0006] For power electronic converters, in order to accurately obtain the transient waveforms of the high-frequency switching devices, the entire circuit must be solved with a sufficiently low step size, which consumes additional computing resources and computing time, seriously reducing simulation efficiency and slowing down the design process.

[0007] With the development of machine learning technology, some studies have used neural networks to approximate the nonlinear relationship between simulation input and output, and then called the neural network to obtain the transient voltage or current of the switching device. However, this method of treating the converter model as a "black box" and directly mapping the output transient waveform through input parameters has the following problems: (1) The generalization ability over a wide range of operating conditions is difficult to meet the simulation requirements; (2) The neural network is very complex, occupies a lot of resources, and takes a long time to calculate, which limits its application in simulation, especially real-time simulation;

[0008] In view of the above analysis, the technical problems existing in the existing power electronic simulation technology are summarized as follows: 1. The contradiction between simulation accuracy and computing resource consumption and simulation time, especially in the simulation of fast transient processes of high-frequency switching devices, this contradiction is particularly prominent. 2. The contradiction between simulation accuracy and simulation precision. 3. It is difficult to ensure and further improve the accuracy of real-time simulation transient waveforms with existing methods. 4. In order to obtain the transient waveforms of high-frequency devices, the overall simulation efficiency of power electronic converters is reduced. 5. The existing machine learning-based methods mainly construct the mapping relationship between simulation input parameters and output transient waveforms, which has high network complexity, long calculation time and poor generalization ability.

[0009] In summary, the present invention proposes a method for improving the accuracy of power semiconductor transient simulation waveforms at a relatively low cost, which is crucial for the development of power electronics simulation and power electronics technology. Summary of the Invention

[0010] In response to the problems existing in the prior art, the present invention provides a method for optimizing the transient simulation accuracy of power semiconductors based on machine learning.

[0011] The present invention is implemented as follows: a method for optimizing the transient simulation accuracy of power semiconductors based on machine learning, the method comprising:

[0012] S1, Analysis and design of power semiconductor switch transient waveform simulation accuracy optimization strategy;

[0013] S2, design machine learning neural network architecture;

[0014] S3, obtain the dataset;

[0015] S4, training the network, optimizing the network structure and training parameters, and saving the best network model;

[0016] S5, processes low-precision transient waveforms based on the trained optimal network model to improve the accuracy of transient waveforms;

[0017] S6, designs a simulation accuracy optimization unit for power electronic converter simulation.

[0018] Furthermore, the transient waveform accuracy optimization strategy described in S1 is inspired by the application of neural networks in image super-resolution. Using neural networks, low-precision waveforms are super-reconstructed to obtain high-precision waveforms. Low-precision simulation consumes less computing resources and takes less time, thus avoiding the resource consumption and solution time of further iterations. Therefore, it is necessary to fully consider resource and time consumption to minimize the cost of direct iterative solutions, which also facilitates the subsequent application of this method in various embedded devices.

[0019] Furthermore, the neural network model structure designed in S2 is constructed to have the characteristics of being flexible and simple, wherein a convolutional network is adopted to improve the applicability to different inputs, and storage resource consumption is reduced by reusing convolutional layers with the same structure. Computing resources can be sacrificed to complete convolution operations at each layer in parallel to improve simulation speed and reduce simulation solution time.

[0020] Furthermore, the dataset in S3 can be derived from data provided by the model manufacturer, or it can be constructed by directly performing offline simulations of the model under different operating conditions to obtain low-precision waveforms and high-precision waveforms. Dividing the dataset into training, validation, and test sets in a reasonable proportion can enhance the generalization of the model.

[0021] Furthermore, in S4, an optimal mapping model from a low-precision waveform to a high-precision waveform is established through network training.

[0022] Furthermore, the step of improving the accuracy of the transient waveform in S5 does not sacrifice the accuracy of the waveform. For the simulation method using iterative solution, since the numerical accuracy of the current moment depends on the numerical accuracy of the previous moment, the step can simultaneously improve the accuracy and precision of the waveform.

[0023] Furthermore, based on the transient waveform accuracy optimization method in S1 to S5, S6 proposes a simulation transient accuracy optimization unit structure that can be applied to various embedded device processors, providing theoretical support for the subsequent application of the method proposed in the present invention to power electronics real-time simulation practice.

[0024] Another object of the present invention is to provide a machine learning-based power semiconductor transient simulation accuracy optimization system based on the machine learning-based power semiconductor transient simulation accuracy optimization method, the system comprising:

[0025] The simulation solver solves the circuit to obtain the low-precision transient waveform of the high-frequency power semiconductor device as the input of the simulation transient accuracy optimization unit.

