A non-singular terminal sliding mode control parameter collaborative optimization method for a buck converter

Abstract

The application provides a non-singular terminal sliding mode control parameter collaborative optimization method for a Buck converter, comprising the following steps: step 1, constructing a state space average model of the Buck converter with inductance current and capacitor voltage as state variables, and defining system tracking error state variables; step 2, filtering and preprocessing transient simulation output voltage of the Buck converter; step 3, constructing a dynamic multi-objective composite fitness function fusing a load step sensing penalty term; step 4, calculating information entropy of each index by using an entropy weight method; step 5, performing physical mechanism guided Bayesian optimization by using a domain specific kernel function; and step 6, performing closed loop iteration optimization of a particle swarm algorithm by using the dynamic multi-objective composite fitness function after parameter calibration. The application effectively decouples the contradiction between transient response and steady state accuracy, guarantees physical safety from the bottom of the algorithm, and significantly improves the control robustness of the converter under large disturbance.

Description

Technical Field

[0001] This invention belongs to the field of power electronics and intelligent control, and particularly relates to a method for collaborative optimization of non-singular terminal sliding mode control parameters for buck converters. Background Technology

[0002] With the rapid development of modern power electronics technology, buck converters ( As a core component of DC-DC power conversion, buck converters have been widely used in renewable energy grid connection, communication base stations, high-precision microelectronic power supply, and industrial control due to their advantages such as simple topology, high conversion efficiency, and small size. However, buck converters are inherently nonlinear, time-varying systems with discrete switching characteristics. In actual operation, they often face large signal disturbances such as startup transients, drastic input voltage fluctuations, and load abrupt changes. These pose extremely high challenges to the transient response speed, steady-state accuracy, and robust stability of their control systems.

[0003] Traditional buck converter control often employs a fixed linear control strategy (such as...). This type of control performs reasonably well under steady-state and small-signal disturbances, but it struggles to cope with changes in system dynamics when faced with large-scale operating condition switching or large-signal disturbances, often exhibiting drawbacks such as excessively long settling times or even system instability. To overcome the limitations of traditional linear control, non-singular terminal sliding mode control (NMT) is used. Nonlinear control theories such as ) were introduced. It not only possesses strong robustness to parameter perturbations and external disturbances, but also guarantees that the system state converges without singularities within a finite time. However, A significant bottleneck exists in the practical engineering applications of this technology: its dynamic performance is extremely sensitive to controller parameters (such as sliding mode coefficient, nonlinear exponent, hysteresis bandwidth, etc.), and there are complex and strong coupling relationships among these parameters, which constitute a very challenging non-convex multi-objective trade-off problem.

[0004] In recent years, with the development of intelligent control technology and optimization algorithms, more and more research has focused on using heuristic algorithms (such as particle swarm optimization). Differential Evolution (etc.) The controller parameters are tuned end-to-end. In this type of multi-objective parameter optimization, the core challenge is how to balance the system's rise time, overshoot, and steady-state error. Existing intelligent optimization methods generally adopt a static weight allocation strategy when constructing the fitness function, that is, assigning a globally fixed weight coefficient to each performance indicator. However, the performance requirements of a physical system are completely different at different stages of evolution—in the early stage, it needs to quickly approach the target (emphasizing response speed), in the middle stage, it needs to suppress oscillations (emphasizing control overshoot), and in the late stage, it needs to accurately lock on (emphasizing steady-state error). The static weight mechanism that runs all over the world is simply unable to adapt to the stage-specific physical requirements of the system's evolution.

[0005] As a crucial component of power electronics technology, the transient response and steady-state disturbance rejection capability of buck converters under complex operating conditions are of paramount importance to power supply quality. Although existing technologies have achieved certain results in parameter optimization using traditional particle swarm optimization (PSO) or differential evolution (DE) algorithms, the difficulty in accurately coordinating the static weights of multiple objectives easily leads to mutual constraints and conflicts among performance indicators, causing the algorithms to easily get trapped in local optima.

