A method and apparatus for regulating a dual active bridge converter based on PGNN
By constructing a hybrid model based on a physical guided neural network, the prediction error problem at the ZVS boundary of the dual active bridge converter was solved, achieving high-precision phase shift and loss prediction and improving the control effect of the converter.
Patent Information
- Application Number
- CN202611226440.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-17
AI Technical Summary
Existing phase-shift control methods for dual active bridge converters are prone to prediction errors at the ZVS boundary due to abrupt changes in switching losses, resulting in low prediction accuracy of loss values and affecting control accuracy.
A hybrid model based on Physically Guided Neural Network (PGNN) is constructed. Data is collected through simulation and experiments to train the initial ZVS boundary and DAB loss PGNN model. The PGNN model is then combined with the physical guidance layer and the data-driven model for iterative optimization to achieve regional prediction and adjustment of loss values.
It improves the prediction accuracy of phase shift and loss values, significantly improves the control accuracy of dual active bridge converters, overcomes the prediction error at the ZVS boundary of traditional methods, and achieves efficient converter operation.
Smart Images

Figure CN123122253A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of power electronics technology, and in particular to a method and apparatus for regulating a dual active bridge converter based on PGNN. Background Technology
[0002] Dual Active Bridge (DAB) converters are widely used in electric vehicles, energy storage systems, and DC microgrids due to their advantages such as bidirectional power transfer, electrical isolation, and wide voltage gain range. To improve converter efficiency, phase-shift control combined with frequency regulation is typically used to achieve zero-voltage switching (ZVS) of the power devices. Accurately predicting the ZVS boundary and converter losses is a prerequisite for achieving efficiency optimization.
[0003] Existing methods mostly employ purely physical-guided neural networks (PGNNs) or purely data-driven models. The former ignores non-ideal factors such as dead zones and parasitic parameters, while the latter lacks physical constraints and uses only a single type of model, resulting in low prediction accuracy of phase shift values. Furthermore, the method does not perform partitioned prediction of converter losses, and prediction errors are easily generated at the ZVS boundary due to abrupt changes in switching losses, resulting in low prediction accuracy of loss values and affecting the control accuracy of dual active bridge converters. Summary of the Invention
[0004] To address the aforementioned problems, in a first aspect, the present invention provides a method for regulating a dual active bridge converter based on a PGNN, comprising: The training operating condition data set, tag phase shift value, and tag loss value of the target dual active bridge converter were obtained through simulation and experimental acquisition. The training operating condition data set includes primary voltage, secondary voltage, transformer turns ratio, output current, switching frequency, output phase shift value, theoretical phase shift value, and theoretical loss value. An initial ZVS boundary PGNN model is constructed. The initial ZVS boundary PGNN model is iteratively optimized using primary side voltage, secondary side voltage, output current, theoretical phase shift value, and tag phase shift value. After the iterative optimization is completed, the final ZVS boundary PGNN model and sample predicted phase shift value are obtained. The theoretical loss values are processed by region based on the primary voltage, secondary voltage, transformer turns ratio, output phase shift value, and sample predicted phase shift value to obtain the full ZVS theoretical loss value, the primary ZVS loss theoretical loss value, and the secondary ZVS loss theoretical loss value. An initial DAB loss PGNN model is constructed. The initial DAB loss PGNN model is then iteratively optimized using primary voltage, transformer turns ratio, secondary voltage, switching frequency, output phase shift value, theoretical loss value of full ZVS, theoretical loss value of primary ZVS loss, theoretical loss value of secondary ZVS loss, and tag loss value. After the iterative optimization is completed, the final DAB loss PGNN model is obtained. A comprehensive loss prediction model is formed by the final ZVS boundary PGNN model and the final DAB loss PGNN model. The actual operating condition data set is obtained and input into the comprehensive loss prediction model to predict the loss, thereby obtaining the loss prediction value during operation. The operation of the target dual active bridge converter is adjusted according to the loss prediction value during operation.
[0005] In one embodiment, the initial DAB loss PGNN model includes a full ZVS loss model, a primary-edge ZVS loss model, and a secondary-edge ZVS loss model; The initial ZVS boundary PGNN model, the full ZVS loss model, the primary edge ZVS loss model, and the secondary edge ZVS loss model all consist of an input layer, a hidden layer, a physical guidance layer, and an output layer.
[0006] In one embodiment, the iterative optimization process of the initial ZVS boundary PGNN model includes: The primary side voltage, secondary side voltage, output current, and theoretical phase shift value are used as boundary feature vectors; The boundary feature vector is input into the initial ZVS boundary PGNN model. The hidden layer generates the boundary deviation based on the boundary feature vector. The theoretical phase shift value is added to the boundary deviation after being mapped by the physical guidance layer to obtain the sample predicted phase shift value. The boundary data-driven loss is calculated based on the sample predicted phase shift value and the tag phase shift value. The boundary physical guidance loss is calculated based on the sample predicted phase shift value and the theoretical phase shift value. The boundary data-driven loss and the boundary physical guidance loss are weighted and added together to obtain the total boundary loss. The initial ZVS boundary PGNN model is iteratively optimized using the Adam optimizer until the total boundary loss converges to the corresponding preset threshold, and the final ZVS boundary PGNN model is output.
[0007] In one embodiment, the boundary physical guidance loss is calculated based on the sample predicted phase shift value and the theoretical phase shift value, including: Set the maximum physical limit phase shift value, and add the theoretical phase shift value to the maximum physical limit phase shift value to obtain the physical upper limit of the phase shift value; The difference between the theoretical phase shift value and the sample predicted phase shift value is input into the ReLU function to calculate the first function value, and the difference between the sample predicted phase shift value and the physical upper limit of the phase shift value is input into the ReLU function to calculate the second function value. The physical penalty term corresponding to the sample is obtained by weighted summing of the first function value and the second function value, and the physical penalty terms of all samples are summed and averaged to obtain the boundary physical guidance loss.
[0008] In one embodiment, the region-based processing includes: The voltage gain is calculated based on the primary voltage, secondary voltage, and transformer turns ratio. When the output phase shift value is greater than or equal to the sample predicted phase shift value, the theoretical loss value is corrected by the primary side turn-off loss and the secondary side turn-off loss to obtain the full ZVS theoretical loss value. When the output phase shift value is less than the sample predicted phase shift value and the voltage gain is greater than 1, the theoretical loss value is corrected by the primary side turn-on and turn-off losses and the secondary side turn-off losses to obtain the theoretical loss value of the primary side ZVS loss. When the output phase shift value is less than the sample predicted phase shift value and the voltage gain is less than 1, the theoretical loss value is corrected by the primary side turn-off loss and the secondary side turn-on and turn-off loss to obtain the theoretical loss value of the secondary side ZVS loss.