[0026] The simulation accuracy optimization unit mainly includes four parts: data reading, caching, neural network and interpolation;

[0027] Furthermore, the neural network consists of three parts: a multiplication array, an addition tree, and an activation function. The convolution operation requires the use of a large number of multipliers. Therefore, the multipliers can be reasonably reused between and within each layer to save computing resources. To this end, the network weights and biases trained on the computer are loaded into the memory unit of the processor. For network layers with the same structure, only hardware resources need to be occupied once, and the corresponding weights and biases are called when performing calculations. Of course, more hardware resources can also be used in parallel to reduce computing time.

[0028] The controller is responsible for controlling the operating logic and timing of each sub-module, and plays a decisive role in whether the calculation process can be completed.

[0029] Transient waveform output: output optimized high-precision power semiconductor simulation transient waveform.

[0030] Another object of the present invention is to provide a real-time simulation of power electronics, which replaces the further iteration from low precision to high precision in high-precision simulation calculations with a flexible, simple, and time-saving neural network model, thereby reducing computing resource consumption and accelerating simulation speed, making it possible to apply the steps of the power semiconductor transient simulation accuracy optimization method based on machine learning to real-time simulation of power electronics, thereby improving the transient accuracy of real-time simulation.

[0031] Another object of the present invention is to provide a method for improving the efficiency of offline simulation of a high-frequency converter system, characterized in that the method uses a transient waveform data set of a high-frequency semiconductor switching device constructed through a one-time multi-operating condition simulation to train a neural network model with high generalization, so that in subsequent simulations, only a low-precision step size is required to complete the simulation of the entire converter system. The high-precision transient waveform of the high-frequency semiconductor switching device can be obtained separately through the network model, thereby improving the simulation efficiency.

[0032] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0033] First, the method for improving the transient waveform accuracy of power semiconductor devices based on machine learning disclosed in the present invention fully considers the existing technical problems and the defects of existing solutions, conforms to the development trend of the times, adopts the method of artificial intelligence machine learning, and creatively proposes the idea of ​​using neural networks to establish a low-precision transient waveform to a high-precision transient waveform mapping model, designs a simple accuracy optimization strategy, and replaces the process of further iteration of low-precision waveforms to high-precision waveforms with a lightweight neural network model, thereby reducing computing resource consumption and solution time, and improving the accuracy of transient waveforms. In the embodiment, only an average network training time of 510s is required. Compared with the original low-precision waveform, the average MSE error of the IGBT switch transient waveform after accuracy optimization is reduced by more than 85%, the average RRMSE error is reduced by more than 64%, the average energy error in the turn-on process is reduced by more than 85%, and the average energy error in the turn-off process is reduced by more than 93%, with significant results.

[0034] Second, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0035] (1) The technical solution of the present invention solves a technical problem that people have long been eager to solve but have never been able to solve successfully:

[0036] The contradiction between accuracy and computing resources and simulation time consumption in power electronics simulation, as well as the contradiction between precision and accuracy, is resolved. The method proposed in this paper improves transient waveform accuracy and reduces simulation solution time with less computing resource consumption.

[0037] To accurately capture the transient state of high-frequency switching devices in power electronic converters, the entire system may need to be solved at a lower step size, reducing simulation efficiency. The method proposed in this paper allows the power electronic converter to be solved only at a low-precision step size, and then improves the transient waveform accuracy of the high-frequency switching devices through a trained neural network model, avoiding the need for high-precision solution of the entire system and improving the simulation efficiency of the power electronic converter system.

[0038] Due to resource and time constraints, existing parallelization solutions struggle to further improve the real-time simulation accuracy of high-frequency switching devices. The lightweight neural network model structure and waveform precision optimization unit designed by this invention, which can be used in various embedded devices, can replace the complex iterative process of converting low-precision waveforms to high-precision waveforms, offering great potential for real-time simulation.

[0039] The existing methods of using neural networks to assist power electronics simulation cannot directly address the pain points. The network is complex, occupies a lot of resources, has weak generalization ability, and takes a long time to calculate. Inspired by the deep learning super-resolution technology in the image field, the present invention proposes the idea of ​​using neural networks to establish a mapping from low-precision transient waveforms to high-precision transient waveforms. Since some structures of the designed neural network can be reused and some calculations can be parallelized, the accuracy of power semiconductor transient simulation waveforms can be improved at a lower cost of resources and time, and the simulation solution time can be reduced. In the application of super-resolution technology in two-dimensional data processing, the introduction of machine learning technology has effectively improved the accuracy.

[0040] The transient waveform accuracy optimization unit designed in the present invention can be applied to various embedded devices, providing a new technical support for real-time simulation of power electronics and a theoretical basis for applying lightweight neural networks to real-time simulation to further improve the accuracy of device-level real-time simulation.

[0041] (3) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0042] Applicable to a variety of power semiconductor devices: The present invention is not only applicable to high-frequency switching devices such as SiC MOSFET and IGBT, but can also be extended to the simulation optimization of other power semiconductor devices.