[0006] Furthermore, even when some advanced control strategies attempt to introduce dynamic weight allocation to break static constraints, the shape hyperparameters of their dynamic weight functions (such as decay rate and Gaussian width) still face the challenge of scientifically calibrating them. Traditional hyperparameter optimization methods (such as grid search or standard Bayesian optimization) are mostly pure mathematical black-box optimizations, completely detached from the underlying physical laws and dynamic control theory of power electronic converters. Especially in non-singular terminal sliding mode control, the process of system state convergence to the sliding surface is accompanied by strict dynamic energy decay, while the Gaussian process surrogate model of standard Bayesian optimization (e.g., using conventional radial basis kernel functions) cannot accurately map this physically guided convergence trend. This not only leads to limited accuracy in predicting control performance costs and low global optimization efficiency of the surrogate model, but also easily causes the generated hyperparameter combinations to deviate from the stability domain of the physical system, triggering control instability. Therefore, how to deeply integrate the underlying physical stability mechanism of the converter with Bayesian optimization, an advanced machine learning algorithm, and propose a hyperparameter optimization framework guided by a physical mechanism is the key to breaking through the bottleneck of existing intelligent control parameter tuning technology. Summary of the Invention

[0007] Purpose of the invention: In order to solve the problems of difficulty in coordinating multi-objective static weights, easy getting trapped in local optima, and high dependence on human experience in the optimization of control parameters of existing NTSMC, this invention proposes a collaborative tuning method for non-singular terminal sliding mode control parameters for buck converters. The aim is to achieve optimal end-to-end controller parameter configuration without subjective bias, and significantly improve the dynamic tracking performance and robust stability of buck converters.

[0008] This method includes the following steps:

[0009] Step 1: Construct a state-space average model of the buck converter with inductor current and capacitor voltage as state variables, and define the system tracking error state variables.

[0010] Step 2: Perform filtering preprocessing on the transient simulation output voltage of the buck converter to eliminate high-frequency switching ripple interference. Then, extract heterogeneous physical performance indicators, including the settling time with millisecond-level physical dimensions. Overshoot in percentage units and steady-state error ;

[0011] Step 3: Divide the algorithm optimization cycle into different stages according to physical laws, configure nonlinear dynamic weight functions with topology reshaping capabilities for each performance index, and construct a dynamic multi-objective composite fitness function that integrates load step perception penalty terms.

[0012] Step 4: Calculate the information entropy of each indicator using the Entropy Weight Method (EWM), and objectively calibrate the base amplitudes of the dynamic weighting functions corresponding to the steady-state time in a purely data-driven manner. The base amplitude of overshoot and the fundamental magnitude of steady-state error

[0013] Step 5: To overcome the limitations of traditional pure mathematical black-box optimization, the multi-objective composite fitness value is used as the observation input of the Gaussian process surrogate model. A domain-specific kernel function is constructed based on the non-singular terminal sliding mode Lyapunov energy function and derivative decay rate correction of the buck converter. The domain-specific kernel function is used to perform physical mechanism-guided Bayesian optimization (BO) to globally search for the optimal set of shape hyperparameters of the dynamic weight function. Where a is the exponential decay rate, is the characteristic length scale of the Gaussian function, and b is the growth exponent of the convex growth model;

[0014] Step 6: Utilize the parameter-calibrated dynamic multi-objective composite fitness function to perform closed-loop iterative optimization using the particle swarm optimization algorithm. In each iteration, substitute the generated candidate parameter set into the steady-state switching frequency calculation equation based on the power MOSFET evaluation model for physical constraint verification. Forcefully eliminate solutions that lead to physically unrealizable solutions at the limiting frequency, and finally derive the optimal parameters that satisfy the hardware constraints. And configured in the digital controller, where For the optimal synovial coefficient, It is the optimal nonlinear exponent. This represents the optimal hysteresis bandwidth.

[0015] In step 1, to achieve finite-time convergence and avoid singularities, a non-singular terminal sliding mode is constructed. Define the set of strongly coupled control parameters to be optimized. ,in The sliding mode coefficient is , It is a non-linear exponent. Hysteresis bandwidth is the limit of the switching frequency of a power metal-oxide-semiconductor field-effect transistor (MOSFET).

[0016] In step 2, to eliminate the high-frequency switching frequency during the hard switching process of the buck converter... To address high-order harmonic interference, a sliding window integral low-pass filter is used for physical preprocessing. The mapping formula is as follows:

[0017] ,

[0018] in, The transient output voltage is the filtered voltage, and t is the current time variable. For integration variables, For the switching cycle, This is the transient output voltage. The differential symbol is used; the processed data is then input into the minimax normalization mapping.