[0009] In one embodiment, the iterative optimization process of the initial DAB loss PGNN model includes: The primary voltage, transformer turns ratio, secondary voltage, switching frequency, and output phase shift value are used as the operating condition feature vector. The first loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the full ZVS. The second loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the primary ZVS loss. The third loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the secondary ZVS loss. The first loss feature vector is input into the full ZVS loss model. The hidden layer generates the first loss deviation based on the first loss feature vector. The full ZVS theoretical loss value is mapped by the physical guidance layer and added to the first loss deviation to obtain the loss prediction value of the first training process. The first loss data-driven loss is calculated based on the first training process loss prediction value and the label loss value. The first loss physical guidance loss is calculated based on the first training process loss prediction value and the full ZVS theoretical loss value. The first loss data-driven loss and the first loss physical guidance loss are weighted and added together to obtain the first loss total loss. The second loss feature vector is input into the primary edge ZVS loss model, and the third loss feature vector is input into the secondary edge ZVS loss model. The same calculation process as the full ZVS loss model is used to obtain the total loss of the second loss and the total loss of the third loss, respectively. The initial DAB loss PGNN model is iteratively optimized using the Adam optimizer until the total loss of the first loss, the total loss of the second loss, and the total loss of the third loss all converge to the corresponding preset thresholds, thus obtaining the final DAB loss PGNN model.
[0010] In one embodiment, the actual operating condition data set is input into the comprehensive loss prediction model to predict the loss and obtain the predicted loss value during operation, including: The actual working condition data set is input into the final ZVS boundary PGNN model to predict the phase shift value and obtain the actual predicted phase shift value. Based on the actual predicted phase shift value, the region is determined to obtain the region type, which includes full ZVS region, primary edge ZVS lost region and secondary edge ZVS lost region. The actual operating condition data set is input into the final DAB loss PGNN model, which includes a full ZVS loss model, a primary-edge ZVS loss model, and a secondary-edge ZVS loss model. When the region type is a full ZVS region, loss prediction is performed using the full ZVS loss model; when the region type is a primary-edge ZVS loss region, loss prediction is performed using the primary-edge ZVS loss model; when the region type is a secondary-edge ZVS loss region, loss prediction is performed using the secondary-edge ZVS loss model. The predicted loss value during operation is output.
[0011] Secondly, the present invention provides a PGNN-based dual active bridge converter adjustment device for implementing the PGNN-based dual active bridge converter adjustment method, the device comprising: The training data acquisition module is used to obtain the training operating condition data set, tag phase shift value, and tag loss value of the target dual active bridge converter through simulation and experimental acquisition. The training operating condition data set includes primary voltage, secondary voltage, transformer turns ratio, output current, switching frequency, output phase shift value, theoretical phase shift value, and theoretical loss value. The ZVS boundary PGNN model training module is used to construct the initial ZVS boundary PGNN model. It iteratively optimizes the initial ZVS boundary PGNN model using primary side voltage, secondary side voltage, output current, theoretical phase shift value, and label phase shift value. After the iterative optimization is completed, the final ZVS boundary PGNN model and sample predicted phase shift value are obtained. The regional processing module is used to process the theoretical loss value into regions based on the primary voltage, secondary voltage, transformer turns ratio, output phase shift value and sample predicted phase shift value, to obtain the full ZVS theoretical loss value, the primary ZVS loss theoretical loss value and the secondary ZVS loss theoretical loss value. The DAB loss PGNN model training module is used to construct the initial DAB loss PGNN model. Iterative optimization of the initial DAB loss PGNN model is performed using primary voltage, transformer turns ratio, secondary voltage, switching frequency, output phase shift value, theoretical loss value of full ZVS, theoretical loss value of primary ZVS loss, theoretical loss value of secondary ZVS loss, and tag loss value. After the iterative optimization is completed, the final DAB loss PGNN model is obtained. The DAB operation adjustment module is used to obtain the actual operating condition data set by the comprehensive loss prediction model composed of the final ZVS boundary PGNN model and the final DAB loss PGNN model. The actual operating condition data set is input into the comprehensive loss prediction model to predict the loss and obtain the loss prediction value during operation. The operation adjustment is performed on the target dual active bridge converter based on the loss prediction value during operation.
[0012] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned PGNN-based dual active bridge converter adjustment method.
[0013] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned PGNN-based dual active bridge converter adjustment method.
[0014] The present invention has the following beneficial effects: 1. By constructing a ZVS boundary PGNN model and a DAB loss PGNN model that include a physical guidance layer, the ZVS boundary PGNN model uses the theoretical phase shift value as an identity mapping bypass through the physical guidance layer. The neural network only learns the residual deviation caused by non-ideal factors such as dead zone and parasitic parameters. The combination of the two improves the prediction accuracy of the phase shift value. The theoretical loss value is processed by dividing the predicted phase shift value into regions. The DAB loss PGNN model calls the corresponding sub-model to predict the loss value based on the regional division result. This overcomes the prediction error caused by the sudden change in switching loss at the ZVS boundary in the traditional single model, improves the prediction accuracy of the loss value, and thus enables high-precision control of DAB based on the predicted loss value.
[0015] 2. In the ZVS boundary PGNN model, the hidden layers only need to learn the small deviation between the actual ZVS boundary and the theoretical boundary, while the physical guidance layer directly transmits the theoretical phase shift value to the output layer through a parameter-free identity mapping. The model approximates the physical theoretical solution in the early stages of training, significantly accelerating convergence and improving the generalization ability under sparse samples. An asymmetric physical penalty mechanism is designed to address the dead zone effect and hardware limits. The theoretical phase shift value is used as the physical lower bound, and the maximum physical limit phase shift value calculated based on the highest frequency and dead time is used as the elastic upper bound, forming an elastic wall. This ensures that the model's predicted value is neither lower than the theoretical value nor excessively exceeds the physical upper limit, thereby accurately compensating for the deviation caused by parasitic parameters and other non-ideal factors, and significantly improving the prediction accuracy of the phase shift value.