[0043] Reduce simulation costs. By reducing the computing resources and time required for high-precision simulation, the simulation costs are significantly reduced, which is especially suitable for the design and optimization of large-scale power electronic systems.

[0044] Improved design efficiency: The present invention can quickly and accurately obtain the transient waveform of high-frequency power semiconductor devices, shortening the R&D cycle of power electronic systems and improving design efficiency.

[0045] In summary, the present invention successfully resolves the two-way contradiction in power electronics simulation technology through innovative technical solutions, and has significant technical advantages, economic benefits, and social benefits. This invention not only fills the technical gap at home and abroad, but also provides new ideas and methods for the development of power electronics simulation technology, with broad application prospects and important promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a method for optimizing power semiconductor transient simulation accuracy based on machine learning provided by an embodiment of the present invention;

[0047] Figure 2 This is a structural diagram of a power semiconductor transient simulation accuracy optimization system based on machine learning provided by an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of transient waveforms of IGBT switches under different low-precision working conditions provided by an embodiment of the present invention;

[0049] Figure 4 Schematic diagram comparing low-precision simulated current waveform, high-precision simulated waveform, and sampled high-precision simulated waveform during the IGBT turn-on process under certain working conditions provided by an embodiment of the present invention;

[0050] Figure 5 Schematic diagram comparing a low-precision simulated current waveform, a high-precision simulated waveform, and a sampled high-precision simulated waveform during an IGBT turn-off process under certain working conditions provided by an embodiment of the present invention;

[0051] Figure 6 Schematic diagram comparing low-precision simulated voltage waveform, high-precision simulated waveform, and sampled high-precision simulated waveform during the IGBT turn-on process under certain working conditions provided by an embodiment of the present invention;

[0052] Figure 7 Schematic diagram comparing a low-precision simulated voltage waveform, a high-precision simulated waveform, and a sampled high-precision simulated waveform during an IGBT turn-off process under certain working conditions provided by an embodiment of the present invention;

[0053] Figure 8 Schematic diagram of the point-to-point mapping convolutional neural network structure provided by an embodiment of the present invention;

[0054] Figure 9 Schematic diagram of the current model training process for the turn-on process provided by an embodiment of the present invention;

[0055] Figure 10 Schematic diagram of the current model training process for the shutdown process provided by an embodiment of the present invention;

[0056] Figure 11Schematic diagram of the voltage model training process for the turn-on process provided by an embodiment of the present invention;

[0057] Figure 12 Schematic diagram of the voltage model training process for the shutdown process provided by an embodiment of the present invention;

[0058] Figure 13 1 is a schematic diagram comparing a low-precision current waveform, a high-precision current waveform, and a current waveform after precision optimization during a turn-on process provided by an embodiment of the present invention;

[0059] Figure 14 1 is a schematic diagram comparing a low-precision current waveform, a high-precision current waveform, and a current waveform after precision optimization during a shutdown process provided by an embodiment of the present invention;

[0060] Figure 15 1 is a schematic diagram comparing a low-precision voltage waveform, a high-precision voltage waveform, and a voltage waveform after precision optimization during a turn-on process provided by an embodiment of the present invention;

[0061] Figure 16 This is a schematic diagram comparing the low-precision voltage waveform, high-precision voltage waveform, and voltage waveform after precision optimization during the turn-on process provided by an embodiment of the present invention.

[0062] Figure 17 Schematic diagram comparing a network output waveform and a high-precision waveform of a current waveform in a peak region during a turn-on process under certain working conditions provided by an embodiment of the present invention;

[0063] Figure 18 Schematic diagram of the upsampling convolutional neural network structure provided by an embodiment of the present invention;

[0064] Figure 19 It is a schematic diagram comparing the network output waveform and the high-precision waveform of the current waveform in the peak area of ​​the turn-on process under certain working conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] like Figure 1 As shown, an embodiment of the present invention provides a method for optimizing power semiconductor transient simulation accuracy based on machine learning, the method comprising:

[0067] S1, analysis and design of power semiconductor switch transient waveform accuracy optimization strategy;

[0068] S2, design machine learning neural network architecture;

[0069] S3, obtain the dataset;

[0070] S4, training the network, optimizing the network structure and training parameters, and saving the best network model;

[0071] S5, processes low-precision transient waveforms based on the trained optimal network model to improve the accuracy of transient waveforms;

[0072] S6, designs a simulation accuracy optimization unit for power electronic converter simulation.

[0073] This method first addresses the issue of insufficient accuracy in transient waveform simulations of power semiconductor switches by conducting an in-depth analysis and design of a simulation accuracy optimization strategy (S1). By analyzing existing direct methods for improving waveform accuracy, a precision optimization strategy combining direct interpolation and machine learning is proposed. This ensures that the proposed method can achieve more accurate waveform accuracy at a lower cost. This step provides guidance for the design and optimization of subsequent machine learning models.