[0019] Step 3, which involves configuring nonlinear dynamic weighting functions with topology reshaping capabilities for each performance index, specifically includes:

[0020] Exploration Phase: Constructing dynamic weights for the exponential decay model with respect to the steady-state time:

[0021] ,

[0022] in, For the dynamic weighting function with respect to the stationary time, To normalize the number of iterations, the value of the exponential decay rate 'a' is controlled by the inherent resonant angular frequency of the converter's LC filter. This is to avoid sliding mode control saturation and limit cycle oscillation under continuous conduction mode (CCM); where L represents inductance and C represents capacitance.

[0023] Transition phase (corresponding to oscillation suppression): Dynamic weights of the Gaussian activation model are constructed for the overshoot to form a transient penalty barrier to selectively exclude inferior solutions with high-frequency oscillations.

[0024] ,

[0025] in, For the dynamic weighting function of overshoot, Represents the natural constant. The mathematical expectation of the Gaussian activation model is... Let V be the variance of the Gaussian activation model;

[0026] Utilization phase (corresponding to steady-state accuracy): Dynamic weights for the convex growth model are constructed based on the steady-state error, and fine-tuning is performed in the later stages of convergence by forming a local strong attraction basin.

[0027] ,

[0028] in, This is a dynamic weighting function for steady-state error.

[0029] In step 3, the formula for the dynamic multi-objective composite fitness function that incorporates the load step perception penalty term is:

[0030] ,

[0031] in, Let be the value of the dynamic multi-objective composite fitness function in the k-th iteration, where k is the current iteration number; The stationary time corresponding to the k-th iteration is respectively Overshoot and steady-state error Dynamic weighting coefficients; For load disturbance sensing triggers, when the first derivative of the state variable... When the charge discharge rate exceeds the threshold, the activation value is 1. and The robustness penalty coefficient; This refers to the transient voltage drop depth. The moment when the load step change occurs. For recovery time, This is the reference voltage.

[0032] Step 5 includes:

[0033] Define the Lyapunov energy function for nonsingular terminal sliding mode control. Construct domain-specific kernel functions :

[0034] ,

[0035] Where exp represents the natural exponential function. and These represent input samples of hyperparameters with different shapes in a Bayesian optimization Gaussian process surrogate model; The characteristic length scale of the kernel function; This is the penalty coefficient; and Represent the Lyapunov energy function value obtained under the corresponding hyperparameter input sample x and the corresponding hyperparameter input sample x, respectively. The Lyapunov energy function value obtained from the evaluation; the domain-specific kernel function Embedded in a Gaussian process surrogate model, by evaluating the set of shape hyperparameters The expected improvement is verified through iterative sampling, and the optimal combination that minimizes the expected control cost is found through global search.

[0036] Step 6 includes: estimating the average steady-state off frequency based on the internal switching boundary dynamic equations of the Buck converter. ; Inverse calculation to form the hysteresis bandwidth of candidate parameters Forced physical hardware constraint boundaries:

[0037] ,

[0038] in, Where D is the DC input voltage, and D is the duty cycle. The steady-state thermal resistance of the heat sink, For steady-state operating current, The rise time of the power MOSFET is... The fall time of the power MOSFET. The highest junction temperature threshold for chip safety. For ambient temperature, This is the on-resistance of the power MOSFET;

[0039] If the mandatory physical hardware constraint boundary is not met, the fitness value of the particle is forcibly set to positive infinity, triggering the algorithm penalty elimination mechanism to avoid falling into pure algorithm optimization and ignoring the hardware constraints.

[0040] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0041] The present invention also provides a computer-readable storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0042] The present invention also provides a buck converter system, wherein the system is internally configured with a digital controller, and the digital controller operates non-singular terminal sliding mode control logic, wherein the control logic uses control parameters. The result is obtained by offline calculation and burning using the method described above.

[0043] The present invention has the following beneficial effects:

[0044] (1) Temporally Decoupled Multi-Objective Dynamic Weight Adjustment Mechanism: The core of this invention is to break away from the static weight penalty mechanism in traditional heuristic algorithms and introduce stage-aware logic. Through exponential decay function, Gaussian activation function and convex growth function, conflicting control objectives (such as fast response and damped stability) are isolated and individually dominated within a specific physical evolution window. This mechanism reshapes the fitness function terrain of parameter search in real time, effectively preventing mutual interference and constraints between various performance indicators, and completely avoiding the problem of local optima stagnation that is prone to occur in traditional algorithms.

