Double-closed-loop model-free control method for high-gain DC-DC converter based on data driving

Abstract

The invention discloses a double-closed-loop model-free control method for a high-gain DC-DC converter based on data driving. The method comprises the following steps: acquiring input duty ratio, output voltage and current data of the DC-DC converter under an open-loop condition, and filtering; a data twinborn model of the converter is constructed by adopting NARX, and dynamic characteristics of the converter are accurately represented through open-loop training and closed-loop conversion; a double-closed-loop neural network controller composed of a voltage outer loop controller and a current inner loop controller is designed, the voltage outer loop controller and the current inner loop controller are both three-layer feed-forward networks, input comprises errors, error integral terms and error differential terms, and nonlinear mapping capacity is achieved; an ADP algorithm is adopted, a performance index function is minimized, and a batch gradient descent method is combined to automatically optimize the weight and bias parameters of the controller; and the trained controller parameters are imported into a data processor to realize real-time high-performance control of the DC-DC converter. The method does not need to depend on an accurate mathematical model, parameters are automatically optimized, the dynamic performance is excellent, the method is suitable for various DC-DC converter topologies, and engineering implementation is easy.

Description

Technical Field

[0001] This invention relates to the field of high-gain DC-DC converter control, and in particular to a data-driven dual-closed-loop model-free control method for high-gain DC-DC converters. Background Technology

[0002] With the rapid development of industries such as new energy vehicles and photovoltaic power generation, high-gain DC-DC converters are playing an increasingly important role in energy conversion systems. To achieve high-gain and high-efficiency energy conversion, various boost modules are introduced into the topology, leading to a significant increase in system modeling complexity.

[0003] Traditional model-based control methods suffer from the following problems when dealing with complex high-gain converters:

[0004] First, modeling is difficult. High-gain DC-DC converters typically include multiple sub-modules such as coupled inductors, active clamping circuits, and voltage multiplier units. Solving the state equations is complex, and actual circuits contain parasitic parameters and component nonlinearities, leading to significant deviations between theoretical models and actual systems.

[0005] Secondly, they are highly sensitive to parameters. Traditional model-based control methods such as PID control, model predictive control (MPC), and linear quadratic regulator (LQR) are sensitive to changes in system parameters. When component parameters change or operating conditions change, control performance deteriorates significantly.

[0006] Secondly, dynamic performance is limited. Single-voltage loop control relies solely on output voltage feedback, resulting in slow dynamic response and large overshoot when facing high-frequency disturbances or rapid load changes, making it difficult to meet the requirements of new energy systems for rapid dynamic adjustment.

[0007] Finally, controller parameter tuning is difficult. Traditional controller parameter design relies on expert experience and repeated experiments, resulting in long design cycles and high costs.

[0008] Existing model-free control methods, such as model-free adaptive control (MFAC), while not relying on an exact model, involve complex parameter design processes and still have room for improvement in control performance. In recent years, neural network control has been applied in DC-DC converter control due to its powerful nonlinear mapping capabilities; however, existing neural network control methods have the following shortcomings:

[0009] (1) There is a lack of accurate system equivalent models for parameter updates;

[0010] (2) They often use deep neural network structures, which require a large amount of training data and have high computational complexity;

[0011] (3) The controller parameters from simulation training are difficult to directly transfer to the actual hardware system;

[0012] (4) The single voltage loop control structure limits further improvement of dynamic performance.

[0013] Therefore, there is an urgent need for a DC-DC converter control method that does not require a precise mathematical model, automatically updates parameters, has excellent dynamic performance, and is easy to implement in engineering. Summary of the Invention

[0014] The main objective of this invention is to overcome the aforementioned deficiencies in the prior art and propose a data-driven, high-gain DC-DC converter dual-loop model-free control method. By constructing a NARX data twin model and a dual-loop neural network controller, and combining an adaptive dynamic programming algorithm to automatically optimize the controller parameters, high-performance control of the DC-DC converter is achieved.

