CLLLC resonant converter control method based on RBF neural network active disturbance rejection

By combining RBF neural networks and active disturbance rejection controllers, real-time voltage regulation of the CLLLC resonant converter under load changes and parameter perturbations was achieved, solving the output voltage fluctuation problem and improving dynamic response and voltage stability.

CN119030330APending Publication Date: 2024-11-26UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411198881.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-08-23
Filing Date
2024-08-29
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing CLLLC resonant converters are difficult to control stably under load changes and parameter perturbations, and the parameters of the active disturbance rejection controller are complex to adjust and difficult to optimize in real time.

Method used

An active disturbance rejection controller based on an RBF neural network is adopted. The observer gain is tuned online through the neural network, and combined with a nonlinear state observer and error feedback control, the system parameters are adjusted in real time and disturbances are suppressed.

Benefits of technology

It achieves rapid stabilization and low ripple of the output voltage of the CLLLC resonant converter, shortens the dynamic response time, and improves the voltage regulation performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119030330A_ABST
    Figure CN119030330A_ABST
Patent Text Reader

Abstract

The invention discloses a CLLLC resonant converter control method based on RBF neural network active disturbance rejection. The invention aims to solve the problem that an ADRC (active disturbance rejection controller) under a resonant converter formed by CLLLC (cell logic link control) is complicated in parameter setting process and difficult in real-time adjustment, so that the output voltage of a system is difficult to accurately track instruction voltage. Therefore, the invention provides an ADRC parameter real-time adjustment method based on an RBF (radial basis function) neural network, an actual CLLLC resonant converter output voltage and a reference instruction voltage are taken as an input layer of the neural network, a radial basis function neural network of an output layer structure is combined with an active disturbance rejection controller, a parameter adjuster is designed, and then the system is adjusted in real time. The method not only simplifies the complex process of traditional parameter setting, but also has the adaptive ability of online real-time setting, thereby significantly improving the control precision and response speed of the system. A simulation result verifies that the method has a remarkable effect in the aspect of control performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of power electronic converter control, and in particular relates to a CLLLC resonant converter control method based on RBF neural network self-interference rejection. Background Art

[0002] At present, with the widespread application of new energy vehicles, large-scale DC charging technology has become the development trend of new energy vehicle charging. Due to the characteristics of high transmission efficiency and high power density, CLLLC resonant converter is widely used in vehicle battery chargers. However, load mutation and parameter perturbation will cause the output voltage to fluctuate, and the stable control of the output voltage becomes a problem.

[0003] Active disturbance rejection control (ADRC) has been widely used in complex systems because it does not rely on accurate mathematical models. Although ADRC has many advantages, its algorithm is complex and has many parameters. Many parameters still need to be determined through empirical trial and error, such as error feedback gain and compensation factor. This method is time-consuming and labor-intensive, and there are many system disturbance factors, making it difficult to achieve real-time parameter adjustment, so it is difficult to find the optimal parameters.

[0004] Therefore, the key problem to be solved by the present invention is to study a method that can adaptively adjust the observer gain online to meet the needs of its controller to track the target output voltage and increase its dynamic response time. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a design method for an active disturbance rejection controller based on an RBF (radial basis function, RBF) neural network, which optimizes the active disturbance rejection controller ADRC and combines the RBF neural network with the optimized active disturbance rejection controller ADRC to achieve real-time debugging of system parameters. The present invention achieves the above technical objectives through the following technical means.

[0006] In order to achieve the above-mentioned purpose of the invention, the overall technical solution adopted by the present invention is active anti-disturbance control. Due to the complexity of parameter setting, an online neural network is used to adjust the parameters of the linear extended state observer used. The converter system modeling error part, internal disturbance and external disturbance are taken as the total disturbance. Finally, through real-time estimation and compensation, active suppression and reduction of disturbances are achieved, thereby tracking the converter voltage.

[0007] A CLLLC resonant converter control method based on RBF neural network self-disturbance rejection, wherein the controller needs to design parameters of compensation factor, controller gain and observer gain.

[0008] CLLLC circuit topology is as follows Figure 1 As shown, according to the small signal model, the transfer function describing the effect of switching frequency disturbance on output voltage can be written as: Among them, A and B are coefficient matrices determined by topological parameters, and C outputs the state column vector.