[0074] In S2, a machine learning neural network architecture suitable for processing power semiconductor transient waveform data is designed. This network architecture takes into account the temporal and nonlinear characteristics of waveform data and may employ architectures such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), combined with appropriate activation and loss functions. Subsequently, in S3, a dataset is constructed by collecting a large amount of low- and high-precision transient simulation data. This ensures that the training data is representative and diverse enough to enable the network to accurately predict waveform details under various operating conditions.

[0075] In S4, the obtained dataset is used to train the designed neural network. The network structure and training parameters (such as learning rate, batch size, and number of iterations) are continuously adjusted to minimize the preset loss function and optimize model performance. During training, the optimal network model is determined through cross-validation and loss curve monitoring, and saved for subsequent use. This process ensures the model's excellent generalization and high-precision waveform prediction capabilities, providing strong support for the refined correction of low-precision simulation results.

[0076] In the S5 stage, the trained optimal network model is used to process low-precision transient waveforms. Data correction and compensation are performed on the original simulation results to improve the accuracy of waveform details. Finally, in S6, a simulation accuracy optimization unit for power electronic converter simulation is designed. This unit integrates machine learning models into the overall simulation process to dynamically optimize waveform accuracy in different simulation scenarios. The overall system improves simulation accuracy through intelligent algorithms, significantly improving the limitations of traditional simulation methods in highly dynamic, nonlinear transient processes, thereby providing more accurate data support for the design and optimization of power electronic equipment.

[0077] The transient waveform accuracy optimization strategy described in S1 is inspired by the application of neural networks in image super-resolution. Using neural networks, low-precision waveforms are super-reconstructed to obtain high-precision waveforms. Low-precision simulations consume fewer computing resources and take shorter simulation times, thus avoiding the resource and solution costs of further iterations. Therefore, it is important to fully consider resource and time consumption to minimize the cost of direct iterative solutions, facilitating the subsequent application of this method in various embedded devices.

[0078] The neural network model structure designed in S2 should be as lightweight, flexible and simple as possible, with some structures being reusable and some calculations being parallelized, so as to reduce resource consumption and increase simulation speed.

[0079] The dataset in S3 can be sourced from data provided by the model manufacturer, or directly constructed by performing offline simulations of the model under different operating conditions to obtain low-precision and high-precision waveforms. Dividing the dataset into training, validation, and test sets in a reasonable proportion can enhance the generalization of the model.

[0080] In the S4, an optimal mapping model from a low-precision waveform to a high-precision waveform is established through network training.

[0081] The step of improving the accuracy of the transient waveform in S5 does not sacrifice the accuracy of the waveform. For the simulation method using iterative solution, since the numerical accuracy of the current moment depends on the numerical accuracy of the previous moment, this step can simultaneously improve the accuracy and precision of the waveform.

[0082] Based on the transient simulation accuracy optimization method in S1 to S5, S6 proposes a transient simulation accuracy optimization unit structure that can be applied to various embedded device processors, providing theoretical support for the subsequent application of the method proposed in the present invention to power electronics real-time simulation practice.

[0083] 1. Transient waveform simulation accuracy optimization strategy

[0084] Methods for improving waveform accuracy generally include direct interpolation and curve fitting. However, given the non-ideal characteristics of transient processes, such as nonlinearity and discontinuity, direct interpolation struggles to guarantee waveform accuracy, while curve fitting struggles to find an appropriate time function. Machine learning neural networks excel at describing the mapping relationship between low-precision and high-precision data points. However, establishing an accurate one-to-many point mapping requires a complex network structure and long training times, potentially exceeding the cost of direct iterative solutions. Therefore, consideration is given to combining simple direct interpolation with neural networks, which have the ability to represent nonlinear relationships.

[0085] As discussed above, for device models using iterative solvers, low-precision waveforms often lack high accuracy. Assuming a low-precision simulation step size of h1 and a high-precision simulation step size of h2, a neural network is used to establish a point-to-point mapping between low-precision waveform data points and the high-precision waveform at the low-precision sampling points (i.e., the high-precision waveform is sampled with a step size of h1). Therefore, after the low-precision waveform passes through the network, a high-accuracy waveform with a step size of h1 is generated. Then, through linear interpolation, a high-precision waveform with a step size of h2 is generated, thereby improving both accuracy and precision. However, this approach is likely to miss key points in rapidly changing regions of the waveform, such as peaks, resulting in significant errors in the interpolated waveform. To avoid increasing the complexity of the feature extraction unit, the high-precision waveform sampling step size is reduced to, for example, h1 / 2 in these regions. Then, an upsampling layer is added to the neural network, establishing a mapping from n points to 2n-1 points. Theoretically, with a sufficiently small step size, all key points can be mapped, but this further increases the network complexity.