[0015] The present invention adopts the following technical solution:

[0016] A data-driven, high-gain DC-DC converter dual-loop model-free control method includes:

[0017] S1. Construct a neural network controller to control the output voltage of the high-gain DC-DC converter; the neural network controller consists of a current inner-loop neural network controller I-NNC and a voltage outer-loop neural network controller V-NNC, with the voltage error as the input to the voltage outer-loop neural network controller. Voltage integral term and voltage differential term The output is the current reference value. The input to the current inner-loop neural network controller is the current error. Current integral term and the current differential term The output is the duty cycle. The voltage outer loop neural network controller network has a single hidden layer with three neurons. The activation function of the hidden layer is the hyperbolic tangent function, and the activation function of the output layer is the Purelin function. The current inner loop neural network controller network has a single hidden layer with three neurons. The activation function of the hidden layer is the hyperbolic tangent function, and the activation function of the output layer is the sigmoid function.

[0018] S2, Construct and train a high-gain DC-DC converter twin model with a nonlinear autoregressive external input NARX model; the NARX model has three hidden layers, each containing three neurons, with the hidden layer activation function being the hyperbolic tangent function and the output layer activation function being the Purelin function; the external input data of the NARX model is the duty cycle. The output is a predicted value of the output voltage and input current. and ;

[0019] S3. Based on the neural network controller, a converter closed-loop circuit is built on Simulink. The adaptive dynamic programming (ADP) algorithm is used to update the I-NNC and V-NNC parameters, set the performance evaluation index functions of I-NNC and V-NNC, and initialize the I-NNC and V-NNC parameters; the parameters include weights and biases.

[0020] S4, keeping the parameters of V-NNC fixed, iterate the weights and biases of I-NNC based on the performance evaluation index function of I-NNC using the batch gradient descent method;

[0021] S5. Update the weights and biases of I-NNC based on the iteration results, and select multiple initialization points to repeat S4 to S5 until the performance evaluation index function value calculated based on the performance evaluation index function of I-NNC reaches the preset target value or the iteration stops after a given number of iterations, thus obtaining the optimal current inner loop neural network controller.

[0022] S6. Keep the weights and biases of the optimal current inner loop neural network controller fixed, and iterate the weights and biases of the V-NNC based on the performance evaluation index function of the V-NNC using the batch gradient descent method.

[0023] S7. Update the weights and biases of V-NNC based on the iteration results, and select multiple initialization points to repeat S6 to S7 until the performance evaluation index function value based on V-NNC reaches the preset target value or the iteration stops after a given number of iterations, thus obtaining the optimal voltage outer loop neural network controller.

[0024] S8 outputs the weights and biases of the optimal current inner-loop neural network controller and the optimal voltage outer-loop neural network controller, and controls the DC-DC converter based on the parameters of the optimal neural network controller.

[0025] The preferred steps for constructing and training a nonlinear autoregressive NARX model with external input are as follows:

[0026] Construct a NARX model, wherein the NARX model has a duty cycle As input data, the predicted output voltage value and the predicted input current value As the output; simultaneously, the NARX model output, after a time delay, becomes the input of the model itself, and its input-output relationship is expressed as: Where d(k) and d(kn) represent the duty cycle values ​​of the controller at times k and kn, respectively; and These represent the output voltage values ​​of the converter at times k and km, respectively; and represents the input current value of the converter at times k and km, respectively, and m and n represent the delay step size, respectively; This represents the nonlinear mapping relationship of the NARX model; the NARX model is set to have 3 hidden layers, each containing 3 neurons, with the hidden layer activation function being the hyperbolic tangent function, and the output layer activation function being a linear function;

[0027] A physical model of the converter circuit was built and operated in open loop under varying input duty cycle and output load conditions. Duty cycle, output voltage and input current data were collected, and the collected data were used to train a twin model of a high-gain DC-DC converter.

[0028] The Levenberg-Marquardt algorithm was used to train the NARX model and minimize the mean squared error. The closed-loop NARX model conversion was completed, and the model with the best validation results was selected as the final converter twin model.

[0029] Preferably, the performance evaluation index functions for the current inner-loop neural network controller I-NNC and the voltage outer-loop neural network controller V-NNC are expressed as follows: The dual-loop performance evaluation index functions are expressed as follows:

[0030] I-NNC performance evaluation index function It is expressed as follows:

[0031] ;

[0032] Where i represents the number of iterations;

[0033] V-NNC performance evaluation index function It is expressed as follows:

[0034] ;

[0035] in, This represents the input current sampled at time k. This represents the reference current at time k. This represents the output voltage sampled at time k; The reference voltage at time k is represented; M represents the total number of sampling points for input current and output voltage.