[0009] Considering the CLLLC resonant converter as a second-order control object, the transfer function from switching frequency to output voltage can be written as: In the formula, d is the external disturbance, b is the control gain, and f contains the sum of external and internal disturbances. Part b is known, and the known part is b 0 .

[0010] The system can be rewritten as: in: w is the expansion state of the system.

[0011] Design state variable x 1 =y, x 3 =w, then the state space expression is:

[0012] In order to estimate the state x 1 ,x 2 ,x 3 , the nonlinear state observer of the system can be designed as e is the observation error, z1 is the y observation signal; y is the actual value of the voltage signal output by the resonant converter; β 1 , β 2 , β 3 is the state feedback gain; α1, α2 are the nonlinear factors of the active disturbance rejection controller ADRC, and δ is the linear interval of the power function

[0013] Where fal(e,a,δ) is the nonlinear function expression

[0014] After the nonlinear state observer compensation, the error feedback control rate is designed. The control rate design is as shown in Formula 3 Where k p , k d The gain of the control rate is an adjustable parameter.

[0015] Aiming at the difficulty in adjusting the observer gain parameters, a neural network is used to perform parameter tuning online. The first step is to construct a neural network with 16 input layer nodes, 3 hidden layer nodes, and 1 output layer node.

[0016] The hidden layer neurons use Gaussian function, which is expressed as follows: Where X = [x 1 ,x 2 ,x 3 ] is the network input vector, C j =[c j1 ,c j2 ,…c jn ] is the center vector of the jth node in the network, |||| represents the distance between the input vector and the center vector, and bj is the base width vector of the jth node;

[0017] Linear weighting is used from the hidden layer to the output layer, and the expression is as follows: where ω=[ω 1 ,ω 2 ,…ω m ] T is the weight vector of the network output layer;

[0018] The performance indicator function is selected as: Among them: y(k) is the actual output of the system, y m (k) is the output of the radial basis neural network; the optimization method used is the gradient descent method, and its algorithm expressions are as follows:

[0019] Among them: X=[u(k),y(k),y(k-1)] T is the input vector of the radial basis neural network, η is the learning rate, and α is the momentum factor;

[0020] After the RBF neural network obtains the fitting output of the identified object, a real-time Jacobian matrix can be obtained. The partial derivative of the radial basis neural network output to input is the Jacobian information, and its expression is:

[0021] After training, the neural network will get an identification output y m (k), and the identification error is very small. On this basis, it can be considered that the identification output y m (k) is equivalent to the actual output y(k);

[0022] Then the undetermined controller parameter β is adjusted in real time, and the performance index function is selected as: In the formula, r(k), y(k) are the input and output signals of the system;

[0023] The weight coefficient β uses the gradient descent method, and the formula is as follows:

[0024] The above is the technical solution for neural network parameter tuning of this design. In the process of system identification, the Jacobian information shown in formula (17) is generally used as the partial derivative of the output to the input to characterize the sensitivity change of the output to the input, and also reflects the optimal linear approximation of a differentiable equation to a given point. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 , a circuit diagram of a CLLLC resonant converter in the present invention;

[0026] Figure 2 , an equivalent circuit diagram of the CLLLC resonant converter in the present invention;

[0027] Figure 3 , the control block diagram based on neural network auto-disturbance rejection in the present invention;

[0028] Figure 4 , the simulated output voltage response waveform diagram of the present invention;

[0029] Figure 5 , the local waveform diagram of the simulated output voltage response in the present invention;

[0030] Figure 6 , the simulated output voltage ripple waveform diagram of the present invention; DETAILED DESCRIPTION

[0031] The technical solution will be clearly and completely described below in conjunction with the preferred examples and drawings of the present invention. It should be understood that the preferred examples are only for illustrating the present invention, not for limiting the scope of protection of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0032] The present invention provides a CLLLC resonant converter control method based on RBF neural network auto-disturbance rejection. The overall technical solution adopted by the present invention is auto-disturbance rejection control. The nonlinear extended state observer inside the controller can estimate the total disturbance including load mutation and model deviation in real time and compensate for it. The observer gain is combined with the neural network to design a parameter regulator, which eliminates the difficulty of manual parameter adjustment. The auto-disturbance rejection controller obtained by calculation is used as voltage regulation to adjust the system deformity in real time.