[0016] 3. The DAB loss PGNN model includes three independent sub-networks: full ZVS, primary edge ZVS loss, and secondary edge ZVS loss. For each sub-network, a loss feature vector is constructed, consisting of the operating condition feature vector and the theoretical loss value of the corresponding region. This allows the physical prior to be used as an additional input rather than just as a loss constraint. Each sub-network independently adopts a residual structure including a hidden layer and a physical guidance layer, as well as a composite loss function of data-driven loss plus physical guidance loss. This enables accurate partitioning modeling of the loss model at the ZVS boundary transition, effectively overcoming the problems of poor accuracy of a single physical model in the soft-hard switching transition region and overfitting of a pure data model in sparse regions, and significantly improving the prediction accuracy of loss values. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 An architecture diagram for a hybrid physical-data model; Figure 3 A schematic diagram illustrating the iterative optimization process of the initial ZVS boundary PGNN model; Figure 4 This is a schematic diagram of the iterative optimization process of the initial DAB loss PGNN model; Figure 5 This is a structural diagram of the device according to an embodiment of the present invention. Detailed Implementation
[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.
[0020] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.
[0021] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.
[0022] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0023] Reference Figure 1 The present invention provides a dual active bridge converter adjustment method based on PGNN, including steps S100-S500.
[0024] S100 obtains the training operating condition data set, tag phase shift value, and tag loss value of the target dual active bridge converter through simulation and experimental acquisition. The training operating condition data set includes primary voltage, secondary voltage, transformer turns ratio, output current, switching frequency, output phase shift value, theoretical phase shift value, and theoretical loss value.
[0025] In some embodiments, high-precision simulation and experimental measurement data are acquired for the dual active bridge (DAB) converter. A DAB converter model containing real physical characteristics is built using PLECS simulation software. The model includes the nonlinear junction capacitances of IGBTs and MOSFETs. Simulations included body diode characteristics, a fixed dead time of 200ns, a thermal model of a power switch, and simulations of magnetic component losses. The SimulationScript module automatically scanned a preset wide operating range: primary voltage... Secondary voltage Output current For each static operating point, a binary search algorithm is used, starting from the theoretical boundary. The success of soft switching is determined based on the drain-source voltage state at the instant the switching transistor is turned on. The phase shift value is adjusted until the minimum phase shift value that achieves ZVS is found. The convergence tolerance is set to... The minimum phase shift value is the tag phase shift value. Simultaneously, the steady-state loss value is directly obtained as the tag loss value using the periodic steady-state analysis mode of the simulation solver. .
[0026] Theoretical phase shift value According to voltage gain The voltage gain is calculated and defined as follows: in, The primary voltage, denoted as secondary voltage, and n as the transformer turns ratio.
[0027] when At that time, the theoretical phase shift value is ;when hour, This formula is derived from the ideal ZVS boundary conditions of a dual active bridge converter, without considering non-ideal factors such as dead zone and parasitic parameters.
[0028] Theoretical loss value The physical loss model is used to calculate losses, including conduction losses and switching losses of power devices, as well as iron and copper losses of magnetic components. Because the physical model uses linear approximations and simplifying assumptions (such as ignoring dead zones and assuming linear capacitance), there are systematic deviations from the actual losses near the ZVS boundary and in the hard-switching region, which need to be compensated for using a neural network.
[0029] In some embodiments, the initial DAB loss PGNN model includes a full ZVS loss model, a primary-edge ZVS loss model, and a secondary-edge ZVS loss model. The initial ZVS boundary PGNN model, the full ZVS loss model, the primary edge ZVS loss model, and the secondary edge ZVS loss model all consist of an input layer, a hidden layer, a physical guidance layer, and an output layer.
[0030] In some embodiments, the zero-voltage switching (ZVS) boundary of a dual active bridge converter is affected by non-ideal factors such as dead-zone effect, parasitic parameters, and nonlinear junction capacitance during actual operation, resulting in significant deviations from the ideal physical model. Simultaneously, converter losses exhibit abrupt changes at the ZVS boundary, making accurate prediction difficult with a single model. This method introduces a Physics-Guided Neural Network (PGNN) as a physical guidance layer, fusing the physical model with the data-driven model to construct a hybrid physical-data model.
[0031] When there is a set of target variables Physically related input variables When dealing with data, there are generally two methods to solve the problem of model input-output relationships. One method is to train a neural network model based on the data. To generate estimates of the target variable. Another approach is to use physics-based numerical models. The target variable is obtained by simulating the physical relationship between the input variable and the target variable. Because of simplifications or omissions in physical relationships within the model, the target variable may not be fully represented, leading to discrepancies between the model and observations. Therefore, the fundamental goal of hybrid physics-data models is to combine... and This allows them to complement each other's strengths.
[0032] physical model and neural networks A simple approach to combining them is to use the simulation output based on the physical model. As an additional input to the neural network model, along with the input variables They are all fed into the network. This constitutes the following basic hybrid physical-data model:
[0033] in, This represents a physical-guided neural network model. For the expanded input vector This represents the final predicted value of the PGNN model. In this model, if the output value is based on a physics-based model... It is accurate, and consistent with Observation results If they match perfectly, then the hybrid physics-data model can learn the physical model and directly output it. .if When there is a difference between the input variable and the actual value, the input variable can be extracted from the input driver space. We learn features to supplement these differences, thereby reducing the bias of the physical fundamental model.
[0034] A schematic diagram of the hybrid physics-data model is shown below. Figure 2 As shown, the output of the physical model serves as another input variable to the neural network model, and is compared with the original input variable. The input is combined to form the final input, which is then processed by the neural network model to output the final result. The standard approach to training hybrid physics-data models is to minimize their model predictions. The learning objective of a neural network is typically expressed as the formula shown below: (This refers to the learning loss on the training set while maintaining low model complexity.)