[0036] Preferably, the weights and biases of the neural network controller are iteratively applied using the batch gradient descent method, with the iterative formula expressed as follows:

[0037] ;

[0038] in, Indicates the first The weights of the neural network in the next iteration; This represents the weights of the neural network at the i-th iteration; The learning rate is typically between 0 and 1. This represents the performance evaluation index function value at the i-th iteration; This represents the gradient value of the performance evaluation index function with respect to the weights at the i-th iteration; Indicates the first The bias of the neural network at the next iteration; This represents the bias of the neural network at the i-th iteration; This represents the gradient value of the performance evaluation index function with respect to the bias at the i-th iteration.

[0039] Preferably, the gradient values ​​of the performance evaluation index function with respect to the weights and the gradient values ​​of the performance evaluation index function with respect to the bias in the i-th iteration of I-NNC are expressed as follows:

[0040] ;

[0041] in,

[0042] ;

[0043] In the formula, Indicates the total number of input current sampling points; This represents the input current sampled at time k; This represents the duty cycle at time k; This represents the input-output mapping relationship of a feedforward neural network;

[0044] The gradient values ​​of the performance evaluation index function with respect to the weights and the performance evaluation index function with respect to the biases during the i-th iteration of V-NNC are expressed as follows:

[0045] ;

[0046] Where N represents the total number of output voltage sampling points; This represents the output voltage sampled at time k; This represents the duty cycle at time k; This represents the reference current at time k; Represents the reference current at time k. With input current difference.

[0047] Preferred, and It represents the cumulative gradient of the batch of data, not the gradient of a single data point.

[0048] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) This invention realizes dual closed-loop control of DC-DC converter through data-driven method. Compared with the traditional control method, it is not limited by the mathematical model of the controlled object. The control method does not need to know the state space expression of the controlled system. It only needs to construct the NARX data twin model through the input and output data of the system to realize high-performance control of the system. This method solves the problems of complex controller parameter solution and difficulty in accurately establishing system model in traditional control methods.

[0050] (2) The present invention adopts a dual closed-loop neural network control structure, which significantly improves the dynamic response speed and anti-interference capability of the system compared with dual closed-loop PI control. Experiments show that under the reference voltage jump condition, the dynamic response time is shortened by more than 30% compared with single voltage loop control, and the voltage overshoot is reduced by 8.6%. Under the load change condition, the settling time is shortened by 25%, and the overshoot is reduced by 4.2%.

[0051] (3) The adaptive dynamic programming method of the present invention sets an appropriate performance index function with the aim of minimizing the performance index function, and updates the weights and biases of the neural network controller by combining the batch gradient descent algorithm, and finally obtains the optimal neural network controller; compared with the traditional method, the overshoot is smaller and the response time is faster.

[0052] (4) The controller parameters trained based on the NARX data twin model can be directly applied to the actual hardware system. The simulation results are highly consistent with the experimental results, avoiding repeated parameter tuning and improving the efficiency of engineering implementation.

[0053] (5) The present invention adopts a shallow neural network structure, which significantly reduces the amount of computation and storage requirements compared with deep neural networks, and is easy to implement real-time control on low-cost DSPs or MCUs.

[0054] (6) The method of the present invention is not limited to a specific DC-DC converter topology, but is applicable to various topologies such as Buck, Boost, Buck-Boost, and high-gain converters, and has good versatility and scalability. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the high-gain DC-DC converter topology according to an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the dual closed-loop neural network controller structure according to an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of a closed-loop circuit according to an embodiment of the present invention;

[0058] Figure 4This is a schematic diagram of NARX model fitting in an embodiment of the present invention;

[0059] Figure 5 A flowchart illustrating the ADP algorithm for updating neural network controller parameters in an embodiment of the present invention;

[0060] Figure 6 The control methods of this invention are compared with those of single-loop NNC and dual-closed-loop PI control, showing the voltage following effect.

[0061] Figure 7 The diagram shows the voltage jump tracking effect of single-loop NNC and dual-closed-loop PI control in the embodiment of the present invention.

