Control system and method for realizing soft switching of resonant converter in full load range

Abstract

The invention relates to the technical field of resonant converter control, in particular to a control system and method for realizing full-load range soft switching of a resonant converter. Collecting current, voltage, temperature and time sequence data of a wide temperature zone and different aging stages, and extracting a two-dimensional time-frequency energy spectrum to construct an enhanced data set; training a mixed time sequence neural network based on the data set, fusing a time domain waveform, a time frequency graph and temperature, predicting a next period resonance dynamic state under candidate control parameters, and giving a soft switching confidence coefficient; in each control period, the switching frequency or the phase shift adjustment amount is used as a variable, an optimization problem is solved, and the optimal control amount is output in real time; kalman filtering is utilized to identify equivalent resonant inductance and capacitance parameters, and the equivalent resonant inductance and capacitance parameters are fed back to a prediction and optimization link as dynamic characteristics, so that adaptive correction of parameter drift is realized, an optimal soft switching working point is continuously tracked and maintained, efficiency and reliability are improved, and the service life is prolonged.

Description

Technical Field

[0001] This application relates to the field of resonant converter control technology, specifically a control system and method for realizing soft switching of a resonant converter across the entire load range. Background Technology

[0002] Resonant converters are widely used because they enable soft switching of power transistors, which is crucial for reducing switching losses and improving converter efficiency and power density.

[0003] In practical applications, the optimal soft-switching operating point of a resonant converter is not a fixed static point. It changes dynamically with variations in load, input voltage, ambient temperature, and other operating conditions. Furthermore, throughout the converter's lifespan, due to the aging of key components (such as resonant inductors and resonant capacitors), this optimal operating point will undergo irreversible and slow drift.

[0004] Existing control methods often exhibit limitations when dealing with such complex, multivariable, and life-cycle dynamic changes. These methods struggle to track an optimal operating point that is itself constantly drifting in real time and with high precision. As a result, the soft-switching performance and efficiency of the converter gradually decline after deviating from the ideal design conditions or after long-term operation, failing to maintain optimal operating conditions across the entire load range and life cycle. Therefore, how to achieve a control system that can continuously track and maintain the dynamically changing optimal soft-switching operating point across a wide operating range and the entire life cycle is a current technical challenge in the field of resonant converter technology.

[0005] In view of this, this application proposes a control system and method for realizing soft switching of a resonant converter across the entire load range. Summary of the Invention

[0006] To achieve the above objectives, this application provides a control system and method for realizing soft switching of a resonant converter across the entire load range, the specific technical solution of which is as follows:

[0007] A control method for achieving soft switching across the entire load range of a resonant converter includes:

[0008] The original datasets of resonant current, resonant voltage, device temperature and switching timing are collected. The resonant current and voltage waveforms in the original datasets are processed by time-frequency transformation to generate a two-dimensional time-frequency energy spectrum, and an enhanced dataset is constructed with the original dataset.

[0009] The sequence prediction model is trained based on the augmented dataset. The sequence prediction model takes historical time-domain waveforms, corresponding time-frequency features and current device temperature data as inputs to predict the resonant dynamics under given candidate control parameters, and calculates the soft-switching confidence based on the predicted resonant dynamics.

[0010] In each control cycle, different candidate control parameters are input into the sequence prediction model to obtain the soft-switching confidence corresponding to each candidate control parameter. An optimization problem with the goal of maximizing the soft-switching confidence is solved, and the optimal control parameters are output under the constraint of the safe operating area.

[0011] During each soft-switching cycle, the current values ​​of the equivalent resonant inductance and capacitance are identified in real time based on the resonant current and voltage waveforms.

[0012] The identified equivalent resonant inductance and capacitance are added as new dynamic features and input into the sequence prediction model in the next control cycle. This allows the prediction results to be adjusted according to parameter drift, the control output to be dynamically corrected, and the optimal soft-switching operating point to be continuously tracked.

[0013] Preferably, the resonant current, resonant voltage, device temperature, and switching timing of the resonant converter are collected, and the switching timing includes frequency and dead time;

[0014] Continuous wavelet transform is performed on the resonant current and resonant voltage waveforms acquired in each switching cycle. After wavelet transform, the squared modulus of the wavelet coefficients is calculated to obtain a two-dimensional matrix, which is the two-dimensional time-frequency energy spectrum of the current and voltage signals. In the two-dimensional time-frequency energy spectrum, time is the horizontal axis and frequency is the vertical axis, and the gray value represents the energy density of the signal at the time and frequency points.

[0015] The generated two-dimensional time-frequency energy map is structurally integrated with the original dataset to construct an enhanced dataset.

[0016] Preferably, a hybrid temporal neural network is constructed as the sequence prediction model. The hybrid temporal neural network structure integrates convolutional neural networks and long short-term memory networks. The convolutional neural network is used to process two-dimensional time-frequency energy maps, and the long short-term memory network is used to process time-series data.

[0017] The inputs to the hybrid time-series neural network include: time-domain waveform sequences of resonant current and resonant voltage during historical switching cycles, a two-dimensional time-frequency energy spectrum sequence corresponding to the historical time-domain waveform sequences, the device temperature measured during the current control cycle, and candidate control parameters for prediction.

[0018] Preferably, in the sequence prediction model, the two-dimensional time-frequency energy map of the historical period is fed into the convolutional neural network branch as image data. The convolutional neural network is composed of multiple convolutional layers, activation function layers and pooling layers stacked together. Deep spatial features that can characterize the local time-frequency structure and non-stationary features of the signal are extracted from the two-dimensional time-frequency energy map.

[0019] After being processed by a convolutional neural network, each two-dimensional time-frequency energy spectrum is encoded into a low-dimensional feature vector, forming a sequence of feature vectors.

[0020] The input to the sequence prediction model also includes scalar data of the current device temperature and candidate control parameters; the device temperature and candidate control parameters are concatenated with the feature vector sequence extracted by the convolutional neural network and the time-domain waveform sequence of the resonant current and resonant voltage at the feature level to form a composite feature sequence;

[0021] The composite feature sequence is fed into the long short-term memory network, which is responsible for learning the complex dynamic law of the working state of the entire resonant converter evolving over time.

[0022] After being decoded by the fully connected layer, the final hidden state output of the Long Short-Term Memory network predicts the complete resonant current waveform and resonant voltage waveform of the next switching cycle under the action of candidate control parameters.