[0035] in, The total number of training samples, For sample index, For the empirical loss function, For the first The model prediction value for each sample. For the first The true value of each sample Measuring model complexity To balance hyperparameters, the effectiveness of any such training procedure is limited by the size of the labeled training set, which is often small in many practical problems. Therefore, it cannot be guaranteed that the model trained by minimizing equations will conform to the trend of the physical model's response. Therefore, this study introduces a loss function based on a physical model to guide the learning of the neural network training model towards consistency with the physical mechanism. The following equation is used to represent the target prediction value. Other input variables The physical relationship between them:
[0036] in and Representing the general form of physics-based equations, which can involve both and Algebraic operations (such as laws of kinematics) can also involve their partial derivatives (such as the Navier-Stokes equations for fluid dynamics or the Schrödinger equation for computational chemistry). These physics-based equations must satisfy the same criteria as other loss function terms: continuous and differentiable. One way to measure whether model predictions violate these physics-based equations is to evaluate the following physics-based loss function:
[0037] in This represents the rectified linear unit function, used to handle inequality constraints. Because... No target variable needed It uses actual observations, therefore, compared to traditional loss functions, it can even be evaluated on unlabeled data instances, including... The complete learning objective of PGNN can be expressed as:
[0038] in These are hyperparameters that determine the relative weights between minimizing physical inconsistency, empirical loss, and model complexity. For example, in this embodiment of the invention, for the ZVS boundary model, take... The same value is used for the loss model. Since PGNN assumes that known physical laws also apply to any unknown data instance, using the physical consistency of the model output as a learning objective helps to achieve better generalization performance, even when the training data is small and cannot fully represent the whole.
[0039] The complete learning objective of PGNN can be solved using various algorithms, such as stochastic gradient descent and its variants, which have achieved great success in training deep neural networks. The automatic differentiation procedure provided in standard deep learning packages can also easily calculate the physical loss function. The gradient of the model parameters makes neural networks an ideal choice for incorporating physically based loss functions into the learning objectives of data science models.
[0040] An initial ZVS boundary PGNN model and an initial DAB loss PGNN model are established based on a hybrid physics-data model to achieve efficiency optimization control. The initial ZVS boundary PGNN model consists of an input layer, hidden layers, a physical guidance layer, and an output layer. The input layer receives the normalized boundary feature vector; the hidden layer is a three-layer fully connected feedforward neural network with 40 neurons per layer, and the activation function is the hyperbolic tangent function. The physical guiding layer is a parameterless identity mapping layer that maps theoretical phase shift values. The value is passed directly to the output layer, where it is added to the output of the hidden layer to obtain the final predicted value; the output layer has no activation function.
[0041] The initial DAB loss PGNN model consists of three independent subnetworks: the full ZVS loss model, the primary edge ZVS loss model, and the secondary edge ZVS loss model. Each subnetwork consists of an input layer, a hidden layer, a physical guide layer, and an output layer. The hidden layer structure is the same as that of the ZVS boundary PGNN model (three fully connected layers, 40 neurons per layer, Tanh activation), and the physical guide layer is also a parameterless identity mapping.
[0042] The S200 constructs an initial ZVS boundary PGNN model. Iterative optimization of the initial ZVS boundary PGNN model is performed using primary-side voltage, secondary-side voltage, output current, theoretical phase shift value, and tag phase shift value. After the iterative optimization is completed, the final ZVS boundary PGNN model and sample predicted phase shift value are obtained.
[0043] In some embodiments, the iterative optimization process of the initial ZVS boundary PGNN model includes: The primary side voltage, secondary side voltage, output current, and theoretical phase shift value are used as boundary feature vectors; The boundary feature vector is input into the initial ZVS boundary PGNN model. The hidden layer generates the boundary deviation based on the boundary feature vector. The theoretical phase shift value is added to the boundary deviation after being mapped by the physical guidance layer to obtain the sample predicted phase shift value. The boundary data-driven loss is calculated based on the sample predicted phase shift value and the tag phase shift value. The boundary physical guidance loss is calculated based on the sample predicted phase shift value and the theoretical phase shift value. The boundary data-driven loss and the boundary physical guidance loss are weighted and added together to obtain the total boundary loss. The initial ZVS boundary PGNN model is iteratively optimized using the Adam optimizer until the total boundary loss converges to the corresponding preset threshold, and the final ZVS boundary PGNN model is output.
[0044] In some embodiments, the iterative optimization process of the initial ZVS boundary PGNN model is as follows: Figure 3 As shown, all input variables are subjected to min-max normalization and mapped to a dimensionless interval. , primary voltage Secondary voltage Output current and theoretical phase shift value As boundary feature vectors, the normalized boundary feature vectors are input into the initial ZVS boundary PGNN model. The hidden layer generates boundary bias based on the boundary feature vectors. This deviation characterizes the difference between the actual ZVS boundary and the theoretical boundary (caused by dead zone effect, nonlinear junction capacitance, stray inductance, etc.). Theoretical phase shift value After identity mapping by the physical guidance layer, the boundary deviation is... Add them together to obtain the sample predicted phase shift value. :
[0045] The model weights are initialized using a Xavier uniform distribution to maintain consistent variance between layers. Simultaneously, the bias terms of the output layer in the residual branches (i.e., hidden layers) and the weight matrix of the last layer are forcibly initialized to zero. Under this initialization strategy, the untrained neural network outputs... This allows the model to be near the physical theoretical solution in the early stages of training, significantly accelerating convergence and improving robustness under conditions with few samples.
[0046] Training employs a composite loss function, with the loss driven by boundary data. and boundary physical guidance loss The weighted summation is used, and the boundary data-driven loss employs mean squared error to measure the difference between the predicted phase shift value and the label phase shift value. in The superscript represents the number of samples in the training batch. Indicates the first One sample.
[0047] In some embodiments, the boundary physical guidance loss is calculated based on the sample predicted phase shift value and the theoretical phase shift value, including: Set the maximum physical limit phase shift value, and add the theoretical phase shift value to the maximum physical limit phase shift value to obtain the physical upper limit of the phase shift value; The difference between the theoretical phase shift value and the sample predicted phase shift value is input into the ReLU function to calculate the first function value, and the difference between the sample predicted phase shift value and the physical upper limit of the phase shift value is input into the ReLU function to calculate the second function value. The physical penalty term corresponding to the sample is obtained by weighted summing of the first function value and the second function value, and the physical penalty terms of all samples are summed and averaged to obtain the boundary physical guidance loss.