[0033] Reference Figure 2 , a mathematical model of the CLLLC resonant converter is established. The steps include obtaining the equivalent circuit by extending the description function method and writing the nonlinear circuit equation according to Kirchhoff's law:

[0034] Specifically, the equivalent circuit diagram of the CLLLC resonant converter is as follows: Figure 2 As shown, where L r1 , L r2 are the resonant inductors on the primary and secondary sides, C r1 , C r2 are the resonant capacitors on the primary and secondary sides, L m is the excitation inductance, n is the transformer ratio, i 1 is the resonant current of the primary side resonant network, u 1 is the voltage of the primary side resonant capacitor, i 2 is the resonant current of the secondary side resonant network, u 2 is the voltage of the secondary side resonant capacitor, i m is the magnetizing inductor current, u AB and u CD is the square wave voltage on the primary and secondary sides, C o is the output filter capacitor, R L is the load, U o is the output voltage.

[0035] Further, the steps of constructing the transfer function from switching frequency to output voltage and its reduced-order transfer function expression include: the transfer function from system input to output can be expressed as: describing the switching frequency disturbance Corresponding output voltage impact The transfer function of the CLLLC resonant converter can be obtained by reducing the order:

[0036] A CLLLC resonant converter control method based on RBF neural network self-disturbance rejection, the state error feedback control rate is designed as Among them, k p , k d is the adjustable parameter of the tracker, r is the target voltage value of the output;

[0037] According to the active disturbance rejection control method, the system transfer function needs to be rewritten as:

[0038] Where y is the output voltage of the CLLLC resonant converter, u is the input signal that drives the switch, and b 0 The input gain of the known part, is the internal disturbance of the system, w is the external disturbance, and f is the total disturbance of the CLLLC resonant converter system, including the output y and the first-order derivative of the output Input u and the first derivative of the input As well as unknown disturbances in actual control, such as parameter perturbations, sampling errors, unmodeled dynamics of equipment, and other uncertainties.

[0039] Take the state variable x 1 =y, x 3 =fFor this second-order controlled object, a third-order linear extended state observer is designed: Among them, β 1 ,β 2 ,β 3 is the tracking error gain coefficient of the linear extended state observer, z 1 ,z 2 ,z 3 is the output of the linear extended state observer, which is the state variable x 1 ,x 2 ,x 3 is an estimate of , and y is the actual value of the observed quantity.

[0040] Step a: Build a three-layer neural network and get the output y(k) of the neural network;

[0041] The number of input layer nodes of the radial basis neural network is 3, the number of hidden layer nodes is 3, and the number of output layer nodes is 1; the hidden layer neurons use Gaussian functions, and their expressions are as follows:

[0042] Where: X = [x 1 ,x 2 ,x 3 ] is the network input vector, C j =[c j1 ,c j2 ,…c jn] is the center vector of the jth node in the network, ||·|| represents the distance between the input vector and the center vector, b j is the base width vector of the jth node;

[0043] Linear weighting is used from the hidden layer to the output layer, and the expression is as follows: y m =w T h=w 1 h 1 +w 2 h 2 +...+w m h m (twenty three)

[0044] where ω=[ω 1 ,ω 2 ,…ω m ] T is the weight vector of the network output layer;

[0045] Step b: Optimize the network parameter center vector, base width vector and weight vector of the hidden layer of the neural network; the performance index function selected here is:

[0046] Among them, y(k) is the actual output of the system, y m (k) is the output of the radial basis neural network. The optimization method used is the gradient descent method, and its algorithm expressions are shown in formulas (11), (12), and (13) respectively;

[0047] Where X = [x 1 ,x 2 ,x 3 ] is the input vector of the radial basis neural network, η is the learning rate, and a is the momentum factor;

[0048] The Jacobian information expression of the radial basis neural network output is:

[0049] Step c: using the Jacobian information obtained in step b to adjust the weight coefficients in the nonlinear error feedback control module in the ADRC in real time;

[0050] The performance index function selected here is:

[0051] Where: r(k) is the input signal of the system, and y(k) is the output signal of the system.

[0052] The weight coefficient β is adjusted using the gradient descent method, and the formula is as follows:

[0053] in: is the Jacobian information obtained in step b, and η is the learning rate.