[0062] Figure 8 The diagram shows the load change tracking effect of single-loop NNC and dual-closed-loop PI control in the embodiment of the present invention. Detailed Implementation

[0063] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0064] This embodiment presents a data-driven, high-gain DC-DC converter dual-closed-loop model-free control method, comprising the following steps.

[0065] Step 1: Data Acquisition. Acquire the input and output data of the DC-DC converter under open-loop conditions; set the input voltage and load of the DC-DC converter to remain constant; start the converter in open-loop mode, controlling the duty cycle to randomly jump within the allowable range, with a jump time interval of 0.05~0.2s; acquire the duty cycle during the converter's operation. Output voltage and input current Data, its Indicates the first Each sampling time point; the collected data is filtered and smoothed to eliminate high-frequency noise.

[0066] Step 2: Construct a dual-loop neural network controller. This involves building a voltage outer-loop neural network controller (V-NNC) and a current inner-loop neural network controller (I-NNC) to form a dual-loop control structure. Both the current inner-loop controller (I-NNC) and the voltage outer-loop controller (V-NNC) employ a three-layer feedforward neural network structure. Their input layer is designed to contain three neurons, theoretically capable of receiving the error, the error integral term, and the error derivative term, respectively. In this embodiment, to verify the core control concept and simplify the design, only the error and its integral term are used as the actual input to the controller.

[0067] Specifically, the input layer of the current inner loop controller I-NNC contains 3 neurons, with the input being the current error and the current error integral; the hidden layer contains 3 neurons, with the hyperbolic tangent function as the activation function; and the output layer contains 1 neuron, with the activation function being... Function that outputs the duty cycle ;

[0068] The voltage outer loop controller V-NNC has an input layer containing three neurons, with the input being the voltage error and the voltage error integral; a hidden layer containing three neurons, with the hyperbolic tangent function as the activation function; and an output layer containing one neuron, with the activation function being... Function, input current reference value .

[0069] Step 3: Construct the NARX data twin model. Based on the collected input and output data, a data twin model of the DC-DC converter is constructed using a nonlinear autoregressive neural network (NARX) with external input. The NARX model structure parameters are determined: the input delay step size *m* and the output delay step size *n*, with values ​​ranging from 2. The hidden layer activation function is the hyperbolic tangent function; the output layer activation function is a purely linear function. Open-loop training of the NARX model is performed: the collected data is divided into training and test sets in a 7:3 ratio. The input-output relationship of the NARX model in the open-loop state is expressed as: The Levenberg-Marquardt algorithm is used to train the NARX model to minimize the mean square error. The closed-loop NARX model conversion is completed: after training, the open-loop NARX model is converted to closed-loop mode. The model output y*(k) is then fed back as the model input after a delay, forming a closed-loop data twin model. Model validation is performed: the predictive performance of the closed-loop NARX model is validated on the test set to ensure that the data twin model can accurately represent the dynamic characteristics of the actual converter.

[0070] Step 4: Controller parameter update based on adaptive dynamic programming. The adaptive dynamic programming (ADP) algorithm is used to optimize the weights and biases of the I-NNC and V-NNC respectively. This includes the following steps.

[0071] Define the performance index function: For I-NNC, define the current performance evaluation index function value. For V-NNC, define the voltage performance evaluation index function value. .

[0072] Set the training reference signal: Use a square wave signal as the reference signal; the square wave frequency should be much higher than the control frequency. Update the I-NNC parameters: Initialize the I-NNC weights. and bias ; Construct a simulation system in Simulink that includes a closed-loop NARX model, I-NNC, and V-NNC, keeping the V-NNC parameters fixed; Run the simulation and collect data to calculate the initial performance metrics; Update the I-NNC parameters using the batch gradient descent algorithm; Rerun the simulation to determine whether the convergence condition is met or the maximum number of iterations is reached; If convergence is not achieved, continue iterating; if convergence is achieved, save the optimized I-NNC parameters.

[0073] Perform V-NNC parameter update: With I-NNC parameters fixed, use the same ADP algorithm to update V-NNC parameters.

[0074] Stability verification: After training, verify the local stability of the dual-loop neural network control system.