[0048] In some embodiments, the design of the boundary physical guidance loss is based on physical priors. Due to the existence of dead time, the phase shift value required to actually achieve ZVS is usually greater than or equal to the calculated value of the ideal model. Therefore, the theoretical phase shift value constitutes the physical lower bound of the actual boundary. At the same time, the actual phase shift value has a physical upper bound due to the limitations of the maximum switching frequency and dead time. A maximum physical limit phase shift value is set. (by the highest frequency) With dead time (The calculated basic offset of 0.06 was calibrated after considering the influence of nonlinear junction capacitance). The theoretical phase shift value was compared with... Adding them together gives the physical upper limit of the phase shift value. :
[0049] For each sample, calculate the physical penalty term. : in The first term uses a linear penalty. The first term represents the weighting coefficient, which can be set to 1.0 for example. The second term uses a squared penalty to form a flexible wall; the gradient is small when there is a slight deviation to tolerate unmodeled parasitic biases, while the penalty increases sharply when there is severe divergence. The weighting coefficient, for example, can be taken as 0.5. The boundary physical guidance loss is obtained by summing and averaging the physical penalty terms over all samples. :
[0050] The total boundary loss is obtained by weighted summation of the two losses. : Weighting coefficients This is used to balance the accuracy of data fitting with physical consistency.
[0051] The AdamW optimizer was used to iteratively optimize the parameters of the initial ZVS boundary PGNN model, with the initial learning rate set to [value missing]. In each iteration, the total boundary loss is calculated and the weights are updated via backpropagation. Experiments show that the loss function value decreases with the number of iterations, reaching a minimum after approximately 100 iterations. The initial magnitude is small, with subsequent changes being slow. In practice, training is stopped after 300 to 500 iterations, outputting the final ZVS boundary PGNN model. Simultaneously, for each sample in the training set, the predicted phase shift value output by the model is recorded. .
[0052] S300 processes the theoretical loss value into regions based on the primary voltage, secondary voltage, transformer turns ratio, output phase shift value, and sample predicted phase shift value to obtain the full ZVS theoretical loss value, the primary ZVS loss theoretical loss value, and the secondary ZVS loss theoretical loss value.
[0053] In some embodiments, the region-based processing includes: The voltage gain is calculated based on the primary voltage, secondary voltage, and transformer turns ratio. When the output phase shift value is greater than or equal to the sample predicted phase shift value, the theoretical loss value is corrected by the primary side turn-off loss and the secondary side turn-off loss to obtain the full ZVS theoretical loss value. When the output phase shift value is less than the sample predicted phase shift value and the voltage gain is greater than 1, the theoretical loss value is corrected by the primary side turn-on and turn-off losses and the secondary side turn-off losses to obtain the theoretical loss value of the primary side ZVS loss. When the output phase shift value is less than the sample predicted phase shift value and the voltage gain is less than 1, the theoretical loss value is corrected by the primary side turn-off loss and the secondary side turn-on and turn-off loss to obtain the theoretical loss value of the secondary side ZVS loss.
[0054] In some embodiments, since the converter loss exhibits abrupt changes at the ZVS boundary (the turn-on loss is essentially eliminated when soft switching is implemented, and additional turn-on loss needs to be considered when it is not implemented), it is necessary to divide the operating space into different regions according to the ZVS boundary and calculate the theoretical loss value for each region independently as the physical guidance input for subsequent sub-networks.
[0055] First, calculate the voltage gain based on the primary voltage, secondary voltage, and transformer turns ratio. : Output phase shift value Phase shift value of the sample prediction Compare and combine voltage gain Perform regional processing, theoretical loss value ,in This represents the total power of the switching losses of the power devices. This represents the total power loss during the conduction of the power device. For iron loss, This is due to copper loss.
[0056] when At this time, both the primary and secondary sides achieve ZVS, and the turn-on loss is negligible. Turn-off loss is achieved through the primary side. and secondary side shutdown loss Theoretical loss value Make corrections, so that Obtain the theoretical loss value of the entire ZVS. The calculation of primary-side turn-off loss and secondary-side turn-off loss is based on the parameters in the switching device datasheet:
[0057] in It is the energy consumed by the switching device during the turn-off process. It is the voltage of the switching device at the moment of turn-off. It is the current of the switching device at the moment of turn-off. This represents the inductor current at a specific moment.
[0058] when and At that time, the primary edge did not achieve ZVS, but the secondary edge did. The switching losses are determined by the primary edge's turn-on and turn-off operations. and secondary side shutdown loss Correct the theoretical loss value, so that Obtain the theoretical loss value of the original edge ZVS loss. The primary-side turn-on and turn-off losses include both turn-on and turn-off losses:
[0059] when and At that time, the secondary edge did not achieve ZVS, but the primary edge did. This is achieved by turning off the loss through the primary edge. and secondary side turn-on and turn-off losses Correct the theoretical loss value, so that Obtain the theoretical loss value of secondary edge ZVS loss. The secondary-side turn-on and turn-off losses are approximated using a MOSFET linearization model:
[0060] in It is the drain-source voltage of the MOSFET at the moment of turn-on. It is the drain current of the MOSFET at the moment of turn-on. It refers to the turn-on time of the MOSFET. f It is the switching frequency.
[0061] Through the above regional processing, the theoretical loss values corresponding to three different regions are obtained, which will be used as the physical guidance inputs for three independent sub-networks in step four.
[0062] The S400 constructs an initial DAB loss PGNN model. Iterative optimization is then performed on the initial DAB loss PGNN model using primary voltage, transformer turns ratio, secondary voltage, switching frequency, output phase shift value, theoretical loss value of full ZVS, theoretical loss value of primary ZVS loss, theoretical loss value of secondary ZVS loss, and tag loss value. After the iterative optimization is completed, the final DAB loss PGNN model is obtained.
[0063] In some embodiments, the iterative optimization process of the initial DAB loss PGNN model includes: The primary voltage, transformer turns ratio, secondary voltage, switching frequency, and output phase shift value are used as the operating condition feature vector. The first loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the full ZVS. The second loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the primary ZVS loss. The third loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the secondary ZVS loss. The first loss feature vector is input into the full ZVS loss model. The hidden layer generates the first loss deviation based on the first loss feature vector. The full ZVS theoretical loss value is mapped by the physical guidance layer and added to the first loss deviation to obtain the loss prediction value of the first training process. The first loss data-driven loss is calculated based on the first training process loss prediction value and the label loss value. The first loss physical guidance loss is calculated based on the first training process loss prediction value and the full ZVS theoretical loss value. The first loss data-driven loss and the first loss physical guidance loss are weighted and added together to obtain the first loss total loss. The second loss feature vector is input into the primary edge ZVS loss model, and the third loss feature vector is input into the secondary edge ZVS loss model. The same calculation process as the full ZVS loss model is used to obtain the total loss of the second loss and the total loss of the third loss, respectively. The initial DAB loss PGNN model is iteratively optimized using the Adam optimizer until the total loss of the first loss, the total loss of the second loss, and the total loss of the third loss all converge to the corresponding preset thresholds, thus obtaining the final DAB loss PGNN model.