[0023] After obtaining the predicted resonant current waveform and resonant voltage waveform for the next cycle, the soft-switching confidence is calculated. The soft-switching confidence is used to quantify the extent to which soft-switching is achieved.

[0024] Preferably, a multi-objective optimization problem is constructed, with two optimization objectives: maximizing the soft-switching confidence and maximizing the operating efficiency. The operating efficiency is obtained by calculating the ratio of the predicted input power to the output power. The predicted input power is obtained by integrating the predicted input voltage and input current over one soft-switching cycle, and the input current is directly related to the resonant current.

[0025] The two optimization objectives are merged into a unified performance index function by weighted summation.

[0026] During the optimization process, it is ensured that all candidate solutions are within the safe operating range of the resonant converter, and the corresponding constraints are set based on the predicted waveform.

[0027] The optimization process of a multi-objective problem is to traverse all candidate control parameters that satisfy the safety constraints in the candidate adjustment set and find the optimal control parameter that maximizes the comprehensive performance index function.

[0028] The optimal control parameters obtained will be used as the actual control command for the next control cycle and output to the drive circuit.

[0029] Preferably, at the end of each soft-switching cycle, waveform data of the resonant current and resonant capacitor voltage within that cycle are acquired, and feature points are accurately extracted using digital signal processing algorithms, including the zero-crossing time of the resonant current, the peak time and amplitude of the resonant current, and the peak time and amplitude of the resonant capacitor voltage.

[0030] The specific steps of the digital signal processing algorithm include: For the resonant current waveform, the digital signal processing algorithm identifies the zero-crossing time by detecting the change in the sign of adjacent sampling points, and at the same time finds the local extreme points in the waveform, that is, it determines the peak time and corresponding amplitude of the resonant current by the position of the sign change of the first derivative; similarly, the digital signal processing algorithm processes the resonant capacitor voltage waveform in the same way to find the peak time and amplitude of the resonant capacitor voltage.

[0031] The extracted timestamps and amplitudes constitute a feature vector describing the dynamic behavior of the switching cycle.

[0032] Preferably, the Kalman filter algorithm is applied to estimate parameters, a state-space model describing the dynamics of the resonant circuit is established, the parameters to be identified are used as state variables, and the reciprocals of the parameters are selected as state vectors.

[0033] The observation equations of the Kalman filter algorithm are constructed, which associate state variables with measurable physical quantities. Based on the electromagnetic relationship of the resonant circuit, a linear observation model is established using waveform integration.

[0034] Preferably, at the end of each control cycle, an iterative calculation of the Kalman filter is performed, including two stages: prediction and update. In the prediction stage, the state vector and covariance of the current control cycle are predicted based on the results of the previous control cycle. In the update stage, the predicted values ​​are corrected using the observations of the current control cycle, the Kalman gain is calculated, and the optimal posterior estimate of the state vector of the current control cycle is obtained. By taking the reciprocal, the identification values ​​of the equivalent resonant inductance and capacitance in the current control cycle can be obtained.

[0035] Preferably, at the end of each control cycle, the equivalent resonant inductance and capacitance identified online are used as new dynamic features; the dynamic features are integrated with the input data of historical time-domain waveforms and two-dimensional time-frequency energy maps to form an enhanced input tensor for prediction of the next control cycle;

[0036] The enhanced input tensor is fed into the sequence prediction model to obtain real-time calibrated prediction results;

[0037] The control output decision is dynamically corrected based on the calibrated prediction results, and the optimal soft-switching operating point is tracked in a dynamic manner.

[0038] A control system for implementing soft switching of a resonant converter across the entire load range, which is used to implement the control method for implementing soft switching of a resonant converter across the entire load range, includes: a data feature construction module, a timing model training module, an online optimization decision module, a parameter identification module, and a dynamic closed-loop correction module;

[0039] The data feature construction module is used to collect the original datasets of resonant current, resonant voltage, device temperature and switching timing, perform time-frequency transformation on the resonant current and voltage waveforms in the original dataset to generate a two-dimensional time-frequency energy spectrum, and construct an enhanced dataset with the original dataset.

[0040] The timing model training module trains a sequence prediction model based on an augmented dataset. The sequence prediction model takes historical time-domain waveforms, corresponding time-frequency features, and current device temperature data as inputs to predict the resonant dynamics under given candidate control parameters, and calculates the soft-switching confidence based on the predicted resonant dynamics.

[0041] The online optimization decision module inputs different candidate control parameters into the sequence prediction model in each control cycle to obtain the soft-switching confidence corresponding to each candidate control parameter, solves the optimization problem with the goal of maximizing the soft-switching confidence, and outputs the optimal control parameters under the constraint of the safe operating area.

[0042] The parameter identification module identifies the current values ​​of the equivalent resonant inductance and capacitance in real time based on the resonant current and voltage waveforms during each soft-switching cycle.

[0043] The dynamic closed-loop correction module uses the identified equivalent resonant inductance and capacitance as new dynamic features, and inputs them into the sequence prediction model in the next control cycle to adjust the prediction results according to parameter drift, dynamically correct the control output, and continuously track the optimal soft-switching operating point.

[0044] The beneficial effects of this application are as follows: By introducing data from a wide temperature range and aging stage and performing continuous wavelet transform, this application obtains time-frequency characteristics that are more sensitive to non-stationary resonant dynamics; the enhanced dataset can significantly improve the robustness of the model to parameter drift, temperature changes and noise, and provide more comprehensive and reliable prior information for subsequent prediction and control.

[0045] This application constructs a hybrid temporal neural network that integrates time-domain waveforms, time-frequency energy spectra, and temperature to improve the prediction accuracy and generalization ability of the next cycle resonant dynamics. It also outputs soft-switching confidence simultaneously, enabling the controller to achieve the possibility of ZVS / ZCS through energy evaluation, and providing a measurable basis for optimal control decision-making.

[0046] This application performs constrained online optimization within each control cycle, comprehensively considering soft-switching confidence and efficiency targets to ensure the optimal switching frequency or phase shift is selected in real time within the safe operating range. This closed-loop strategy can quickly adapt to load and input disturbances, reduce switching losses and suppress stress, thereby improving the overall energy efficiency and stability of the system.