[0086] 2 Neural Network Model Structure

[0087] The transient waveforms of high-frequency switching devices often have distinct characteristics. For example, the current waveform during the IGBT turn-on process always rises rapidly before gradually falling back to a stable value, possibly with spikes or oscillations. Convolutional neural networks (CNNs) use multiple layers of convolution and pooling to gradually learn high-level features from low-level features. This hierarchical feature learning capability enables them to capture complex patterns in waveforms. Furthermore, compared to other neural networks, CNN convolution kernels share parameters across input data, significantly reducing the model's parameter count and computational burden. Their lightweight nature makes them suitable for deployment on embedded devices, laying an important foundation for applying this method to real-time simulation.

[0088] Another advantage of CNNs is that one-dimensional convolution has no fixed requirements for the length of the input data; the length of the output data is determined by the convolution parameters. This allows the same model to be used for processing different numbers of sampling points and different regions, and can also adjust single-point-to-single-point or single-point-to-multipoint mapping by adjusting parameters. Therefore, neural networks are composed of multiple layers of convolutional networks, and upsampling networks only require adjusting the parameters of one or more layers. The specific network depth and convolution parameters require continuous debugging to find the optimal solution for specific tasks.

[0089] 3. Get the dataset

[0090] In order to enhance the generalization of the model, offline simulation software is used to obtain low-precision and high-precision transient waveforms under various working conditions of the model and complete the sampling of high-precision waveforms. The working conditions need to cover the entire working range, and the data set is randomly divided into training set, validation set and test set.

[0091] 4 Network Training

[0092] Training parameters such as the learning rate, optimizer, and number of training rounds, like network parameters, need to be constantly debugged. An early stopping strategy is used during training. That is, if the loss does not improve significantly after a certain number of consecutive rounds, the training process is terminated early. This strategy can significantly save training time and computing resources.

[0093] 5. Optimization of transient waveform simulation accuracy

[0094] The low-precision waveforms in the test set are optimized for accuracy, and the optimized high-precision waveforms are analyzed to adjust the neural network structure parameters or neural network training parameters.

[0095] 6 Simulation Accuracy Optimization Unit Design

[0096] The lightweight network makes this method easily applicable to processors of various embedded devices. The simulation solver obtains a low-precision transient waveform as input to the optimization unit, and a high-precision waveform is obtained through calculation.

[0097] The hardware circuitry of the simulation accuracy optimization unit primarily consists of four parts: data reading, caching, neural network, and controller. The neural network consists of a multiplication array, an addition tree, and an activation function. Convolution operations require a large number of multipliers, so multipliers can be reused across and within layers to conserve computing resources. To this end, the network weights and biases trained on the computer are loaded into the processor's memory unit. For network layers with the same structure, only hardware resources are required once, with the corresponding weights and biases being called upon during computation. Furthermore, more hardware resources can be used in parallel to reduce computation time.

[0098] The controller is the most important part of this unit. It is responsible for controlling the operating logic and timing of each sub-module, and plays a decisive role in whether the calculation process can be completed.

[0099] Specific application fields or related products of the present invention:

[0100] like Figure 2 As shown, an embodiment of the present invention provides a power semiconductor transient simulation accuracy optimization system based on machine learning based on the power semiconductor transient simulation accuracy optimization method based on machine learning, the system comprising:

[0101] The simulation solver solves the circuit to obtain the low-precision transient waveform of the high-frequency power semiconductor device as the input of the simulation transient accuracy optimization unit.

[0102] The simulation accuracy optimization unit mainly includes four parts: data reading, caching, neural network and interpolation;

[0103] The controller is responsible for controlling the operating logic and timing of each module, and plays a decisive role in whether the calculation process can be completed.

[0104] Transient waveform output: output optimized high-precision power semiconductor simulation transient waveform.

[0105] An embodiment of the present invention provides a power electronics simulation, which replaces the further iteration from low precision to high precision in high-precision simulation calculations with a flexible, simple, and time-saving neural network model, thereby reducing computing resource consumption and accelerating simulation speed, making it possible to apply the steps of the power semiconductor transient simulation accuracy optimization method based on machine learning to real-time power electronics simulation, thereby improving the transient accuracy of real-time simulation.

[0106] Evidence related to the technical effects achieved by the embodiments of the present invention:

[0107] The present invention takes a device-level IGBT simulation model as an example to introduce the whole process of optimizing the accuracy of IGBT switch transient simulation waveform.

