Neural network based on-line control method for single-switch dc-dc converter

Abstract

The method is a neural network-based online control method for a single-switch DC-DC converter, comprising: constructing a neural network model comprising a neural network training structure, a neural network training mode and a neural network training algorithm; constructing a neural network controller with the single-switch DC-DC converter as a control object and with a converter output voltage as a control target; collecting running state data and control data of the single-switch DC-DC converter in real time, obtaining gradients online, and realizing online training of the neural network and online control of the converter. The method does not need to model the converter, nor does it need a large number of data samples to perform offline training of the neural network, but realizes online acquisition of a control law (weights and biases) in the running process of the converter, achieves the purpose of real-time control, and the neural network controller can cope with the influence of step changes in input voltage or output load, has good robustness and dynamic response performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of converter control, in particular to an online control method and system of a single-switch DC-DC converter based on a neural network. BACKGROUND

[0002] DC-DC converters are the core components of power conversion technology in modern new energy industry applications. Among them, single-switch DC-DC converters are widely used in photovoltaic power generation, wind power generation, new energy electric vehicles, etc. due to their small size, light weight, simple structure, high efficiency, etc. At present, single-switch DC-DC converters are commonly used in the terminal load power supply interface of power conversion systems, and their function is to convert a specific DC voltage into a DC voltage required by the user end, and their performance directly determines the power supply indicators of the load end. In communication power supplies, the randomness of the load of microprocessors and communication devices fluctuates widely, which puts forward higher requirements for the control strategy of the converter.

[0003] The traditional control method of single-switch DC-DC converters is to model and analyze them using small signal average models, and then obtain the target switching signal using linear feedback control methods. This control method is based on known models, and the design of the control parameters of the controller is complex. At the same time, in actual applications, the output response speed is slow, and only a certain range of disturbances can exhibit good control performance. Seeking a more effective control method for single-switch DC-DC converters has important practical significance for improving the real-time performance and stability of DC voltage stabilization systems and improving energy conversion efficiency.

[0004] In recent years, neural network technology has been widely concerned. It does not need to establish an accurate mathematical model, only needs to set the number of hidden layer neurons and activation function reasonably, and can approximate any actual function. Therefore, it has certain advantages in dealing with the nonlinearity and uncertainty of control system. In the paper "Design of Single Switch DC-DC Converter Based on BP Neural Network" (DOI:10.16628 / j.cnki.2095-8188.2015.23.011) published in 2015, it is proposed to use BP neural network to identify the offline model of single switch DC-DC converter and use it for control, which verifies the feasibility of neural network model as a controller. However, the problem is that neural network needs a large amount of experimental data for offline training, although it has strong generalization ability, but the algorithm takes more time and has weak universality. In the patent "Single Switch DC-DC Converter Output Voltage Control Method Based on ELM-PID" (Patent No: CN111082660A), the trained ELM output is used to adaptively adjust the PID parameters, and then sent to the PID controller to realize the control of the converter. Due to its special algorithm design, the system has fast dynamic response speed. However, it still needs to model the converter, and the ELM needs to be trained under the condition that the model is known. How to design an online model-free neural network controller is a problem to be solved. SUMMARY

[0005] To solve the above problems, the present application proposes an online control method and system for single switch DC-DC converter based on neural network. The designed neural network controller can be trained online, and at the same time, it can control the single switch DC-DC converter in real time. It has good robustness and dynamic performance for step changes of input reference voltage and output load.

[0006] In one aspect, an online control method for single switch DC-DC converter based on neural network, comprising:

[0007] Constructing a neural network model including neural network training structure, neural network training method and neural network training algorithm;

[0008] Constructing a neural network controller taking single switch DC-DC converter as control object and taking converter output voltage as control target; the neural network controller includes neural network model and loss function;

[0009] Real-time acquisition of single switch DC-DC converter running state data and control data, online acquisition of gradient, realization of online training of neural network and online control of converter.

[0010] Preferably, the neural network training structure adopts a feedforward neural network structure; the neural network training method adopts a supervised learning method, continuously adjusts the weight and bias parameters, and completes the learning process; the neural network training algorithm adopts a back propagation algorithm, and the gradient descent method is used to perform multiple compound operations on the nonlinear function to reproduce the input-output mapping relationship.

[0011] Preferably, the neural network model is a single-layer neuron, and only includes an input layer and an output layer; the input layer is composed of three neurons, and the input signals of the three neurons are input voltage v in , output voltage v o , and transformer inductor current i L ; the output layer is composed of a single neuron, and the output d of the single neuron is used as a PWM modulation signal of the single-switch DC-DC converter.