[0047] This application utilizes zero-crossing and peak information in conjunction with Kalman filtering for parameter identification, enabling real-time estimation of equivalent inductance and capacitance under noise and measurement uncertainties. This method exhibits fast convergence and high accuracy, and can reflect parameter changes caused by device thermal drift and aging online, providing a reliable basis for model adaptive updates.

[0048] This application injects the online-identified equivalent inductance and capacitance as dynamic features into the sequence prediction model, enabling adaptive correction of parameter drift during prediction and optimization. This allows the controller to continuously track the optimal soft-switching operating point, reducing misjudgments and control jitter, and maintaining long-term stable operation with high efficiency and low stress.

[0049] The technical solution presented in this application achieves soft-switching adaptive control covering the entire load range, balancing efficiency, reliability, and safety. Compared to traditional parameter-based or empirical strategies, it can significantly reduce switching and conduction losses, thermal stress, and EMI, and extend device lifespan. It maintains high efficiency and steady-state performance even under input fluctuations, temperature drift, and aging conditions, making it suitable for applications with stringent requirements for high power density and high reliability, such as server power supplies, photovoltaic inverters, and industrial power supplies. Attached Figure Description

[0050] Figure 1 A flowchart of a control method for implementing soft switching of a resonant converter across the entire load range is provided in this application;

[0051] Figure 2 Flowchart of the time series model training method provided in this application;

[0052] Figure 3 The flowchart of the online optimization decision-making method provided in this application;

[0053] Figure 4 The parameter identification flowchart provided for this application;

[0054] Figure 5 Flowchart of the dynamic closed-loop correction method provided in this application;

[0055] Figure 6 This application provides a control system structure diagram for implementing soft switching of a resonant converter across the entire load range. Detailed Implementation

[0056] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0059] Example 1

[0060] Reference Figures 1 to 5 This is the first embodiment of the present application, such as Figure 1 As shown, a control method for achieving soft switching of a resonant converter across the entire load range is provided.

[0061] Step 1: Collect and construct a raw dataset covering a wide temperature range and different aging stages, including switching current, resonant voltage, device temperature, and switching timing. The switching current is the resonant current. Perform continuous wavelet transform on the resonant current and voltage waveforms of each cycle in the raw dataset to generate a two-dimensional time-frequency energy spectrum, and insert it into the raw dataset to construct an enhanced dataset.

[0062] This step aims to construct a comprehensive and information-rich augmented dataset. The augmented dataset not only includes the original time-domain waveform data of the resonant converter under a wide operating range, but also extracts deep features that can finely characterize the dynamic changes of the system through signal processing techniques, providing a high-quality data foundation for subsequent training of high-precision sequence prediction models.

[0063] In a circuit simulation platform (such as PLECS or Simulink / Simscape), a high-fidelity resonant converter simulation model is built, which includes thermal models of key components and parameter drift characteristics. The core of this model lies in the detailed modeling of major semiconductor devices, including MOSFETs and diodes, as well as passive components, including resonant inductors and resonant capacitors.

[0064] For MOSFETs, the thermal model integrates the calculation of conduction loss and switching loss, and is based on the device's instantaneous junction temperature. Dynamically adjust its on-resistance Threshold voltage And switching time parameters. For resonant inductors and capacitors, the high-fidelity resonant converter simulation model includes their characteristic curves as a function of temperature, as well as simulations of aging effects caused by long-term operation, such as capacitance decay and increase in equivalent series resistance (ESR), and the decrease in permeability of the core material due to temperature rise, which in turn causes changes in inductance. Through this high-fidelity model, the electrical and thermodynamic behavior of the converter under different ambient temperatures, different load levels, and different health conditions can be systematically simulated.

[0065] Based on the constructed high-fidelity resonant converter simulation model, experimental data was collected using a physical prototype to calibrate and verify the simulation model. Temperature sensors and high-precision current and voltage probes were placed on the physical prototype, and tests were conducted at the same operating conditions as the simulation settings (covering a wide load range from 10% light load to 110% overload, different input voltages, and a wide temperature range from -40℃ to 85℃). Actual data were collected on resonant current, resonant voltage, surface temperature of key components (such as MOSFET case and core), and switching timing (frequency and dead time, etc.).

[0066] By comparing experimental and simulation data, and fine-tuning the thermal resistance network parameters, material temperature coefficients, and aging model coefficients in the simulation model, the root mean square error between the simulation output and experimental measurements is minimized, ensuring that the simulation model accurately reproduces the real physical process. After calibration, a large-scale scan simulation is performed using the validated simulation model to generate a raw dataset covering multiple switching cycles. This dataset contains operational data of the converter throughout its entire lifecycle and across all operating conditions. This combination of simulation and experimentation ensures both the breadth and efficiency of data acquisition, as well as the authenticity and accuracy of the data.

[0067] To deeply explore the non-stationary dynamic characteristics of the signal caused by factors such as parameter drift contained in the original dataset, this step analyzes the resonant current within each switching cycle. and resonant voltage The waveform undergoes continuous wavelet transform (CWT). CWT provides a joint time-frequency representation of the signal, and is particularly suitable for analyzing non-stationary, time-varying signals. The mathematical definition of CWT is as follows:

[0068] ;

[0069] in: In the scale factor Translation factor The wavelet coefficients calculated at the specified location have amplitudes that reflect the energy intensity of the signal in the corresponding time and frequency regions. This represents the original one-dimensional time-domain signal being analyzed, i.e., the resonant current within a single cycle. or resonant voltage ; It is a time integral variable; The mother wavelet function is an oscillating waveform with local support characteristics. In this step, the complex Morlet wavelet is selected, as it has a Gaussian window shape in both the time and frequency domains, providing good time-frequency resolution. Represents the complex conjugate of the mother wavelet function; It is a scaling factor, inversely proportional to the frequency of the signal, used to scale the mother wavelet to match signal components of different frequencies; It is a translation factor, representing the translation position of the mother wavelet on the time axis, used to locate a specific time point in the signal.

[0070] For each cycle and After the waveform is subjected to CWT, the squared modulus of the wavelet coefficients is calculated. This yields a two-dimensional matrix, which is the two-dimensional time-frequency energy spectrum of the signal.