[0054] In order to prove the correctness and effectiveness of the solution of the present invention, the simulation model was built in MATLAB / Simulink, and the simulation parameters are shown in Table 1:

[0055] The system operation process is as follows Figure 4 ;

[0056] From the above simulation results, it can be seen that the neural network ADRC basically has no overshoot. Figure 5 It can be seen that its adjustment time is 0.09ms faster than that of the traditional ADRC, and its dynamic performance is better. Figure 6 By comparing the waveforms, it can be seen that the output voltage ripple is lower than 0.621V than the traditional anti-disturbance control ripple, and the steady-state value is better.

[0057] The present invention cites specific embodiments to illustrate the principles and implementation methods of the present invention. Professionals and technicians in this field will realize that the above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. The general principles defined in this article can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in this article, but should conform to the widest scope consistent with the principles and novel features disclosed in this article. All equivalent changes or modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims

1. A CLLLC resonant converter control method based on RBF neural network self-disturbance rejection, characterized in that: Provided is a resonant converter control method for online adjustment of auto-disturbance rejection control parameters; a CLLLC resonant converter control method based on RBF (radial basis function, RBF) neural network auto-disturbance rejection, online adjustment of observer gains β1, β2, β3 so that the output voltage signal can track the command voltage signal; the method includes the following steps. Step 1: Use the transfer function from the switching frequency to the output voltage of the CLLLC resonant converter to establish a second-order control object. Where, d is the external disturbance, b is the gain of the control quantity, It is the comprehensive characteristics of the unknown time-varying dynamics and external disturbances in the system. Step 2: Design a second-order ADRC and select the input u = f s , output y = u o , state variables Expand the sum of the internal and external disturbances of the system into a new state variable x3, and record The expanded linear system is as follows. The state error feedback control rate is designed to be. Among them, k p , k d is the adjustable parameter of the tracker, and r is the target output voltage value. The third-order linear extended state observer (LESO) takes the state variable x1 = y, x3=fFor this second-order controlled object, the observer is designed as. Among them, β1, β2, β3 are the tracking error gain coefficients of the linear extended state observer, z1, z2, z3 are the outputs of the linear extended state observer, which are estimates of the state variables x1, x2, x3, and y is the actual value of the observed quantity. Step 3: The nonlinear function coefficients, β1, β2, β3, of the state observer in ADRC (active disturbance rejection control) are adjusted in real time, and the hidden layer activation function selects a Gaussian function with excellent performance, and the expression is as follows. In the formula, c i =[c i1 ,c i2 ,…c iN ] T is the Gaussian center point of the i-th hidden layer neuron, b i is the base width length of the i-th hidden layer node, which is used to adjust the sensitivity of the neuron. ||·|| is the Euclidean vector norm, which represents the distance between the input neuron and the center point vector. The core of the RBF network model design is to determine the hidden layer basis function parameter c i 、b i Link weight w ij . The gradient descent method is used to learn and train the network model parameters.

2. The control method of a CLLLC resonant converter based on RBF neural network self-disturbance rejection according to claim 1 is characterized in that: The active disturbance rejection controller design specifically includes: The steps to construct the transfer function from switching frequency to output voltage and its reduced-order transfer function expression include: The transfer function from system input to output can be expressed as: Describing the switching frequency disturbance Corresponding output voltage impact The transfer function of the CLLLC resonant converter can be obtained by reducing the order. The system transfer function is simplified to: in, is the expanded state of the system. Generally, part b can be considered to be known, and the known part is bo.

3. The control method of a CLLLC resonant converter based on RBF neural network self-disturbance rejection according to claim 2 is characterized in that: The neural network anti-disturbance control specifically includes. Construct a three-layer neural network and obtain the output y(k) of the neural network; the number of input layer nodes is 12, the number of hidden layer nodes is 6, and the number of output layer nodes is 1; the hidden layer neurons use Gaussian function, which is expressed as follows: Where: X = [x1, x2, x3] is the network input vector, C j =[c j1 ,c j2 ,…c jn ] is the center vector of the jth node in the network, ||·|| represents the distance between the input vector and the center vector, b j is the base width vector of the jth node; The performance indicator function is selected as: Adjust the parameters that need to be debugged of ADRC so that the actual value y(k) of the voltage signal of the resonant converter at time k is close to the given signal v(k) of the voltage signal of the resonant converter at time k. According to the gradient descent algorithm, the increment △n(k) of the parameter n that needs to be debugged is obtained; Substituting y(k) with the tracking value ym(k) of y(k) yields △n(k): The parameter n that needs to be debugged after adjustment at time k is obtained through calculation.