[0075] Step 5: Import the trained I-NNC and V-NNC parameters into a data signal processor or microcontroller; implement a dual closed-loop control structure in the control program: sample the output voltage. and input current Calculate the voltage error and its integral, and input V-NNC to obtain the current reference value. Calculate the current error and its integral, and input it into I-NNC to obtain the control duty cycle. The duty cycle signal is output to the PWM module to drive the power switching transistor.

[0076] See Figure 1 As shown in the embodiment, the high-gain DC-DC converter controlled by the method of the present invention consists of five parts: a DC power supply, a coupling inductor, a voltage multiplier unit, an active clamping circuit, and an output load. The DC power supply is... Coupled inductors include magnetizing inductors. Leakage inductance The voltage multiplier unit includes two switched capacitors. , and two diodes , The active clamping circuit includes an auxiliary switch. and clamping capacitor Output load is In addition, it also includes the main switch. and its parasitic capacitance Cr, output diode .

[0077] For details, see Figure 2 As shown, the dual closed-loop neural network controller of the present invention (including a current inner-loop neural network controller I-NNC and a voltage outer-loop neural network controller V-NNC) uses e(t) transformed by the hyperbolic tangent function and (Corresponding to I-NNC, representing current error) and current integral term ; corresponding to V-NNC, which represents the voltage error. and voltage integral term Using this as input, mapping the input data to the range of 0 to 1 helps eliminate differences in dimensions and value ranges between different features. V-NNC output current reference signal , with input current The result of the subtraction is input to I-NNC, and I-NNC outputs a high-gain DC-DC control signal with a duty cycle d(k).

[0078] For details, see Figure 3 The figure shows a NARX (Nonlinear Autoregressive External Input) closed-loop circuit model built according to an embodiment of the present invention, used for the iterative process of adaptive dynamic programming. The NARX external input data in the figure represents the duty cycle. The output consists of predicted values ​​for the output voltage and input current. and Meanwhile, the NARX model output, after a time delay, becomes the model's own input. The input-output relationship can be expressed as:

[0079] Where d(k) and d(kn) represent the duty cycle values ​​of the controller at times k and kn, respectively; and These represent the output voltage values ​​of the converter at times k and km, respectively; and represents the input current value of the converter at times k and km, respectively, and m and n represent the delay step size, respectively.

[0080] The training process of the NARX model will be described below:

[0081] a, Build as Figure 1 The experimental platform shown keeps the circuit's input DC voltage and load constant.

[0082] b. Open-loop start-up circuit, which allows the duty cycle to jump within the allowable random jump range, making it convenient to observe the dynamic response of the output voltage;

[0083] c, Duty cycle during the data collection experiment Output voltage and input current The data serves as the training dataset for the NARX neural network.

[0084] d. Filter and smooth the data to make the final data a smooth curve, which is beneficial for subsequent network training;

[0085] e. The dataset is split for training and testing respectively. The LM (Levenberg-Marquardt) algorithm is used to train the open-loop NARX model to obtain the minimum mean square error of output voltage and input current.

[0086] f. After the open-loop NARX model is trained, it is transformed into a closed-loop network architecture. The prediction performance of the closed-loop NARX model on the validation set is checked to obtain the minimum mean square error of the prediction results.

[0087] The trained closed-loop NARX model has high fitting accuracy for the transients of input current and output voltage of high-gain DC-DC converters. Then, the adaptive dynamic programming algorithm (ADP) can be used to update the neural network parameters.

[0088] For details, see Figure 4 As shown, the steps for updating neural network parameters using the adaptive dynamic programming algorithm are as follows.

[0089] a. Start the closed-loop circuit. After the circuit stabilizes, collect d(k) and v in each control cycle. o (k), i in (k) data. Combined , Processing these data yields e(t). and performance evaluation index function value .

[0090] b. Based on the collected and calculated data, the parameters of the neural network controller are iterated. Here, the batch gradient descent algorithm is used for iteration, and the iterative formulas for weights and biases are:

[0091] ;

[0092] in, Indicates the first The weights of the neural network in the next iteration; This represents the weights of the neural network at the i-th iteration; The learning rate is typically between 0 and 1. This represents the performance evaluation index function value at the i-th iteration; This represents the gradient value of the performance evaluation index function with respect to the weights at the i-th iteration; Indicates the first The bias of the neural network at the next iteration; This represents the bias of the neural network at the i-th iteration; This represents the gradient value of the performance evaluation index function with respect to the bias at the i-th iteration.