[0071] In a two-dimensional time-frequency energy map, time is the horizontal axis and frequency (or scale) is the vertical axis. Color or grayscale value represents the energy density of the signal at that time-frequency point. For example, slight shifts in resonant frequency, the appearance or disappearance of high-frequency oscillations, and waveform distortions may be difficult to detect in the time domain waveform, but in the time-frequency energy map, they will appear as clear visual features such as the curvature of energy ridges, the appearance of energy clusters in specific time-frequency regions, or changes in intensity. This method of upscaling one-dimensional time-series signals to two-dimensional image features can greatly enrich the expressive power of information, allowing subtle dynamic changes caused by temperature variations and device aging to be quantified and highlighted.

[0072] The generated two-dimensional time-frequency energy maps are structurally integrated with the original dataset to construct the final enhanced dataset. For each sample in the original dataset (representing one or more consecutive switching cycles), its data structure is expanded into a tuple containing the following: {historical time-domain waveform sequence, corresponding two-dimensional time-frequency energy map sequence, current device temperature, current switching control timing}. In this way, each data point simultaneously possesses direct measurements (such as temperature and waveform) and in-depth derived features (time-frequency energy map).

[0073] This step constructs a multimodal and high-dimensional augmented dataset. This dataset not only ensures accurate mapping to physical reality through the fusion of simulation and experiment, but more importantly, by introducing time-frequency energy spectra generated by continuous wavelet transform, it makes explicit the non-stationary dynamic characteristics hidden in the time-domain waveforms and closely related to system parameter drift. This provides an extremely rich and robust set of input features for subsequent sequence prediction neural networks, enabling a deeper understanding of the complex nonlinear relationship between system state and control input. This lays a solid data foundation for achieving accurate soft-switching prediction and control across the entire load range.

[0074] Step 2: Train a sequence prediction model based on the augmented dataset. The sequence prediction model adopts a hybrid time-series neural network structure, using historical time-domain waveforms, corresponding two-dimensional time-frequency energy spectra, and current device temperature data as inputs to predict the resonant dynamics under candidate control parameters for the next cycle, and calculates the soft-switching confidence based on the predicted resonant dynamics.

[0075] This step, based on the augmented dataset constructed in step 1, trains a sequence prediction model. The sequence prediction model aims to accurately predict the dynamic response of the resonant converter under given candidate control parameters for the next cycle, and quantitatively evaluates the confidence level of achieving soft switching based on the prediction results, providing a decision-making basis for subsequent online optimization control. See also... Figure 2 The following is a flowchart of the time series model training method provided in this step.

[0076] A hybrid temporal neural network is constructed as a sequence prediction model. The structure of the hybrid temporal neural network integrates a convolutional neural network (CNN) for processing two-dimensional time-frequency energy maps and a long short-term memory network (LSTM) for processing time-series data. The hybrid network architecture can process and effectively fuse multimodal input data in parallel.

[0077] The input to the hybrid temporal neural network consists of three parts, specifically including: the resonant current from N historical switching cycles. With resonant voltage The time-domain waveform sequence; N two-dimensional time-frequency energy spectrum sequences corresponding to the historical waveform sequence; the device temperature measured in the current control cycle. and candidate control parameters used for prediction (For frequency-controlled LLC converters, this parameter is the candidate switching frequency) For phase-shift controlled PSFB converters, the candidate phase shift angle is... ).

[0078] In the sequence prediction model, two-dimensional time-frequency energy maps (as image data) spanning N historical periods are fed into a dedicated CNN branch. This CNN branch consists of stacked convolutional layers, activation function layers (such as ReLU), and pooling layers, and its function is to automatically extract deep spatial features from each map that characterize the local time-frequency structure and non-stationary features of the signal.

[0079] After CNN processing, each two-dimensional time-frequency energy map is encoded into a low-dimensional feature vector. Thus, N maps form a feature vector sequence. Simultaneously, the original time-domain waveform sequence from N historical periods is fed into the LSTM branch to capture the time dependencies in the waveform data. The sequence prediction model also includes the current device temperature as input. and candidate control parameters These two scalar data points, along with the feature vector sequence extracted by the CNN and the time-domain waveform sequence of the resonant current and resonant voltage, are concatenated at the feature level to form a composite feature sequence that integrates time-domain information, frequency-domain information, thermal state information, and control intent.

[0080] The composite feature sequence is then fed into the main LSTM network, which is responsible for learning the complex dynamics of the resonant converter's operating state over time. The final hidden state output of the main LSTM network is decoded through one or more fully connected layers to ultimately predict the candidate control parameters. Under the influence of the action, the complete resonant current waveform of the next switching cycle and resonant voltage waveform .

[0081] To obtain the predicted waveform of the next cycle and Next, the soft-switching confidence level needs to be calculated. This confidence level is a continuous value between 0 and 1, used to quantify the degree to which soft switching is achieved. Taking an LLC converter as an example, the soft-switching condition for its main power switches is zero-voltage turn-on (ZVS).

[0082] ZVS implementation depends on the moment of activation (denoted as...). When the voltage across the switching transistor is zero and the current flowing through it is negative (i.e., the current freewheels through its body diode), the confidence level of ZVS can be based on the predicted resonant current. At the time of opening The value is used to define the confidence level; to obtain a smooth and differentiable confidence function, the Sigmoid function is used for modeling:

[0083] ;

[0084] in, It is the confidence level of ZVS achieved by a single switching transistor; The neural network predicts the resonant current value at the moment the switch turns on in the next cycle; It is a positive constant gain factor used to adjust the steepness of the function; when When it is negative and the absolute value is large, Approaching 1; when When it is the right time, Approaching 0.

[0085] For a resonant converter with a half-bridge or full-bridge structure, there are multiple switching events within a switching cycle. The total soft-switching confidence is a weighted combination of the confidence of all key switching events (such as the turn-on of Q1 and Q2) within the cycle, for example, by taking their geometric mean or arithmetic mean, to comprehensively evaluate the soft-switching performance of the entire cycle.

[0086] The training process of the sequence prediction model adopts supervised learning, and the training loss function is... It consists of two parts: waveform prediction loss and confidence-assisted loss The waveform prediction loss is calculated using the mean square error (MSE) between the predicted waveform and the actual waveform in the dataset, ensuring that the sequence prediction model can accurately reproduce the resonant dynamics. The confidence-assisted loss guides the model to focus more directly on the waveform features related to soft switching. In this way, the network structure of the sequence prediction model can not only learn the overall shape of the waveform, but also focus on learning the key feature points that determine the success or failure of soft switching.