[0045] (2) Data-driven hyperparameter tuning framework without subjective bias: For the first time, information entropy theory (EWM) and Bayesian optimization (BO) are embedded into the top layer of the optimization engine. The bottom layer uses EWM to objectively allocate the benchmark amplitude based on the actual statistical dispersion of the system; the top layer uses the BO surrogate model to globally search for the optimal shape hyperparameter, which completely eliminates the subjective bias of human experience and realizes mathematically rigorous end-to-end hyperparameter tuning.

[0046] (3) Fitness terrain guidance based on nonlinear functions: The carefully designed nonlinear evolution logic (such as Gaussian barrier) in the fitness function can selectively and quickly "reject" or "penalize" the poor solutions with high-frequency oscillations during the transition period. Due to the dynamic sensitivity of Gaussian function and convex growth function, the fitness cost of poor solutions will be rapidly amplified, thereby accelerating the algorithm to concentrate on the global optimal solution region with higher steady-state accuracy and faster convergence in the complex high-dimensional solution space, which significantly improves the efficiency and success rate of the search.

[0047] (4) Excellent transient tracking and steady-state robustness: The dynamic multi-objective collaborative optimization strategy adopted in this invention effectively takes into account the stringent requirements in the parameter optimization of non-singular terminal sliding mode controllers. Compared with traditional control methods, the buck converter optimized in this invention can achieve zero overshoot response in a very short time (milliseconds) when facing large signal disturbances (such as extreme load resistance steps or sudden changes in input power supply), which greatly reduces the output voltage deviation and significantly improves the overall dynamic tracking performance, steady-state accuracy and anti-interference robustness of the power electronic system. Attached Figure Description

[0048] Figure 1 This is a flowchart of the Adaptive Dynamic Weighted Particle Swarm Optimization (ADW-PSO) method of the present invention.

[0049] Figure 2 This is a framework diagram for intelligent optimization of the Buck converter control system parameters in an embodiment of the present invention.

[0050] Figure 3 This is a magnified comparison of the transient response of the output voltage during the startup phase.

[0051] Figure 4 This is a magnified comparison of output voltage recovery under a sudden load change. Detailed Implementation

[0052] This invention provides a method for collaborative optimization of non-singular terminal sliding mode control parameters in a buck converter. For example... Figure 1 (Flowchart of the collaborative optimization method of this invention) and Figure 2 (Buck converter control system parameter intelligent tuning framework diagram) As shown, this method relies on the Adaptive Dynamic Weighted Particle Swarm Optimizer (ADW-PSO) and the underlying physical constraint mechanism, and specifically includes the following steps:

[0053] Step 1, initialization and model building, the fitness function includes multiple performance metrics, including the following steps:

[0054] Step 1.1, based on control requirements, such as Figure 2 The closed-loop control model of the buck converter is shown within the dashed box above. The closed-loop buck circuit of this invention consists of the following key parameters and components:

[0055] Input end DC input voltage source, rated at 10V;

[0056] Energy transfer and switching elements: energy storage inductors The nominal value is 1mH, and the output filter capacitor is... nominal value ;

[0057] Load network: load resistance nominal value ;

[0058] Control and feedback network: reference output voltage 5V; configured with a non-singular terminal sliding mode controller ( It receives voltage error feedback and generates control signals, which are then... Pulse width modulation drives the switching transistor.

[0059] Step 1.2, Initialize the particle swarm algorithm ( ) and the parameter range of the controller. Set the population size. Maximum number of iterations . The controller parameter space to be optimized is set as follows: slip coefficient Nonlinear exponent Hysteresis bandwidth .

[0060] Step 1.3, design the dynamic fitness function, which is defined as:

[0061] ,

[0062] in, For load disturbance sensing triggers, when the first derivative of the state variable... When the charge discharge rate exceeds the threshold, the activation value is 1. This refers to the transient voltage drop depth. For recovery time, For reference voltage, and This is the robustness penalty coefficient. To stabilize the time, For overshoot, This represents the steady-state error. For performance indicators The dynamic weighting coefficients, For performance indicators The dynamic weighting coefficients, For performance indicators The dynamic weighting coefficients.