[0012] Preferably, the input layer and the output layer are connected to each other, and the mapping relationship between the neurons in the layers is as follows:

[0013]

[0014] d(x,ω,b)=S(f(x))

[0015] wherein I represents the number of input layer neurons, ω i represents the weight of the i th neuron of the input layer mapped to the output layer neuron, x i represents the i th input of the input layer, b represents the bias of the neuron of the input layer mapped to the output layer neuron, f(x) is the weighted sum of the inputs, and S represents the output layer activation function, e -f(x) represents the-f(x) th power of e.

[0016] Preferably, the loss function adopts a mean square error, which is used to represent the difference between the predicted value and the true value, and the neural network model parameters are improved by reducing the loss function value, and the loss function is represented as follows:

[0017]

[0018] wherein N is the number of samples, y i is the predicted value, and Y i is the true value; according to the actual control condition, the predicted value is the output voltage v o of the single-switch DC-DC converter, and the true value is the output reference voltage v ref ; in online training, each sampling time corresponds to an output predicted value, that is, the number of samples N is 1.

[0019] Preferably, the BP algorithm is used to obtain the online gradient; according to the gradient descent method and the chain rule, the online gradient is represented as follows:

[0020]

[0021] in, and △b k The weights ω are respectively i The change in bias b, where the superscript k and the k in parentheses represent the iterative sampled or calculated value at time k, η is the learning rate; L(k) represents the loss function; v o (k) represents the output voltage at time k; d(k) represents the PWM modulation signal at time k; f k (x) represents the weighted sum of all inputs at time k; b represents the weights that map the i-th neuron in the input layer to the neuron in the output layer at time k-1; k-1 This represents the bias that maps the input layer neurons to the output layer neurons at time k-1;

[0022] The online gradient acquisition method is as follows: at each sampling time, the required transformer parameters are collected and fed into the above formula for calculation, where, The partial derivatives are calculated directly, as follows:

[0023]

[0024]

[0025] Among them, v ref (k) represents the output reference voltage at time k; v o (k-1) represents the output voltage at time k-1; d(k-1) represents the PWM modulation signal at time k-1; This represents the input signal of the i-th neuron in the input layer at time k.

[0026] Preferably, during online training, a value-based truncation method is used to limit... To control the output range and prevent gradient explosion during iteration, it is expressed as follows:

[0027]

[0028] In the above formula, let For g k Given the interval [+v] in -v in At time k, if gk is less than -v in Set it to -v in If g k Big + V in Set it to +v in .

[0029] The weight ω i The kth iteration process of the weight ω

[0030]

[0031] wherein, ωki represents the weight of the kth iteration process of the input layer i neuron mapping to the output layer neuron; bki represents the bias of the kth iteration process of the input layer i neuron mapping to the output layer neuron. k ωki represents the weight of the kth iteration process of the input layer i neuron mapping to the output layer neuron; bki represents the bias of the kth iteration process of the input layer i neuron mapping to the output layer neuron.

[0032] Preferably, after each iteration, the weight ω i and the bias b are substituted into The new output d of the neural network is calculated, and the next iteration cycle is entered; when the loss function is less than the allowable value, the iteration of the weight and the bias is stopped, which indicates that the online training of the neural network is completed, the single-switch DC-DC converter reaches the target output voltage, and the online control of the system is realized.

[0033] On the other hand, a single-switch DC-DC converter online control system based on a neural network comprises:

[0034] A neural network model construction module is configured to construct a neural network model comprising a neural network training structure, a neural network training method and a neural network training algorithm.

[0035] A neural network controller construction module is configured to construct a neural network controller taking a single-switch DC-DC converter as a control object and taking a converter output voltage as a control target; the neural network controller comprises a neural network model and a loss function.

[0036] An online training module is configured to collect running state data and control data of the single-switch DC-DC converter in real time, to obtain gradients online, and to realize online training of the neural network and online control of the converter.

[0037] The present application has the following innovative points and advantages compared with the prior art:

[0038] The present application does not need to model the single-switch DC-DC converter, and does not need to perform offline training of the neural network with a large number of data samples, but realizes online acquisition of a control law (including a weight and a bias) in the running process of the converter, so as to achieve real-time control, and therefore has strong adaptability; the neural network controller of the present application can cope with the influence of input voltage or output load fluctuation, and has good robustness and dynamic response performance. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1Flow chart of the neural network-based online control method of the single-switch DC-DC converter according to the embodiment of the present application;

[0040] Figure 2 Structure block diagram of the neural network-based online control method of the single-switch DC-DC converter according to the embodiment of the present application;

[0041] Figure 3 Detailed flow chart of the neural network-based online control method of the single-switch DC-DC converter according to the embodiment of the present application;

[0042] Figure 4 Simulink simulation schematic diagram of the neural network-based online control system of the Buck converter according to the embodiment of the present application;