[0108] In the embodiment, the original data set is directly obtained through offline power electronics simulation software. The switching transient characteristics of the IGBT are mainly related to the circuit input voltage and input current. The input voltage VCC range is 400-1200V, and the input current IL range is 200-1000A. In order to make the trained model have good generalization, the training set data must cover the complete working range. The voltage step is set to 16V and the current step is set to 16A, then there are a total of 2601 working combinations. The low-precision transient simulation step h1 is set to 50ns, and the high-precision transient simulation step h2 is set to 5ns. Therefore, the number of data points of the high-precision waveform is ten times the number of low-precision data points. The simulation is run to obtain the switching transient waveforms under two accuracies under different working conditions. Figure 3 The low-precision IGBT voltage and current turn-on and turn-off transient waveforms under different working conditions are displayed.

[0109] The high-precision waveform is sampled according to the step size h1 of the low-precision waveform to obtain a sampled high-precision waveform. At this time, the number of data points of the sampled high-precision waveform is the same as the number of data points of the low-precision waveform. Figure 4 、 Figure 5 、 Figure 6 、 Figure 7The comparison diagrams of the low-precision waveform, original high-precision waveform, and high-precision waveform after sampling of four transient waveforms under the working conditions of VCC=400V and IL=200A respectively show the following conclusions: 1. Since the model is solved in an iterative manner, the accuracy of the low-precision waveform is also very poor. 2. The high-precision waveform after sampling is very close to the original high-precision waveform after linear interpolation, and the accuracy error is very small. Only in some areas with large slope changes, such as Figure 4 The peaks of the medium waveform have relatively large errors due to insufficient sampling points in the peak area. Therefore, by accurately mapping the sample points of the low-precision waveform directly to the sample points of the high-precision waveform after sampling, and then using linear interpolation, we can optimize the accuracy while ensuring the accuracy of the iteratively solved high-precision waveform. If we want to further obtain a more accurate high-precision waveform, we can increase the number of sampling points and map the low-precision sample points to more sample points of the high-precision waveform, but this will inevitably increase the network complexity and training resource consumption.

[0110] The low-precision one-dimensional data of the four waveforms and the sampled high-precision one-dimensional data were combined into four original data sets, and then the original data sets were randomly divided into training set, validation set and test set in the ratio of 2000:301:300.

[0111] In order to minimize resource consumption, the number of network layers and convolution kernel size are continuously optimized through experiments, and the obtained Figure 8 In the network structure shown in Figure 1, low-precision waveform data passes through a five-layer convolutional neural network in sequence to obtain the predicted high-precision sampling point data. The specific parameters of each convolutional layer are shown in Table 1. It can be seen that convolutional layers 2, 3, and 4 have exactly the same structure, which is conducive to reuse and resource saving.

[0112] During the convolution process, the number of data points remains unchanged through reasonable padding, thereby establishing a single-point to single-point mapping. The relationship between the convolution layer input data length n1 and the output data length n2 is shown in formula (1).

[0113] (1);

[0114] Here, k is the size of the convolution kernel, p is the amount of padding, and s is the convolution stride. Note that the padding method should use "nearest neighbor padding," which fills missing values ​​with the corresponding values ​​of the nearest neighbor samples. This is because the beginning and end of the waveform are both in a steady state. Using the default "zero padding" method would introduce incorrect features, causing the CNN to learn incorrect mapping relationships.

[0115] Table 1 Network parameters

[0116] Number of input channels Number of output channels Convolution kernel size filling step length Convolutional layer 1 1 4 3 1 1 Convolutional layer 2 4 4 3 1 1 Convolutional layer 3 4 4 3 1 1 Convolutional layer 4 4 4 3 1 1 Convolutional layer 5 4 1 3 1 1

[0117] For the four transient waveforms, four network models were trained respectively, with a model learning rate of 0.001, an optimizer of Adam, a loss function of MSELoss, an epoch limit of 1500, and a patience of 20. This means that when the loss does not improve significantly after 20 consecutive rounds, the model is saved and training is stopped. This early stopping strategy can greatly save training time and computing resource consumption. The losses in the training set and test set during the training of the four waveforms are as follows: Figure 9 、 Figure 10 、 Figure 11 、 Figure 12 As shown in the figure, even if only the CPU is used, each model only needs an average of 449 rounds and 510 seconds to complete training.

[0118] The low-precision waveform of the test set is processed by the model to obtain the predicted waveform, and then the predicted waveform is linearly interpolated to the high-precision sample points to obtain the waveform with optimized accuracy. Each transient waveform is randomly selected from a certain working condition to observe the optimization effect, such as Figure 13 、 Figure 14 、 Figure 15 、 Figure 16 As shown, in order to more intuitively evaluate the actual effect of this method, the root mean square error (MSE), relative root mean square error (RRMSE), relative error of energy loss during the turn-on process (ŋ1), and relative error of energy loss during the turn-off process (ŋ2) are used as evaluation indicators, and the calculation formulas are as follows:

[0119] (2);

[0120] (3);

[0121] Where yj^ = {y1^, y2^, …yn^} is the value of each point after the low-precision waveform is interpolated to the high-precision sample point or the value of each point after the low-precision waveform is optimized for accuracy, and yj is the value of each point of the corresponding high-precision waveform.