[0093] c. Calculate the gradient of the objective function with respect to the weights and biases. According to the chain rule, the gradient of the objective function with respect to the weights and biases can be expressed as:

[0094] The gradient values ​​of the performance evaluation index function with respect to the weights and the performance evaluation index function with respect to the biases during the i-th iteration of I-NNC are expressed as follows:

[0095] ;

[0096] ;

[0097] in, Indicates the total number of input current sampling points; This represents the input current sampled at time k; This represents the duty cycle at time k; This represents the input-output mapping relationship of a feedforward neural network;

[0098] The gradient values ​​of the performance evaluation index function with respect to the weights and the gradient values ​​of the performance evaluation index function with respect to the biases during the i-th iteration of V-NNC are expressed as follows:

[0099] ;

[0100] Where N represents the total number of output voltage sampling points; This represents the output voltage sampled at time k; This represents the duty cycle at time k; This represents the reference current at time k; Represents the reference current at time k. With input current difference; and It can be calculated from the input and output of the neural network controller and the model parameters.

[0101] d. Update the weights and biases of the neural network controller based on the obtained gradient data to complete one iteration.

[0102] e. Restart the circuit and repeat the iteration process until the iteration converges or the number of iterations reaches the set maximum number of iterations.

[0103] f, the neural network controller after iteration is the optimal controller, and its parameters are output and recorded. The neural network controller after iteration can be directly used for voltage control of the converter. Specifically, the control signals output by the neural network controller are used to control the switching transistors (MOSFET, IGBT) of the DC-DC converter to turn on and off, thereby controlling the output voltage of the DC-DC converter and achieving optimal control of the DC-DC converter.

[0104] In this embodiment, the weights and bias vectors of the current inner-loop neural network controller I-NNC and the voltage outer-loop neural network controller V-NNC are trained and optimized so that the resulting closed-loop control system satisfies the Lyapunov stability condition near the equilibrium point. The specific analysis process is as follows:

[0105] A) Constructing an augmented state-space model: By combining the state variables of the high-gain DC-DC converter with the error integral and error differential terms of the neural network controller, an augmented state-space equation of the following form is constructed:

[0106] ;

[0107] in, These are the initial state variables of the converter. For the error integral term, This is the error differential term.

[0108] B) Equilibrium point linearization: Determine the equilibrium point of the control system as... ,in and Let be a non-zero constant, and use coordinate transformation:

[0109] ;

[0110] By shifting the equilibrium point to the origin, and linearizing the augmented state-space equations at the origin, we obtain the linearized system equations in the following form:

[0111] ;

[0112] in, This is the state vector after coordinate transformation. The system matrix after linearization is defined as the Jacobian matrix obtained by taking the partial derivatives at the equilibrium points:

[0113] ;

[0114] in, For the system's output error, and , This represents the system's target state. The transformed error integral term, and ; The transformed error differential term, and ; The relationship between each state variable and time;

[0115] C) Stability Criterion: The weights and bias vectors of the neural network controller must satisfy the following conditions to ensure that the control system is locally asymptotically stable near the equilibrium point:

[0116] ;

[0117] in, This indicates the operation of taking the real part of a complex number. This indicates the operation of taking the eigenvalues ​​of a matrix, which requires that the real parts of all eigenvalues ​​of matrix G be negative.

[0118] See Figure 6 As shown, among the three controllers and three control methods, the dual-closed-loop NNC has the fastest dynamic response time and the smallest overshoot during adjustment.