[0064] In some embodiments, the iterative optimization process of the initial DAB loss PGNN model is as follows: Figure 4 As shown, the primary voltage Transformer turns ratio Secondary voltage Switching frequency and output phase shift value As the operating condition feature vector. To construct the hybrid physical-data model, the theoretical loss value is used as an additional input: consisting of the operating condition feature vector and the full ZVS theoretical loss value. The first loss feature vector is formed by the operating condition feature vector and the theoretical loss value of the original edge ZVS. The second loss feature vector is formed by the operating condition feature vector and the theoretical loss value of the secondary ZVS. The third loss feature vector is formed, and all features are also processed by minimax normalization.
[0065] The first loss feature vector is input into the full ZVS loss model, and the hidden layer generates the first loss bias based on this vector. Theoretical loss value of full ZVS After identity mapping by the physical guidance layer Add them together to obtain the predicted loss value for the first training process. : Similarly, the predicted loss value in the second training process The predicted loss value during the third training process .
[0066] The training of each sub-network also employs a composite loss function. Taking the full ZVS loss model as an example: the first loss data-driven loss is calculated based on the loss prediction value and the label loss value from the first training process. :
[0067] The physical guidance loss of the first loss is calculated based on the predicted loss value of the first training process and the theoretical loss value of the full ZVS. : The weighted sum of the two yields the first total loss. : Weighting coefficients Preferred .
[0068] The primary-side ZVS loss model and the secondary-side ZVS loss model use the exact same process, respectively using their theoretical loss values ( and ) and predicted loss values, to obtain the second total loss. and the third loss total loss .
[0069] The AdamW optimizer was used to iteratively optimize the parameters of the initial DAB loss PGNN model, with the initial learning rate set to 1. The three sub-networks are trained simultaneously, with each sub-network's loss function decreasing independently. The threshold for the loss function of each of the three sub-networks is set to... Experiments show that the ZVS region has the largest amount of data and its convergence curve is the most stable; the two hard-switching regions also did not exhibit drastic oscillations, and the loss function values of all three regions decreased to a low level after 400 to 500 iterations. After training, the final DAB loss PGNN model is obtained.
[0070] The model was input with 100 discrete working condition sample points from the validation set for prediction, and the results were compared with the actual loss values from PLECS simulation. The average relative error between the predicted and actual values was: The mean absolute error is In the low-loss region ( ) and high-loss area ( The model maintains high accuracy, demonstrating that it overcomes the problem of overfitting in sparse regions of traditional data-driven methods.
[0071] The S500 consists of a comprehensive loss prediction model composed of the final ZVS boundary PGNN model and the final DAB loss PGNN model. It obtains the actual operating condition data set, inputs the actual operating condition data set into the comprehensive loss prediction model to predict the loss, obtains the loss prediction value during operation, and adjusts the operation of the target dual active bridge converter according to the loss prediction value during operation.
[0072] In some embodiments, the actual operating condition data set is input into the comprehensive loss prediction model to predict the loss and obtain the predicted loss value during operation, including: The actual working condition data set is input into the final ZVS boundary PGNN model to predict the phase shift value and obtain the actual predicted phase shift value. Based on the actual predicted phase shift value, the region is determined to obtain the region type, which includes full ZVS region, primary edge ZVS lost region and secondary edge ZVS lost region. The actual operating condition data set is input into the final DAB loss PGNN model, which includes a full ZVS loss model, a primary-edge ZVS loss model, and a secondary-edge ZVS loss model. When the region type is a full ZVS region, loss prediction is performed using the full ZVS loss model; when the region type is a primary-edge ZVS loss region, loss prediction is performed using the primary-edge ZVS loss model; when the region type is a secondary-edge ZVS loss region, loss prediction is performed using the secondary-edge ZVS loss model. The predicted loss value during operation is output.
[0073] In some embodiments, the comprehensive loss prediction model is used to predict the total loss of the dual active bridge converter in real time under actual operating conditions, and adjust the control parameters of the converter accordingly to make the converter operate in the optimal efficiency state.
[0074] The primary voltage in the actual operating condition data set Secondary voltage and output current Input the final ZVS boundary PGNN model. Internally, this model performs forward propagation: first, it performs min-max normalization on the input features, and then the hidden layers generate the boundary bias. The physical guiding layer will change the theoretical phase shift value. The identity is mapped to the output layer and added to the bias value to output the actual predicted phase shift value.
[0075] The final DAB loss PGNN model contains three sub-networks: the full ZVS loss model, the primary edge ZVS loss model, and the secondary edge ZVS loss model.
[0076] The primary voltage in the actual operating condition data set Transformer turns ratio Secondary voltage Switching frequency and output phase shift value Together with the theoretical loss value corresponding to the selected subnetwork, i.e., the theoretical loss value of the entire ZVS Theoretical loss value of ZVS loss at the original edge Or the theoretical loss value of the secondary side ZVS is lost. This forms the loss feature vector. This vector is then input into the selected sub-network, where forward propagation is performed: the hidden layers generate the loss bias. The physical guidance layer maps the theoretical loss value to the output layer and adds it to the deviation to output the predicted loss value during operation. .
[0077] Obtain predicted values of operating process losses Then, this value is used as the objective function to adjust the control parameters of the dual active bridge converter. In this embodiment, the adjustable control parameter is the phase shift value. and switching frequency The adjustment objective is to minimize the predicted loss value.
[0078] The specific adjustment strategies are as follows: Offline optimization: An optimal control table is pre-stored in the controller, which is obtained by offline scanning of all possible phase shift values. and switching frequency (In this embodiment) , The predicted loss for each combination is calculated using a comprehensive loss prediction model, and the combination that minimizes the loss is identified. Combine and store in a table. During actual runtime, based on the current... , , The optimal control parameters are obtained by looking up the table and then directly applied to the converter.