[0122] (4);

[0123] (5);

[0124] Where t1 and t2 are defined as the time required for the IGBT collector-emitter voltage to reach 90% of its stable value and the time required for the voltage to reach 10% of its stable value, respectively. t3 and t4 are defined as the time required for the IGBT collector-emitter voltage to reach 10% of its final value and the time required for it to reach 90% of its final value, respectively. The numerators above Equations (4) and (5) calculate the turn-on or turn-off losses of the low-precision waveform or the oversampled waveform, respectively, while the denominators below calculate the turn-on / turn-off losses of the corresponding high-precision waveform.

[0125] The aforementioned metrics were calculated for the four waveforms in the test set before and after precision optimization. To ensure consistent scaling, the low- and medium-precision waveforms were directly linearly interpolated to high-precision samples. The average MSE, RRMSE, ŋ1, ŋ2, and improvement percentages for the 300 random operating conditions in the test set are shown in Tables 2, 3, 4, 5, and 6. It can be seen that the waveforms optimized using the proposed method are highly accurate. Compared with the high-precision waveforms obtained by direct iteration, the maximum average RRMSE error occurs in the turn-on voltage waveform, at only 0.84%. Compared to the low-precision waveforms, this represents a significant improvement in accuracy, with the average improvement percentage under the MSE metric exceeding 90%. Overall, the energy loss prediction for the switching process is also more accurate, with energy loss errors for the turn-on and turn-off processes being only 2.21% and 0.22% respectively.

[0126] In summary, the method proposed in the present invention effectively improves the precision and accuracy of transient waveforms in power semiconductor switch simulations at a relatively low cost in resources and time, and resolves two contradictions in current power electronics simulations.

[0127] Table 2 Turn-on process current waveform accuracy optimization accuracy improvement effect,

[0128] index index Waveform MSE RRMSE Low-precision waveform 187.70 2.14% Precision-optimized waveforms 18.27 0. 76% Percentage increase 90.26% 64.49%

[0129] Table 3 Accuracy improvement effect of current waveform optimization during shutdown process

[0130] index index Waveform MSE RRMSE Low-precision waveform 3191.49 14.88% Precision-optimized waveforms 8.53 0. 79% Percentage increase 99.73% 94.69%

[0131] Table 4 Accuracy improvement effect of voltage waveform optimization during turn-on process

[0132] index index Waveform MSE RRMSE Low-precision waveform 129.81 3.60% Precision-optimized waveforms 7.11 0. 84% Percentage increase 94.52% 76.67%

[0133] Table 5 Accuracy improvement effect of voltage waveform optimization during shutdown process

[0134] index index Waveform MSE RRMSE Low-precision waveform 3245.46 8.42% Precision-optimized waveforms 8.24 0. 45% Percentage increase 99.75% 94.66%

[0135] Table 6 Effect of improving energy loss prediction accuracy before and after precision optimization

[0136] index index Waveform <![CDATA[ŋ1]]> <![CDATA[ŋ2]]> Low-precision waveform 15.08% 3.60% Precision-optimized waveforms 2.21% 0.22% Percentage increase 85.34% 93.89%

[0137] In the previous network structure, through reasonable padding, the data length of the convolution process remains unchanged, thereby establishing a single-point mapping between the low-precision waveform and the high-precision waveform at the low-precision sampling point. This method is simple and direct, but it has a great disadvantage: it is easy to miss some key points in the area with a large rate of change, such as the peak point of the transient waveform, resulting in Figure 17 The situation shown, Figure 17 This is a comparison chart of the network model output and the high-precision waveform in the peak area of ​​the IGBT turn-on process current waveform under certain operating conditions. It can be seen that since the peak point of the high-precision waveform does not correspond to the sampling point, interpolation will cause a large error. Accurately predicting peaks is very important for power electronics simulation.

[0138] In response to the above problems, an improvement method is proposed. For areas with large transient waveform change rates, deconvolution (upsampling) is used to establish a mapping from one point to multiple high-precision waveform sample points. The upsampling convolutional neural network structure is as follows: Figure 18 Compared to Figure 8 The improved network changes the fourth convolution unit into an upsampling unit, and the rest of the structure remains unchanged. The upsampling unit consists of a deconvolution layer and an activation function relu. The parameters of the deconvolution layer are shown in the following table. Through reasonable step size and padding, the deconvolution layer doubles the data length, thereby establishing a mapping from 1 point to 3 points.