[0087] This step constructs and trains a sequence prediction model, which, as a high-precision "digital twin," can proactively deduce the converter's electrical response based on historical operating states, current ambient temperature, and a hypothetical future control action. This step also introduces the quantitative metric of soft-switching confidence, transforming a complex physical phenomenon (soft switching) into a definite and optimizable mathematical objective. This enables the controller to predict the consequences of different control decisions and perform quantitative comparisons.

[0088] Step 3: In each control cycle, different candidate switching frequencies or phase shift adjustments are used as variables and input into the sequence prediction process to obtain the resonant dynamic prediction and soft-switching confidence corresponding to each candidate adjustment. The constrained optimization problem with the goal of maximizing the soft-switching confidence and operating efficiency is solved. Under the constraint of the safe operating area, the optimal switching frequency or phase shift adjustment is output in real time.

[0089] This step aims to determine and output the control parameters that enable the resonant converter to achieve optimal overall performance in real time within each control cycle, using the sequence prediction model trained in step 2 to solve a constrained optimization problem. This process translates predictive capability into actual control actions, ensuring that the converter can continuously operate in a highly efficient and soft-switching state under dynamically changing operating conditions and its own parameter drift. See also... Figure 3 The flowchart below shows the online optimization decision-making method provided for this step.

[0090] Specifically, at the beginning of each control cycle, a set of candidate adjustment values ​​for the control parameters of the next cycle is defined. Taking switching frequency modulation as an example, the switching frequency control parameter is the switching frequency. Based on the current switching frequency Generate a finite set of discrete candidate frequencies Discrete candidate frequency set This covers a reasonable search range near the current operating point; for example, the switching frequency. ,in, The switching frequency index variable in the candidate frequency set. M is the number of switching frequencies in the candidate frequency set. By adjusting the step size to maximize frequency, this discretization method transforms the complex continuous optimization problem into a discrete search problem that is computationally easier to process in real time in embedded controllers.

[0091] For the candidate adjustment set Each candidate control parameter in (here) Right now This, along with historical state data (including historical waveforms, time-frequency spectra, and current device temperature), is input into the sequence prediction model constructed in step 2. The sequence prediction model will then output the candidate control parameter. The next cycle of the predicted resonant dynamics includes the predicted resonant current waveform. and resonant voltage waveform And the soft-switching confidence calculated based on these predicted waveforms. .

[0092] To achieve optimal overall performance, this step constructs a multi-objective optimization problem. In this constructed multi-objective optimization problem, besides maximizing the soft-switching confidence... In addition, work efficiency was also introduced. As another optimization objective; predicted work efficiency Estimation can be made based on the predicted waveform; for example, by calculating the predicted input power. and known or expected output power The predicted input power can be obtained from the predicted input voltage. and input current In a switching cycle The result is obtained through internal integration, while the input current and resonant current are... Directly related; in the LLC resonant full-bridge topology used, the resonant current Since the output currents of the bridge arms are in the same series branch, the DC-side input current can be considered as the equivalent component of the resonant current drawn from the bus in each switch-on state. Therefore, the input current (and the resulting input power) can be directly derived from the predicted resonant current. It was calculated using the known bridge arm voltage waveform.

[0093] In this way, for each candidate control parameter All of them are associated with a soft-switching confidence level. and prediction efficiency .

[0094] The two optimization objectives are combined into a unified performance index function through a weighted summation. : ;in, Candidate control parameters The corresponding comprehensive performance evaluation value; It is the weighting coefficient of the soft-switching confidence, reflecting the importance attached to the implementation of soft switching; It is a weighting coefficient for work efficiency, reflecting the degree of importance attached to energy efficiency; and All are positive numbers and satisfy the following conditions: Their specific values ​​can be preset according to the application scenario and design requirements of the converter, allowing for a flexible trade-off between pursuing ultimate soft switching and maximum efficiency.

[0095] During the optimization process, it is essential to ensure that all candidate solutions remain within the converter's safe operating area (SOA). Therefore, the multi-objective optimization problem is constrained, primarily based on the predicted waveforms, to prevent device overcurrent or overvoltage. Specific constraints are as follows:

[0096] ;

[0097] in, and It's about candidate control parameters. The constraint function; It is the predicted peak value of the resonant current; It is the predicted peak value of the resonant voltage; and These are the preset maximum allowable current and maximum allowable voltage of the resonant network, which are determined by the specifications of the power devices and passive components.

[0098] The optimization process of a multi-objective problem is to optimize the set of candidate adjustment variables. In the process, iterate through all cases that satisfy the above security constraints. and Candidate control parameters And find one that can make the comprehensive performance index function Maximize the optimal control parameters :

[0099] ;

[0100] in, Indicating in the candidate adjustment set In the middle, find the comprehensive performance index function. The independent variable that takes the maximum value The optimal control parameters obtained by solving (i.e., optimal switching frequency) Or optimal phase shift angle This will be used as the actual control command for the next control cycle, and will be output to the drive circuit by the digital signal processor (DSP) or field programmable gate array (FPGA), thereby completing a closed-loop online optimization control.

[0101] The technical solution developed in this step no longer passively responds to errors, but possesses proactive decision-making capabilities. Within each control cycle, it predicts the consequences of different control strategies and, while ensuring safety, actively selects the control action that maximizes soft-switching performance and efficiency. Based on model prediction and multi-objective optimization methods, this step enables the converter to adapt in real-time to changes in load, input voltage, and parameter drift caused by temperature and aging. This allows it to continuously maintain its position near the optimal operating point across the entire operating range, significantly improving the overall performance, reliability, and lifespan of the resonant converter.

[0102] Step 4: During each soft-switching cycle, obtain the zero-crossing and peak point information of the resonant current and voltage from the original dataset, and use the Kalman filter algorithm to calculate the current values ​​of the equivalent resonant inductance and capacitance in real time for parameter identification.

[0103] The core task of this step is to accurately identify the equivalent resonant inductance in each switching cycle during the operation of the resonant converter. and equivalent resonant capacitance The current value; this process analyzes the key electrical signals of the resonant circuit acquired in real time and applies the Kalman filter algorithm to dynamically track parameter drift caused by factors such as temperature changes and device aging. See also Figure 4 The flowchart provided for parameter identification in this step is as follows.