[0079] Online optimization: If computational resources are sufficient, online optimization can be performed in each control cycle or every few cycles. For example, the golden ratio search or coordinate descent method can be used to fine-tune the operation around the current operating point. and The comprehensive loss prediction model is invoked to assess loss changes, and parameters are adjusted in the direction of decreasing loss. To prevent system instability caused by frequent adjustments, an adjustment dead zone can be set (e.g., only adjusting when the predicted loss is lower than the current operating point). Adjustments are only implemented when the above conditions are met.
[0080] Hybrid strategy: First, use the offline optimal table for coarse adjustment, and then perform small-scale online fine-tuning around the table, balancing fast response and fine optimization.
[0081] The adjustment action is achieved by generating a pulse width modulation (PWM) signal through a digital controller (such as a DSP or FPGA): changing the phase shift value. That is, adjusting the phase difference between the primary and secondary H-bridge drive signals; changing the switching frequency. This involves adjusting the PWM carrier period. The controller will then optimize the... and Write the PWM generator, and the converter will then run under the new control parameters.
[0082] In actual operation, new operating data (actual operating conditions and corresponding true loss values) can be collected periodically to incrementally learn or periodically retrain the comprehensive loss prediction model, enabling the model to adapt to long-term changes such as device aging and temperature drift. The updated model can regenerate the optimal control table, thereby achieving adaptive optimization control.
[0083] Reference Figure 5 This invention provides a PGNN-based dual active bridge converter adjustment device 50 for implementing a PGNN-based dual active bridge converter adjustment method. The device includes: The training data acquisition module 51 is used to obtain the training operating condition data set, tag phase shift value and tag loss value of the target dual active bridge converter through simulation and experimental acquisition. The training operating condition data set includes primary voltage, secondary voltage, transformer turns ratio, output current, switching frequency, output phase shift value, theoretical phase shift value and theoretical loss value. ZVS boundary PGNN model training module 52 is used to construct an initial ZVS boundary PGNN model. The initial ZVS boundary PGNN model is iteratively optimized using primary side voltage, secondary side voltage, output current, theoretical phase shift value, and label phase shift value. After the iterative optimization is completed, the final ZVS boundary PGNN model and sample predicted phase shift value are obtained. The regional processing module 53 is used to process the theoretical loss value into regions based on the primary voltage, secondary voltage, transformer turns ratio, output phase shift value and sample predicted phase shift value, to obtain the full ZVS theoretical loss value, the primary ZVS loss theoretical loss value and the secondary ZVS loss theoretical loss value. The DAB loss PGNN model training module 54 is used to construct the initial DAB loss PGNN model. Iterative optimization of the initial DAB loss PGNN model is performed using primary voltage, transformer turns ratio, secondary voltage, switching frequency, output phase shift value, full ZVS theoretical loss value, primary ZVS loss theoretical loss value, secondary ZVS loss theoretical loss value, and tag loss value. After the iterative optimization is completed, the final DAB loss PGNN model is obtained. The DAB operation adjustment module 55 is used to form a comprehensive loss prediction model composed of the final ZVS boundary PGNN model and the final DAB loss PGNN model, obtain the actual operating condition data set, input the actual operating condition data set into the comprehensive loss prediction model to perform loss prediction, obtain the loss prediction value during operation, and adjust the operation of the target dual active bridge converter according to the loss prediction value during operation.
[0084] This application provides an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the PGNN-based dual active bridge converter regulation method of any of the above schemes.
[0085] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.
[0086] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.
[0087] This application also provides a computer-readable medium storing a computer program that, when executed by a processor, implements the PGNN-based dual active bridge converter regulation method described above. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.
[0088] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.
[0089] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for regulating a dual active bridge converter based on PGNN, characterized in that, include: The training operating condition data set, tag phase shift value, and tag loss value of the target dual active bridge converter were obtained through simulation and experimental acquisition. The training operating condition data set includes primary voltage, secondary voltage, transformer turns ratio, output current, switching frequency, output phase shift value, theoretical phase shift value, and theoretical loss value. An initial ZVS boundary PGNN model is constructed. The initial ZVS boundary PGNN model is iteratively optimized using primary side voltage, secondary side voltage, output current, theoretical phase shift value, and tag phase shift value. After the iterative optimization is completed, the final ZVS boundary PGNN model and sample predicted phase shift value are obtained. The theoretical loss values are processed by region based on the primary voltage, secondary voltage, transformer turns ratio, output phase shift value, and sample predicted phase shift value to obtain the full ZVS theoretical loss value, the primary ZVS loss theoretical loss value, and the secondary ZVS loss theoretical loss value. An initial DAB loss PGNN model is constructed. The initial DAB loss PGNN model is then iteratively optimized using primary voltage, transformer turns ratio, secondary voltage, switching frequency, output phase shift value, theoretical loss value of full ZVS, theoretical loss value of primary ZVS loss, theoretical loss value of secondary ZVS loss, and tag loss value. After the iterative optimization is completed, the final DAB loss PGNN model is obtained. A comprehensive loss prediction model is formed by the final ZVS boundary PGNN model and the final DAB loss PGNN model. The actual operating condition data set is obtained and input into the comprehensive loss prediction model to predict the loss, thereby obtaining the loss prediction value during operation. The operation of the target dual active bridge converter is adjusted according to the loss prediction value during operation.
2. The method for adjusting a dual active bridge converter based on PGNN according to claim 1, characterized in that, The initial DAB loss PGNN model includes the full ZVS loss model, the primary edge ZVS loss model, and the secondary edge ZVS loss model; The initial ZVS boundary PGNN model, the full ZVS loss model, the primary edge ZVS loss model, and the secondary edge ZVS loss model all consist of an input layer, a hidden layer, a physical guidance layer, and an output layer.
3. The method for adjusting a dual active bridge converter based on PGNN according to claim 2, characterized in that, The iterative optimization process of the initial ZVS boundary PGNN model includes: The primary side voltage, secondary side voltage, output current, and theoretical phase shift value are used as boundary feature vectors; The boundary feature vector is input into the initial ZVS boundary PGNN model. The hidden layer generates the boundary deviation based on the boundary feature vector. The theoretical phase shift value is added to the boundary deviation after being mapped by the physical guidance layer to obtain the sample predicted phase shift value. The boundary data-driven loss is calculated based on the sample predicted phase shift value and the tag phase shift value. The boundary physical guidance loss is calculated based on the sample predicted phase shift value and the theoretical phase shift value. The boundary data-driven loss and the boundary physical guidance loss are weighted and added together to obtain the total boundary loss. The initial ZVS boundary PGNN model is iteratively optimized using the Adam optimizer until the total boundary loss converges to the corresponding preset threshold, and the final ZVS boundary PGNN model is output.