[0119] See Figure 7 The figure shows the voltage jump experimental waveforms of the control method of this invention compared to single-voltage-loop NNC control and dual-closed-loop PI control. When the reference voltage jumps from 180V to 140V, the voltage stabilization time is 30ms under single-voltage-loop NNC control, and the current stabilization time is 20ms, with no voltage undershoot. Figure 7 As shown in (a). Under dual closed-loop PI control, the voltage settling time is 34ms with an undershoot of 10V, and the current settling time is 30ms with an undershoot of 2.1A. Figure 7 As shown in (b). Under dual closed-loop NNC control, the voltage settling time is 17ms with an undershoot of 1.5V, and the current settling time is 13ms with an undershoot of 2A. Figure 7 As shown in (c), although the dual-loop NNC has a small amount of voltage and current overshoot, its dynamic response speed is better than that of the single-loop NNC and the dual-loop PI. The dual-loop PI has the largest voltage and current overshoot. This is partly because the PI dual-loop parameter design is based on a small-signal model, which differs from the actual circuit model; and partly because the linearity of PI control leads to its control performance being inferior to that of the dual-loop NNC.

[0120] See Figure 8 The figure shows the load switching waveforms of the control method of this invention compared to single-voltage-loop NNC control and dual-closed-loop PI control. When the load switches from 180Ω to 360Ω, the voltage stabilization time is 30ms and the current is 45ms under single-voltage-loop NNC control, with a voltage overshoot of 28V and a current undershoot of 0.4A. Figure 8As shown in (a). Under dual closed-loop PI control, the voltage settling time is 46ms with an overshoot of 20V, the current settling time is 44ms with an undershoot of 1.4A, as follows. Figure 8 As shown in (b). Under dual closed-loop NNC control, the voltage settling time is 27ms, the overshoot is 12.5V, the current settling time is 20ms, and the current undershoot is 0.8A. Figure 8 As shown in (c), the dual-loop NNC has the fastest dynamic response, and its voltage overshoot and undershoot are lower than the other two control schemes.

[0121] In summary, the optimal control method for DC-DC converters based on adaptive dynamic programming and data-driven principles of this invention can achieve parameter updates for the neural network controller without establishing an accurate model of the DC-DC converter. Compared with traditional control methods, it has advantages such as fast control speed, good dynamic performance, and independence from system models.

[0122] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data-driven based high gain DC-DC converter dual-loop model-free control method, characterized in that, The application comprises the following steps: S1, constructing a neural network controller, using the neural network controller to control the output voltage of the high-gain DC-DC converter; the neural network controller is composed of a current inner loop neural network controller I-NNC and a voltage outer loop neural network controller V-NNC, the input of the voltage outer loop neural network controller is a voltage error , a voltage integral term and a voltage differential term , and the output is a current reference value ; the input of the current inner loop neural network controller is a current error , a current integral term and a current differential term , and the output is a duty cycle ; the network of the voltage outer loop neural network controller is a single hidden layer, the hidden layer includes three neurons, the hidden layer activation function is a hyperbolic tangent function, and the output layer activation function is a Purelin function; the network of the current inner loop neural network controller is a single hidden layer, the hidden layer includes three neurons, the hidden layer activation function is a hyperbolic tangent function, and the output layer activation function is a sigmoid function; S2, construct and train a high-gain DC-DC converter twin model of a nonlinear autoregressive external input (NARX) model with external input; the NARX model has three hidden layers, each hidden layer includes three neurons, the hidden layer activation function is a hyperbolic tangent function, and the output layer activation function is a Purelin function; the external input data of the NARX model is the duty cycle , and the output is the predicted value of the output voltage and the input current and ; S3, based on the neural network controller, a closed-loop circuit of the converter is built on Simulink, the I-NNC and V-NNC parameters are updated by using an adaptive dynamic programming (ADP) algorithm, the I-NNC and V-NNC performance evaluation index functions are set, and the I-NNC and V-NNC parameters are initialized; the parameters include weights and biases; S4, the parameters of the V-NNC are kept fixed, the weights and biases of the I-NNC are iterated based on the performance evaluation index function of the I-NNC by using a batch gradient descent method; S5, the weights and biases of the I-NNC are updated according to the iteration results, and steps S4 to S5 are repeated by selecting multiple initialization points until the performance evaluation index function value calculated based on the performance evaluation index function of the I-NNC reaches a preset target value or the iteration reaches a given iteration number, thereby obtaining an optimal current inner loop neural network controller; S6, the weights and biases of the optimal current inner loop neural network controller are kept fixed, the weights and biases of the V-NNC are iterated based on the performance evaluation index function of the V-NNC by using a batch gradient descent method; S7, the weights and biases of the V-NNC are updated according to the iteration results, and steps S6 to S7 are repeated by selecting multiple initialization points until the performance evaluation index function value based on the performance evaluation index function of the V-NNC reaches a preset target value or the iteration reaches a given iteration number, thereby obtaining an optimal voltage outer loop neural network controller; S8, the weights and biases of the optimal current inner loop neural network controller and the optimal voltage outer loop neural network controller are output, and the DC-DC converter is controlled based on the parameters of the optimal neural network controller.