[0063] Step 2: Construct a two-layer data-driven calibration with dynamic weighting functions and their parameters;

[0064] To achieve stage awareness of the "exploration-transition-utilization" process in the optimization, a nonlinear dynamic weight evolution logic incorporating morphological hyperparameters and basic weights is designed:

[0065] ,

[0066] ,

[0067] ;

[0068] Step 2.1: Benchmark weight calibration is performed using the Entropy Weight Method (EWM). Heterogeneous performance data from the initial random population (N=30) is extracted and then... Perform dimensionless normalization mapping. Calculate information entropy and objectively quantify the baseline amplitudes of various performance parameters.

[0069] Step 2.2: When constructing a Gaussian Process (GP) surrogate model, traditional Bayesian optimization typically uses standard radial basis function (RBF) kernel function or Matern kernel function. These purely mathematical kernel functions only measure the Euclidean distance between sample points in the hyperparameter space, completely detached from the underlying physical convergence law of the transformer.

[0070] To adapt to the dynamic convergence trend of non-singular terminal sliding mode control (NTSMC), this invention innovatively proposes to adopt a function based on Lyapunov energy. and its derivative decay rate The modified domain-specific kernel function. Its specific mathematical expression is constructed as follows:

[0071] ,

[0072] in, and These represent two different sets of shape hyperparameter input samples in the Gaussian process surrogate model (i.e., Different combinations of values); The feature length scale is used to control smoothness; A custom physical energy penalty coefficient; These represent the Lyapunov energy functional indices of the sliding surface obtained through transient simulation evaluation under the corresponding hyperparameter configurations.

[0073] The physical mechanism and technical effect of the kernel function specific to this field lie in: the left half of the formula. The measure of geometric distance in the hyperparameter space is preserved; while the right half of the formula... This is the introduced physical energy correction term. Even if two hyperparameter sample points are geometrically close, if they lead to a change in the physical energy state of the system... A violent mutation (i.e.) When the variance is large, the correction term will quickly reduce the covariance between the two points.

[0074] This mechanism prevents the Bayesian-optimized Gaussian process surrogate model from blindly predicting based solely on mathematical distance; instead, it forces the incorporation of prior physical knowledge of "continuous energy decay." By evaluating the expected improvement (EI) of the shape hyperparameter set through this kernel function and verifying iterative sampling, the algorithm can globally lock the optimal combination that minimizes the overall control cost within the complex, high-dimensional hyperparameter space, following the path with the most robust physical energy decay. This solves the technical problem that traditional black-box optimization easily gets stuck in physically infeasible solutions.

[0075] Step 3, proceed The closed-loop iterative optimization process involves the following steps:

[0076] Step 3.1: Before each iteration begins, determine the current normalized iteration number. Using the parameters determined in step 2, the weight coefficients of the current generation are calculated and updated in real time. .

[0077] Step 3.2, perform transient simulation verification. This involves setting the candidate parameter set for each particle. Mapped to the closed-loop model of the buck converter, real-time data is extracted. The fitness scalar value is then calculated by substituting it into the dynamic fitness function.

[0078] Step 3.3, Physical feasibility verification based on the power MOSFET heat dissipation model. After obtaining the fitness value, the candidate parameters are... Substitute into the steady-state switching frequency estimation equation Secondly, considering the on-resistance of the MOSFET... With switching time characteristics (rise time) descent time ), calculate the steady-state operating current of the system Total theoretical heat dissipation loss :

[0079] ,

[0080] Secondly, the steady-state thermal resistance of the heat sink is introduced. With ambient temperature Establish chip junction temperature Thermodynamic model:

[0081] ,

[0082] To prevent the hardware from burning out, the estimated junction temperature must be strictly lower than the chip's maximum safe junction temperature threshold (i.e., ...). By combining the above three equations, the hysteresis bandwidth directly related to the algorithm parameters can be extracted. This allows us to deduce the mandatory hardware constraint boundaries that ensure physical security:

[0083] ,

[0084] In each iteration of the PSO algorithm, the system automatically selects candidate parameters. and Substitute into the above formula to calculate the safety lower limit. If the h value generated by the particle is less than the constraint boundary, then the solution is determined to have triggered high-frequency loss failure, and its fitness value is forcibly set to positive infinity. The swarm is forced to retreat to a safe physical working zone to seek optimization by severely punishing and eliminating physically unrealizable solutions.