4. The method for adjusting a dual active bridge converter based on PGNN according to claim 3, characterized in that, The boundary physical guidance loss is calculated based on the predicted phase shift value and the theoretical phase shift value, including: Set the maximum physical limit phase shift value, and add the theoretical phase shift value to the maximum physical limit phase shift value to obtain the physical upper limit of the phase shift value; The difference between the theoretical phase shift value and the sample predicted phase shift value is input into the ReLU function to calculate the first function value, and the difference between the sample predicted phase shift value and the physical upper limit of the phase shift value is input into the ReLU function to calculate the second function value. The physical penalty term corresponding to the sample is obtained by weighted summing of the first function value and the second function value, and the physical penalty terms of all samples are summed and averaged to obtain the boundary physical guidance loss.
5. The method for adjusting a dual active bridge converter based on PGNN according to claim 1, characterized in that, Regional processing includes: The voltage gain is calculated based on the primary voltage, secondary voltage, and transformer turns ratio. When the output phase shift value is greater than or equal to the sample predicted phase shift value, the theoretical loss value is corrected by the primary side turn-off loss and the secondary side turn-off loss to obtain the full ZVS theoretical loss value. When the output phase shift value is less than the sample predicted phase shift value and the voltage gain is greater than 1, the theoretical loss value is corrected by the primary side turn-on and turn-off losses and the secondary side turn-off losses to obtain the theoretical loss value of the primary side ZVS loss. When the output phase shift value is less than the sample predicted phase shift value and the voltage gain is less than 1, the theoretical loss value is corrected by the primary side turn-off loss and the secondary side turn-on and turn-off loss to obtain the theoretical loss value of the secondary side ZVS loss.
6. The method for adjusting a dual active bridge converter based on PGNN according to claim 2, characterized in that, The iterative optimization process of the initial DAB loss PGNN model includes: The primary voltage, transformer turns ratio, secondary voltage, switching frequency, and output phase shift value are used as the operating condition feature vector. The first loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the full ZVS. The second loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the primary ZVS loss. The third loss feature vector is composed of the operating condition feature vector and the theoretical loss value of the secondary ZVS loss. The first loss feature vector is input into the full ZVS loss model. The hidden layer generates the first loss deviation based on the first loss feature vector. The full ZVS theoretical loss value is mapped by the physical guidance layer and added to the first loss deviation to obtain the loss prediction value of the first training process. The first loss data-driven loss is calculated based on the first training process loss prediction value and the label loss value. The first loss physical guidance loss is calculated based on the first training process loss prediction value and the full ZVS theoretical loss value. The first loss data-driven loss and the first loss physical guidance loss are weighted and added together to obtain the first loss total loss. The second loss feature vector is input into the primary edge ZVS loss model, and the third loss feature vector is input into the secondary edge ZVS loss model. The same calculation process as the full ZVS loss model is used to obtain the total loss of the second loss and the total loss of the third loss, respectively. The initial DAB loss PGNN model is iteratively optimized using the Adam optimizer until the total loss of the first loss, the total loss of the second loss, and the total loss of the third loss all converge to the corresponding preset thresholds, thus obtaining the final DAB loss PGNN model.
7. The method for adjusting a dual active bridge converter based on PGNN according to claim 1, characterized in that, The actual operating condition data set is input into the comprehensive loss prediction model to predict the loss, and the predicted loss values during operation are obtained, including: The actual working condition data set is input into the final ZVS boundary PGNN model to predict the phase shift value and obtain the actual predicted phase shift value. Based on the actual predicted phase shift value, the region is determined to obtain the region type, which includes full ZVS region, primary edge ZVS lost region and secondary edge ZVS lost region. The actual operating condition data set is input into the final DAB loss PGNN model, which includes a full ZVS loss model, a primary-edge ZVS loss model, and a secondary-edge ZVS loss model. When the region type is a full ZVS region, loss prediction is performed using the full ZVS loss model; when the region type is a primary-edge ZVS loss region, loss prediction is performed using the primary-edge ZVS loss model; when the region type is a secondary-edge ZVS loss region, loss prediction is performed using the secondary-edge ZVS loss model. The predicted loss value during operation is output.
8. A PGNN-based dual active bridge converter adjustment device, used to implement the PGNN-based dual active bridge converter adjustment method as described in any one of claims 1 to 7, characterized in that, The device includes: The training data acquisition module is used to obtain the training operating condition data set, tag phase shift value, and tag loss value of the target dual active bridge converter through simulation and experimental acquisition. The training operating condition data set includes primary voltage, secondary voltage, transformer turns ratio, output current, switching frequency, output phase shift value, theoretical phase shift value, and theoretical loss value. The ZVS boundary PGNN model training module is used to construct the initial ZVS boundary PGNN model. It iteratively optimizes the initial ZVS boundary PGNN model using primary side voltage, secondary side voltage, output current, theoretical phase shift value, and label phase shift value. After the iterative optimization is completed, the final ZVS boundary PGNN model and sample predicted phase shift value are obtained. The regional processing module is used to process the theoretical loss value into regions based on the primary voltage, secondary voltage, transformer turns ratio, output phase shift value and sample predicted phase shift value, to obtain the full ZVS theoretical loss value, the primary ZVS loss theoretical loss value and the secondary ZVS loss theoretical loss value. The DAB loss PGNN model training module is used to construct the initial DAB loss PGNN model. Iterative optimization of the initial DAB loss PGNN model is performed using primary voltage, transformer turns ratio, secondary voltage, switching frequency, output phase shift value, theoretical loss value of full ZVS, theoretical loss value of primary ZVS loss, theoretical loss value of secondary ZVS loss, and tag loss value. After the iterative optimization is completed, the final DAB loss PGNN model is obtained. The DAB operation adjustment module is used to obtain the actual operating condition data set by the comprehensive loss prediction model composed of the final ZVS boundary PGNN model and the final DAB loss PGNN model. The actual operating condition data set is input into the comprehensive loss prediction model to predict the loss and obtain the loss prediction value during operation. The operation adjustment is performed on the target dual active bridge converter based on the loss prediction value during operation.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the PGNN-based dual active bridge converter adjustment method as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the PGNN-based dual active bridge converter adjustment method as described in any one of claims 1 to 7.