[0104] Specifically, at the end of each complete soft-switching cycle, the resonant current within that cycle is acquired. and resonant capacitor voltage The waveform data is used to accurately extract key feature points from these real-time waveform data using digital signal processing algorithms. These feature points include the zero-crossing time of the resonant current, the peak time and amplitude of the resonant current, and the peak time and amplitude of the resonant capacitor voltage. These feature points provide necessary observation information for subsequent parameter identification.

[0105] Digital signal processing algorithms are derived from the acquired resonant current. and resonant capacitor voltage In waveform data, key feature points are quickly and accurately located. For resonant current waveforms, digital signal processing (DSP) algorithms identify zero-crossing moments by detecting changes in the sign of adjacent sampling points, typically using linear interpolation to achieve sub-sampling period accuracy. Simultaneously, DSP algorithms search for local extrema in the waveform, determining the peak current moment and its corresponding amplitude by the position of the sign change of the first derivative (difference). Similarly, DSP algorithms are applied to resonant capacitor voltage waveforms to identify the peak voltage moment and amplitude. These extracted timestamps and amplitudes constitute a set of precisely quantized feature vectors describing the dynamic behavior of the switching cycle.

[0106] To apply Kalman filtering for parameter estimation, a state-space model describing the dynamics of the resonant circuit is first established. In this step, the parameters to be identified are used as state variables. To simplify the model and improve numerical stability, the reciprocals of the parameters are chosen as the state vectors. ,Right now ,in and Representing respectively in the The resonant inductor during one cycle and resonant capacitor The reciprocal, This represents the transpose of a vector.

[0107] Since these parameters change very slowly over a single switching cycle, they can be dynamically modeled as a random walk process, with the following state equation: ;in, It is in the State vector estimate for each cycle; It is the estimated state vector value from the previous period; It is a two-dimensional identity matrix; This is the process noise vector, assumed to be zero-mean Gaussian white noise, with the following covariance matrix: ; The existence of this indicates that the values ​​of inductance and capacitance can undergo small and unpredictable changes due to aging and temperature effects.

[0108] An observation equation is constructed to correlate state variables with measurable physical quantities. Based on the fundamental electromagnetic relationships of resonant circuits, a linear observation model can be established using waveform integration, which offers better robustness to measurement noise compared to differentiation.

[0109] In a time interval Internal, resonant capacitor voltage , The change in quantity and the current flowing through it The integral satisfies the following relation: Similarly, the resonant inductor current , The change in voltage and the integral of the voltage across the terminals satisfy the following relationship: Inductor voltage It can be determined by the input square wave voltage and capacitor voltage The calculation shows that, .

[0110] Therefore, it can be defined in the first... Observation vectors for each period and observation matrix For example, half a resonance period can be selected as the observation time window. Then the observation equation can be written as: ;in, It is the observation vector, whose elements are directly calculated from the acquired waveform data;

[0111] It is the observation matrix, whose elements are obtained by numerical integration of the acquired waveforms; This is the observed noise vector, assumed to be zero-mean Gaussian white noise, with a covariance matrix of... , The magnitude of the value reflects the measurement accuracy of the current and voltage sensors, as well as the quantization error of the analog-to-digital conversion.

[0112] At the end of each cycle, an iterative calculation of the Kalman filter is performed, which includes two phases: prediction and update. In the prediction phase, the state vector and covariance for the current control cycle are predicted based on the results of the previous control cycle. In the update phase, the observations from the current control cycle are used. and To correct the predicted values, calculate the Kalman gain. And finally obtain the optimal posterior estimate of the current control cycle state vector. Finally, by taking the reciprocal, the identification values ​​of the equivalent resonant inductance and capacitance within the current control cycle can be obtained: and ;in, Indicates the first The inductance parameter components in the state vector during each switching cycle; Indicates the first The capacitance parameter component in the state vector during each switching cycle.

[0113] The technical solution in this step enables the resonant converter to acquire self-sensing capability. The resonant converter no longer relies on the nominal parameter values ​​at the initial design stage, but can quantify the actual electrical characteristics of the key passive components of the resonant circuit in real time. This accurate online parameter identification provides crucial, dynamically updated system model information for subsequent control decisions, enabling the control algorithm to truly adapt to the physical changes of the converter itself and achieve robust and efficient control throughout its entire life cycle.

[0114] Step 5: The identified equivalent resonant inductance and capacitance are used as new dynamic features and input into the sequence prediction process in the next control cycle. This is to adjust the predicted resonant dynamics and soft-switching confidence based on parameter drift, adjust the optimization problem solving of subsequent control cycles, dynamically correct the control output, and continuously track the actual optimal soft-switching operating point.

[0115] The purpose of this step is to construct a closed-loop adaptive correction mechanism that combines real-time parameter identification results with feedforward predictive control. By using the resonant element parameters identified online in step 4 as new dynamic features and feeding them back into the sequence prediction process in step 2, real-time calibration of the control model's prediction results is achieved, and the final control output is dynamically corrected. This ensures that the resonant converter can continuously and accurately track the optimal soft-switching operating point that dynamically changes due to parameter drift. See also... Figure 5 The flowchart below shows the dynamic closed-loop correction method provided in this step.

[0116] Specifically, in the first At the end of each control cycle, the parameter identification process in step 4 has output the equivalent resonant inductance identification value for the current cycle. Identification value of equivalent resonant capacitance These two scalar values ​​are combined into a parameter identification vector. In order to control the next cycle (i.e., the first cycle) In the prediction of (number of cycles), this latest information is used to integrate this vector as a new dynamic feature with the original input data.

[0117] The inputs to the original sequence prediction model mainly include historical time-domain waveforms, two-dimensional time-frequency energy maps, and device temperature. In the new control cycle... By incorporating these elements, a more information-rich input tensor is formed, which is used to process the first... The resonant dynamics of the nth cycle are predicted. Specifically, it is used to predict the nth cycle. Periodic dynamic input vector Defined as: ;in, It is used for the first The complete input vector of a sequence prediction model for each control period; Representing up to the A sequence of historical time-domain waveform data from the end of a cycle, containing data from multiple past cycles; Is with The corresponding two-dimensional time-frequency energy spectrum sequence; It is in the The current device temperature data collected in each cycle; It is in the The parameter vector consisting of the equivalent resonant inductance and capacitance values ​​identified in each cycle; This means concatenating various feature vectors or tensors to form a unified input with a higher dimension.