[0139] Table 7 Deconvolution layer parameters

[0140] Number of input channels Number of output channels Convolution kernel size filling step length Deconvolution layer 4 4 3 1 2

[0141] Through the optimization of the upsampling network, Figure 17 The error is improved, and the improved waveform is as follows Figure 19 As shown in the figure, the error in the peak area is significantly reduced, but the mapping accuracy of some points is not high. It can be foreseen that as long as the number of mapping points continues to increase, the error in the peak area can continue to decrease, but the cost is that under the same network structure, the task complexity becomes higher and the mapping accuracy will decrease.

[0142] The network structure proposed in this embodiment can theoretically be used for real-time simulation. For the internal structure of the single-point mapping network, convolution units 2, 3, and 4 are exactly the same. For the upsampling convolution network, convolution units 1, 2, 3, and 5 are exactly the same as the corresponding units in the single-point mapping network. This design provides more flexibility for the simulation calculation process, whether it is to reuse FPGA hardware to save resources or to consume more resources for parallel computing, there are more options.

[0143] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0144] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for optimizing power semiconductor transient simulation accuracy based on machine learning, characterized in that: The method includes: S1: Analyze simulation accuracy optimization strategies. To address the problem of insufficient accuracy in transient waveform simulation of power semiconductor switches and the resource limitations of the hardware platform, a composite waveform accuracy optimization method combining machine learning and interpolation is proposed. S2. Design a machine learning neural network architecture suitable for processing power semiconductor transient waveform sequences. This architecture considers the temporal and nonlinear characteristics of waveform data and uses convolutional neural networks, recurrent neural networks, or a combination thereof to input entire low-precision waveform sequences and output entire high-precision waveform sequences. It also incorporates appropriate activation and loss functions to enhance the network's ability to represent transient characteristics. S3, collects a large amount of low-precision and high-precision transient simulation data to build a data set, ensuring that the training data is sufficiently representative and diverse so that the neural network can accurately predict waveform details under various working conditions; S4, trains a neural network model based on the constructed dataset, optimizes model performance by adjusting the network structure and training parameters, including learning rate, batch size, and number of iterations, and determines the optimal network model with good generalization ability and high-precision prediction capabilities through cross-validation and loss curve monitoring; S5, uses the trained neural network model combined with the interpolation method to correct and compensate the low-precision transient waveform to avoid error accumulation in the peak area, thereby improving the accuracy of the simulation waveform and making the corrected simulation results consistent with the high-precision simulation results; S6, designing a simulation accuracy optimization unit, which contains the trained neural network model and is embedded in the power electronic converter simulation process; The transient simulation accuracy optimization strategy in S1 is inspired by the application of neural networks in the field of image super-resolution. By reconstructing low-precision waveforms, a result close to high-precision simulation is obtained, avoiding the high computational overhead caused by further iterations. Considering the deployment requirements of subsequent models on embedded hardware or FPGA platforms, an n-point-to-n method of mapping low-precision waveform points to low-precision sampling points of high-precision waveforms is adopted, and the remaining points are supplemented by interpolation methods, thereby achieving overall reconstruction of high-resolution waveforms. The convolutional neural network includes at least three reusable convolutional layers with the same structure, which reduces resource consumption through a hardware reuse mechanism. At the same time, the convolutional layers are reusable and support pipeline parallel computing. In S4, based on the convolutional neural network model with n-point to n-point mapping, in order to compensate for the large errors that may occur in the area of ​​waveform spike value mutation, an upsampling convolutional neural network containing a deconvolution layer is further improved. The n low-precision points are mapped to 2n−1 high-precision sampling points, thereby improving the reconstruction accuracy of the waveform in complex areas.

2. The power semiconductor transient simulation accuracy optimization method based on machine learning according to claim 1 is characterized in that: What is established in S4 is the mapping of the low-precision waveform to the high-precision waveform sampling points.

3. A system for optimizing power semiconductor transient simulation accuracy based on machine learning, based on the method for optimizing power semiconductor transient simulation accuracy based on machine learning as claimed in any one of claims 1 to 2, characterized in that: The system includes: a simulation solver, which solves the circuit to obtain low-precision transient waveforms of power semiconductor devices as input to the simulation transient accuracy optimization unit; a simulation accuracy optimization unit, based on S6 and the data processing flow, which mainly includes four parts: data reading, caching, neural network, and interpolation; and a controller, which is responsible for controlling the operating logic and timing of each sub-module and plays a decisive role in whether the calculation process can be completed.

4. A method for improving the offline simulation efficiency of a high-frequency converter system based on the power semiconductor transient simulation accuracy optimization method based on machine learning as described in any one of claims 1-2, characterized in that: The entire converter system only needs to be simulated with a low-precision step size. The high-precision transient waveform of the power semiconductor switch can be obtained through the precision optimization method, thus avoiding the high-precision simulation of the entire system.

Images