[0085] Step 3.4 involves evolution using particle velocity and position update formulas. The specific update formulas are as follows:

[0086] Speed ​​update formula:

[0087] ,

[0088] Position update formula:

[0089] ,

[0090] in It is inertial weight. and It is a learning factor and It is a random number uniformly distributed between [0,1]. It is the historical best position of the i-th particle. It is the optimal position globally for the population.

[0091] Step 3.4: Repeat the fitness evaluation and position update process until the maximum number of iterations is reached. Output the converged optimal controller parameter vector. .

[0092] Step 4: Substitute the obtained optimal controller parameter combination into the system, and output the optimized voltage response image in conjunction with the attached diagram to specifically verify the technical advantages of the present invention.

[0093] (1) Advantages of transient start-up performance: such as Figure 3 (Enlarged comparison of output voltage transient response during startup) As shown in the diagram, compared to the traditional static weight optimization strategy (blue line, with obvious overshoot) and the piecewise weight strategy (yellow line, slow response speed), the system optimized using the physically guided dynamic weight strategy of this invention (red line) has a significantly reduced response time of approximately 2.0 ms, and also suppresses overshoot. The red line smoothly and rapidly reaches the steady-state reference voltage of 5V, which intuitively demonstrates that this method effectively decouples the contradiction between response speed and damping stability.

[0094] (2) The advantages of robustness against load mutations are manifested in: such as Figure 4 (The enlarged comparison diagram of output voltage recovery under load step change is shown.) When the system faces a drastic load step change from 10Ω to 2Ω, the converter (red line) controlled by the optimal parameters obtained in this invention exhibits a minimal output voltage drop (less than 0.05V) and rapidly recovers to the 5V steady-state standard within a very short time (within 1.0ms). Compared to traditional strategies (where the recovery time is significantly longer and the drop is deeper, as shown by the blue and yellow lines), Figure 4 This fully demonstrates that the present invention achieves unexpectedly strong anti-interference robustness under parameter perturbation and extreme load step effects.

[0095] This invention provides a method for collaborative optimization of non-singular terminal sliding mode control parameters for buck converters. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for collaborative optimization of non-singular terminal sliding mode control parameters for a buck converter, characterized in that, Includes the following steps: Step 1: Construct a state-space average model of the buck converter with inductor current and capacitor voltage as state variables, and define the system tracking error state variables. Step 2: Perform filtering preprocessing on the transient simulation output voltage of the buck converter to eliminate high-frequency switching ripple interference. Then, extract heterogeneous physical performance indicators, including the settling time with millisecond-level physical dimensions. Overshoot in percentage units and steady-state error ; Step 3: Divide the algorithm optimization cycle into different stages according to physical laws, configure nonlinear dynamic weight functions with topology reshaping capabilities for each performance index, and construct a dynamic multi-objective composite fitness function that integrates load step perception penalty terms. Step 4: Calculate the information entropy of each indicator using the entropy weight method, and objectively calibrate the base amplitudes of the dynamic weighting functions corresponding to the steady-state time in a purely data-driven manner. The base amplitude of overshoot and the fundamental magnitude of steady-state error Step 5: Using the multi-objective composite fitness value as the observation input to the Gaussian process surrogate model, construct a domain-specific kernel function based on the non-singular terminal sliding mode Lyapunov energy function and derivative decay rate correction of the buck converter. Then, perform physics-guided Bayesian optimization (BO) using the domain-specific kernel function to globally search for the optimal set of shape hyperparameters for the dynamic weight function. Where a is the exponential decay rate, is the characteristic length scale of the Gaussian function, and b is the growth exponent of the convex growth model; Step 6: Utilize the parameter-calibrated dynamic multi-objective composite fitness function to perform closed-loop iterative optimization using the particle swarm optimization algorithm. In each iteration, substitute the generated candidate parameter set into the steady-state switching frequency calculation equation based on the power MOSFET evaluation model for physical constraint verification. Forcefully eliminate solutions that lead to physically unrealizable solutions at the limiting frequency, and finally derive the optimal parameters that satisfy the hardware constraints. And configured in the digital controller, where For the optimal synovial coefficient, It is the optimal nonlinear exponent. This represents the optimal hysteresis bandwidth.