[0118] By real-time identified parameters As a direct input, the sequence prediction model (hybrid temporal neural network) can anchor its prediction process to the actual physical foundation of the converter. The network structure of the sequence prediction model has already learned the resonant parameters during the training phase. and The changes in these parameters are intrinsically linked to resonant dynamics (such as resonant frequency, current-voltage phase relationship, etc.). Therefore, when the input contains precise current parameter values, the sequence prediction model can predict given candidate control parameters (such as switching frequency). The resonant dynamics under these conditions will be more accurate, thus allowing for the calculation of the soft-switching confidence level. This also makes it closer to reality. As a result, the optimization problem constructed in step 3, which aims to maximize the confidence of soft switching, is solved based on a prediction model that has been calibrated in real time, which can significantly improve the effectiveness and reliability of the optimization solution.

[0119] For example, suppose in the first The ... Each cycle evaluates the switching frequency adjustment of two candidates. and Without parameter feedback, sequence prediction models may predict... The corresponding soft-switching confidence level is 0.95, while The value is 0.92, therefore it will be selected. However, the identification results in step 4 show that, due to the increase in temperature, the equivalent resonant inductance identification value... It has increased by 3% compared to the nominal value. When this new... The value is used as After a portion of the data is input into the sequence prediction model, the model re-evaluates the two candidate adjustment parameters based on its internally learned knowledge. The corrected prediction result may become: The confidence level dropped to 0.88 (because the increased inductance caused a shift in the original optimal frequency point), while The confidence level then increases to 0.96. At this point, the controller will dynamically adjust its decision and instead select... This serves as the optimal control output for the next cycle. This process repeats itself within each control cycle, achieving continuous tracking of the optimal operating point.

[0120] This step transforms the sequence prediction model trained on historical data into a self-correcting dynamic model. By establishing a rapid feedback loop from physical state perception (parameter identification) to model prediction correction and control decision optimization, the control strategy can respond in real time to internal parameter drift caused by factors such as temperature and aging. The dynamic correction mechanism constructed in this step ensures that the control output is always the optimal solution for the current true state of the converter. This allows the converter to maintain the highest soft-switching performance and efficiency throughout its entire lifecycle and wide operating range, greatly enhancing the robustness and long-term reliability of the resonant converter.

[0121] Example 2

[0122] Reference Figure 6 This is the second embodiment of the present application, which provides a control system for realizing soft switching of a resonant converter across the entire load range.

[0123] The system includes: a data feature construction module, a time series model training module, an online optimization decision module, a parameter identification module, and a dynamic closed-loop correction module.

[0124] The data feature construction module collects the original datasets of resonant current, resonant voltage, device temperature, and switching timing. It performs time-frequency transformation on the resonant current and voltage waveforms in the original datasets to generate a two-dimensional time-frequency energy spectrum, and constructs an enhanced dataset with the original datasets.

[0125] The timing model training module trains a sequence prediction model based on an augmented dataset. The sequence prediction model takes historical time-domain waveforms, corresponding time-frequency features, and current device temperature data as inputs to predict the resonant dynamics under given candidate control parameters, and calculates the soft-switching confidence based on the predicted resonant dynamics.

[0126] The online optimization decision module inputs different candidate control parameters into the sequence prediction model in each control cycle to obtain the soft-switching confidence level corresponding to each candidate control parameter, solves the optimization problem with the goal of maximizing the soft-switching confidence level, and outputs the optimal control parameters under the constraint of the safe operating area.

[0127] The parameter identification module identifies the current values ​​of the equivalent resonant inductance and capacitance in real time based on the resonant current and voltage waveforms during each soft-switching cycle.

[0128] The dynamic closed-loop correction module uses the identified equivalent resonant inductance and capacitance as new dynamic features, and inputs them into the sequence prediction model in the next control cycle to adjust the prediction results according to parameter drift, dynamically correct the control output, and continuously track the optimal soft-switching operation.

[0129] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0130] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

Claims

1. A control method for achieving soft switching across the entire load range of a resonant converter, characterized in that, include: The original datasets of resonant current, resonant voltage, device temperature and switching timing are collected. The resonant current and voltage waveforms in the original datasets are processed by time-frequency transformation to generate a two-dimensional time-frequency energy spectrum, and an enhanced dataset is constructed with the original dataset. The sequence prediction model is trained based on the augmented dataset. The sequence prediction model takes historical time-domain waveforms, corresponding time-frequency features and current device temperature data as inputs to predict the resonant dynamics under given candidate control parameters, and calculates the soft-switching confidence based on the predicted resonant dynamics. In each control cycle, different candidate control parameters are input into the sequence prediction model to obtain the soft-switching confidence corresponding to each candidate control parameter. An optimization problem with the goal of maximizing the soft-switching confidence is solved, and the optimal control parameters are output under the constraint of the safe operating area. During each soft-switching cycle, the current values ​​of the equivalent resonant inductance and capacitance are identified in real time based on the resonant current and voltage waveforms. The identified equivalent resonant inductance and capacitance are added as new dynamic features and input into the sequence prediction model in the next control cycle. This allows the prediction results to be adjusted according to parameter drift, the control output to be dynamically corrected, and the optimal soft-switching operating point to be continuously tracked.

2. The control method for achieving soft switching of a resonant converter across the entire load range according to claim 1, characterized in that, The resonant current, resonant voltage, device temperature, and switching timing of the resonant converter are collected, including the frequency and dead time. Continuous wavelet transform is performed on the resonant current and resonant voltage waveforms acquired in each switching cycle. After wavelet transform, the squared modulus of the wavelet coefficients is calculated to obtain a two-dimensional matrix, which is the two-dimensional time-frequency energy spectrum of the current and voltage signals. In the two-dimensional time-frequency energy spectrum, time is the horizontal axis and frequency is the vertical axis, and the gray value represents the energy density of the signal at the time and frequency points. The generated two-dimensional time-frequency energy map is structurally integrated with the original dataset to construct an enhanced dataset.

3. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 2, characterized in that, A hybrid temporal neural network was constructed as a sequence prediction model. The hybrid temporal neural network structure integrates convolutional neural networks and long short-term memory networks. The convolutional neural network is used to process two-dimensional time-frequency energy maps, and the long short-term memory network is used to process time-series data. The inputs to the hybrid time-series neural network include: time-domain waveform sequences of resonant current and resonant voltage during historical switching cycles, a two-dimensional time-frequency energy spectrum sequence corresponding to the historical time-domain waveform sequences, the device temperature measured during the current control cycle, and candidate control parameters for prediction.

4. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 3, characterized in that, In the sequence prediction model, the two-dimensional time-frequency energy map of the historical period is fed into the convolutional neural network branch as image data. The convolutional neural network is composed of multiple convolutional layers, activation function layers and pooling layers stacked together. Deep spatial features that can characterize the local time-frequency structure and non-stationary features of the signal are extracted from the two-dimensional time-frequency energy map. After being processed by a convolutional neural network, each two-dimensional time-frequency energy spectrum is encoded into a low-dimensional feature vector, forming a sequence of feature vectors. The input to the sequence prediction model also includes scalar data of the current device temperature and candidate control parameters. The device temperature and candidate control parameters are concatenated with the feature vector sequence extracted by the convolutional neural network and the time-domain waveform sequence of the resonant current and resonant voltage at the feature level to form a composite feature sequence. The composite feature sequence is fed into the long short-term memory network, which is responsible for learning the complex dynamic law of the working state of the entire resonant converter evolving over time. After being decoded by the fully connected layer, the final hidden state output of the Long Short-Term Memory network predicts the complete resonant current waveform and resonant voltage waveform of the next switching cycle under the action of candidate control parameters. After obtaining the predicted resonant current waveform and resonant voltage waveform for the next cycle, the soft-switching confidence is calculated. The soft-switching confidence is used to quantify the extent to which soft-switching is achieved.

5. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 4, characterized in that, Construct a multi-objective optimization problem with two optimization objectives: maximizing soft-switching confidence and maximizing operating efficiency. Operating efficiency is obtained by calculating the ratio of predicted input power to output power. The predicted input power is obtained by integrating the predicted input voltage and input current over one soft-switching cycle, and the input current is directly related to the resonant current. The two optimization objectives are merged into a unified performance index function by weighted summation. During the optimization process, it is ensured that all candidate solutions are within the safe operating range of the resonant converter, and the corresponding constraints are set based on the predicted waveform. The optimization process of a multi-objective problem is to traverse all candidate control parameters that satisfy the safety constraints in the candidate adjustment set and find the optimal control parameter that maximizes the comprehensive performance index function. The optimal control parameters obtained will be used as the actual control command for the next control cycle and output to the drive circuit.

6. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 5, characterized in that, At the end of each soft-switching cycle, the waveform data of the resonant current and resonant capacitor voltage within that cycle are acquired, and feature points are accurately extracted using digital signal processing algorithms, including the zero-crossing time of the resonant current, the peak time and amplitude of the resonant current, and the peak time and amplitude of the resonant capacitor voltage. The specific steps of the digital signal processing algorithm include: For the resonant current waveform, the digital signal processing algorithm identifies the zero-crossing time by detecting the change in the sign of adjacent sampling points, and at the same time finds the local extreme points in the waveform, that is, it determines the peak time and corresponding amplitude of the resonant current by the position of the sign change of the first derivative; similarly, the digital signal processing algorithm processes the resonant capacitor voltage waveform in the same way to find the peak time and amplitude of the resonant capacitor voltage. The extracted timestamps and amplitudes constitute a feature vector describing the dynamic behavior of the switching cycle.

7. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 6, characterized in that, The Kalman filter algorithm is applied to estimate parameters, and a state-space model describing the dynamics of the resonant circuit is established. The parameters to be identified are used as state variables, and the reciprocals of the parameters are selected as state vectors. The observation equations of the Kalman filter algorithm are constructed, which associate state variables with measurable physical quantities. Based on the electromagnetic relationship of the resonant circuit, a linear observation model is established using waveform integration.

8. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 7, characterized in that, At the end of each control cycle, an iterative calculation of the Kalman filter is performed, which includes two stages: prediction and update. In the prediction stage, the state vector and covariance of the current control cycle are predicted based on the results of the previous control cycle. During the update phase, the predicted values ​​are corrected using the observations of the current control cycle, the Kalman gain is calculated, and the optimal posterior estimate of the state vector of the current control cycle is obtained. By taking the reciprocal, the identification values ​​of the equivalent resonant inductance and capacitance in the current control cycle can be obtained.

9. The control method for realizing soft switching of a resonant converter across the entire load range according to claim 8, characterized in that, At the end of each control cycle, the equivalent resonant inductance and capacitance identified online are used as new dynamic features; these dynamic features are integrated with the input data of historical time-domain waveforms and two-dimensional time-frequency energy maps to form an enhanced input tensor for prediction of the next control cycle. The enhanced input tensor is fed into the sequence prediction model to obtain real-time calibrated prediction results; The control output decision is dynamically corrected based on the calibrated prediction results, and the optimal soft-switching operating point is tracked in a dynamic manner.

10. A control system for implementing soft switching of a resonant converter across its entire load range, used to implement the control method for implementing soft switching of a resonant converter across its entire load range as described in any one of claims 1 to 9, characterized in that, include: The system includes a data feature construction module, a time series model training module, an online optimization decision-making module, a parameter identification module, and a dynamic closed-loop correction module. The data feature construction module is used to collect the original datasets of resonant current, resonant voltage, device temperature and switching timing, perform time-frequency transformation on the resonant current and voltage waveforms in the original dataset to generate a two-dimensional time-frequency energy spectrum, and construct an enhanced dataset with the original dataset. The timing model training module trains a sequence prediction model based on an augmented dataset. The sequence prediction model takes historical time-domain waveforms, corresponding time-frequency features, and current device temperature data as inputs to predict the resonant dynamics under given candidate control parameters, and calculates the soft-switching confidence based on the predicted resonant dynamics. The online optimization decision module inputs different candidate control parameters into the sequence prediction model in each control cycle to obtain the soft-switching confidence corresponding to each candidate control parameter, solves the optimization problem with the goal of maximizing the soft-switching confidence, and outputs the optimal control parameters under the constraint of the safe operating area. The parameter identification module identifies the current values ​​of the equivalent resonant inductance and capacitance in real time based on the resonant current and voltage waveforms during each soft-switching cycle. The dynamic closed-loop correction module uses the identified equivalent resonant inductance and capacitance as new dynamic features, and inputs them into the sequence prediction model in the next control cycle to adjust the prediction results according to parameter drift, dynamically correct the control output, and continuously track the optimal soft-switching operating point.

Images