2. The data-driven based high gain DC-DC converter dual-loop model-free control method according to claim 1, wherein, The steps of constructing and training the nonlinear autoregressive exogenous (NARX) model with external input are as follows: The NARX model is constructed, and the duty cycle The output voltage prediction value And the input current prediction value As output; while the NARX model output is delayed and also used as the input of the model itself, and the input-output relationship is represented as: ; Wherein d(k) and d(k-n) represent the duty cycle values of the controller at k and k-n time respectively; And And represent the output voltage values of the converter at k and k-m time respectively; And And represent the input current values of the converter at k and k-m time respectively, and m and n represent the delay step length respectively; Represent the nonlinear mapping relationship of the NARX model; the NARX model is set to have 3 hidden layers, the hidden layer includes 3 neurons, the hidden layer activation function is the hyperbolic tangent function, and the output layer activation function is the linear function; A physical model of the converter circuit is built, the circuit is operated in an open loop under the conditions of a variable input duty ratio and an output load, the duty ratio, output voltage and input current data are collected, and the collected data are used to train a high-gain DC-DC converter twin model; The NARX model is trained by using a Levenberg-Marquardt algorithm to minimize the mean square error; the closed-loop NARX model is converted, and the best model in the verification result is selected as the final converter twin model.

3. The data-driven based high gain DC-DC converter dual-loop model-free control method according to claim 1, wherein, The performance evaluation index functions of the current inner loop neural network controller (I-NNC) and the voltage outer loop neural network controller (V-NNC) are represented as follows: Performance evaluation index function of I-NNC is represented as follows: ； wherein i represents the iteration number; Performance evaluation index function of V-NNC is represented as follows: ； wherein, Ik represents the input current sampled at k, Ik represents the reference current at k, Ik represents the output voltage sampled at k; Ik represents the reference voltage at k; M represents the total number of input current and output voltage sampling points.

4. The data-driven based high-gain DC-DC converter dual-loop model-free control method according to claim 3, wherein, The weights and biases of the neural network controller are iterated by using a batch gradient descent method, and the iteration formula is represented as follows: ； wherein, denotes the weights of the neural network at the i-th iteration; denotes the weights of the neural network at the i-th iteration; denotes the weights of the neural network at the i-th iteration; is a learning rate, typically between 0 and 1 ; denotes the value of the performance evaluation indicator function at the i-th iteration; denotes the gradient of the performance evaluation indicator function with respect to the weights at the i-th iteration; denotes the biases of the neural network at the i-th iteration; denotes the biases of the neural network at the i-th iteration; denotes the biases of the neural network at the i-th iteration; denotes the gradient of the performance evaluation indicator function with respect to the biases at the i-th iteration.

5. The data-driven based high-gain DC-DC converter dual-loop model-free control method according to claim 4, wherein, The gradient value of the performance evaluation index function of the I-NNC on the weights and the gradient value of the performance evaluation index function of the I-NNC on the biases at the i-th iteration of the I-NNC are represented as follows: ； wherein ； In the formula, represents the total number of input current sampling points; represents the input current sampled at k time; represents the duty cycle at k time; represents the input-output mapping relationship of the feedforward neural network; The gradient value of the performance evaluation index function of the V-NNC on the weights and the gradient value of the performance evaluation index function of the V-NNC on the biases at the i-th iteration of the V-NNC are represented as follows: ； wherein N represents the total number of output voltage sampling points; represents the output voltage sampled at time k; represents the duty cycle at time k; represents the reference current at time k; represents the reference current at time k and the input current difference.

6. The data-driven based high-gain DC-DC converter dual-loop model-free control method according to claim 4, wherein, and is the cumulative value of the batch data gradient, not the gradient value of a certain data.

Images