2. The method according to claim 1, characterized in that, In step 1, a non-singular terminal sliding mode is constructed. Define the set of strongly coupled control parameters to be optimized. ,in The sliding mode coefficient is , It is a non-linear exponent. The hysteresis bandwidth is the limit of the switching frequency of the corresponding power metal-oxide-semiconductor field-effect transistor.

3. The method according to claim 2, characterized in that, In step 2, to eliminate the high-frequency switching frequency during the hard switching process of the buck converter... To address high-order harmonic interference, a sliding window integral low-pass filter is used for physical preprocessing. The mapping formula is as follows: ， in, The transient output voltage is the filtered voltage, and t is the current time variable. For integration variables, For the switching cycle, This is the transient output voltage. The differential symbol is used; the processed data is then input into the minimax normalization mapping.

4. The method according to claim 3, characterized in that, Step 3, which involves configuring nonlinear dynamic weighting functions with topology reshaping capabilities for each performance index, specifically includes: Exploration Phase: Constructing dynamic weights for the exponential decay model with respect to the steady-state time: ， in, For the dynamic weighting function with respect to the stationary time, To normalize the number of iterations, the value of the exponential decay rate 'a' is controlled by the inherent resonant angular frequency of the converter's LC filter. Where L represents inductance and C represents capacitance; Transition phase: Dynamic weights for the Gaussian activation model are constructed to address overshoot, forming a transient penalty barrier to selectively exclude inferior solutions with high-frequency oscillations. ， in, For the dynamic weighting function of overshoot, Represents the natural constant. The mathematical expectation of the Gaussian activation model is... Let V be the variance of the Gaussian activation model; Utilization phase: Dynamic weights for the convex growth model are constructed based on the steady-state error, and fine-tuning is performed in the later stages of convergence when a local strong attraction basin is formed. ， in, This is a dynamic weighting function for steady-state error.

5. The method according to claim 4, characterized in that, In step 3, the formula for the dynamic multi-objective composite fitness function that incorporates the load step perception penalty term is: ， in, Let be the value of the dynamic multi-objective composite fitness function in the k-th iteration, where k is the current iteration number; The stationary time corresponding to the k-th iteration is respectively Overshoot and steady-state error Dynamic weighting coefficients; For load disturbance sensing triggers, when the first derivative of the state variable... When the charge discharge rate exceeds the threshold, the activation value is 1. and The robustness penalty coefficient; This refers to the transient voltage drop depth. The moment when the load step change occurs. For recovery time, This is the reference voltage.

6. The method according to claim 5, characterized in that, Step 5 includes: Define the Lyapunov energy function for nonsingular terminal sliding mode control. Construct domain-specific kernel functions : ， Where exp represents the natural exponential function. and These represent input samples of hyperparameters with different shapes in a Bayesian optimization Gaussian process surrogate model; The characteristic length scale of the kernel function; This is the penalty coefficient; and Represent the Lyapunov energy function value obtained under the corresponding hyperparameter input sample x and the corresponding hyperparameter input sample x, respectively. The Lyapunov energy function value obtained from the evaluation; the domain-specific kernel function Embedded in a Gaussian process surrogate model, by evaluating the set of shape hyperparameters The expected improvement is verified through iterative sampling, and the optimal combination that minimizes the expected control cost is found through global search.

7. The method according to claim 6, characterized in that, Step 6 includes: estimating the average steady-state off frequency based on the internal switching boundary dynamic equations of the Buck converter. ; Inverse calculation to form the hysteresis bandwidth of candidate parameters Forced physical hardware constraint boundaries: ， in, Where D is the DC input voltage, and D is the duty cycle. The steady-state thermal resistance of the heat sink, For steady-state operating current, The rise time of the power MOSFET is... The fall time of the power MOSFET. The highest junction temperature threshold for chip safety. For ambient temperature, This is the on-resistance of the power MOSFET; If the mandatory physical hardware constraint boundary is not met, the fitness value of the particle is forcibly set to positive infinity, triggering the algorithm's penalty elimination mechanism.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 7.

10. A buck converter system, characterized in that, The system is internally configured with a digital controller, which runs non-singular terminal sliding mode control logic. The control logic uses control parameters... It is obtained by offline calculation and burning using the method described in any one of claims 1 to 7.

Images