[0043] Figure 5 Simulink simulation output voltage waveform diagram of the neural network-based online control system of the Buck converter according to the embodiment of the present application when the input voltage or load jumps;

[0044] Figure 6 Structure block diagram of the neural network-based online control system of the single-switch DC-DC converter according to the embodiment of the present application. DETAILED DESCRIPTION

[0045] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] Referring to Figure 1 The neural network-based online control method of the single-switch DC-DC converter according to the present application comprises the following steps:

[0047] S101, constructing a neural network model comprising a neural network training structure, a neural network training mode and a neural network training algorithm;

[0048] S102, constructing a neural network controller taking the single-switch DC-DC converter as a control object and taking the converter output voltage as a control target;

[0049] S103, collecting the running state data and the control data of the single-switch DC-DC converter in real time, obtaining the gradient online, realizing the online training of the neural network and the online control of the converter.

[0050] Specifically, referring to Figure 2 The neural network controller comprises a neural network model and a loss function. The neural network model adopts a single neuron structure and only contains an input layer and an output layer. The network input layer is composed of three neurons, which are the input voltage v in , the actual output voltage v o and the converter inductor current i LThe network output layer consists of a single neuron, whose output d serves as the PWM modulation signal for a single-switch DC-DC converter.

[0051] In this embodiment, the input layer and output layer of the neural network model are interconnected, and the mapping relationship between neurons in the layers is as follows:

[0052]

[0053] d(x,ω,b)=S(f(x))

[0054] Where I is the number of neurons in the input layer, ω i x represents the weights that map the i-th neuron in the input layer to the neurons in the output layer. i Let represent the i-th input of the input layer, and b represent the bias of the input layer neuron mapped to the output layer neuron. f(x) is the weighted sum of all inputs, and S represents the output layer activation function; here, the unipolar sigmoid function is chosen. As the activation function. In this embodiment, I is 3.

[0055] The loss function is used to represent the difference between the predicted value and the true value of the neural network output. Reducing the value of the loss function improves the parameters of the neural network model. The control method of this invention uses the mean squared difference as the loss function, which can be expressed as:

[0056]

[0057] Where N is the number of samples, y i Y is the predicted value. i This is the true value. Based on the neural network model structure set here, the network predicts the output voltage v of the single-switch DC-DC converter system. o The actual value is the output reference voltage V. ref In online training, each sampling time corresponds to one output prediction value, so the number of samples N is 1 here.

[0058] See Figure 3 The diagram shows the online control algorithm for a single-switch DC-DC converter based on a neural network, similar to traditional control methods. The neural network controller processes the converter's state variables at time k. Sampling is performed, and the sampled values ​​are fed into the network for calculation. The network output d at that moment is used as the PWM modulation signal for the next moment to control the single-switch DC-DC converter.

[0059] In this embodiment, the loss function evaluates the neural network output at time k, and adjusts the network weights and biases for the next time neural network calculation. This process is essentially online training of neural networks, when the iterative gradient of the weight and bias is small enough, it means that the neural network has completed the approximation of the system dynamic model, and also realizes the control of the system on the converter output voltage.

[0060] The BP algorithm is used for online training, which is essentially solving the optimal parameters under the minimum loss function by gradient descent method and chain rule, so as to realize the iteration of weights and bias b. It can be expressed as:

[0061]

[0062] Among them, and△b k are the change of weight ω i and bias b, superscript k and bracket k represent the iterative sampling value or calculation value at time k, and η is the learning rate. The gradient is obtained online as follows: at each sampling time, the required converter parameters are collected and sent to the above formula for calculation. Four parts of information can be directly calculated by partial derivative, which can be expressed as:

[0063]

[0064] Among them, is the output voltage v o of the single-switch DC-DC converter in the controlled process. The first-order partial derivative of the neural network output d is calculated. If the single-switch DC-DC converter is modeled by using the traditional control method, it can be considered that and the input voltage v in are equivalent at steady state. However, the single-switch DC-DC converter is not modeled in the present invention, and the single-switch DC-DC converter is not in steady state during the control process, so the partial derivative cannot be calculated directly. Here, the difference quotient is used instead of the differential quotient, and the difference between v o and the difference between d are used to approximate this part and calculate The calculation method of

[0065]

[0066] It should be noted that during the neural network training process, the above formula may have a particularly large value at some time. The value truncation method is used to limit the output range to prevent gradient explosion during iteration and cause system instability. It can be expressed as:

[0067]

[0068] In the above formula, let For g k Given the interval [v in -v in At that moment, if g k Less than -v in Set it to -v in If g k Greater than v in Let it be v. in .

[0069] Ultimately, the weight ω i The k-th iteration of bias b can be represented as:

[0070]

[0071] After each iteration, the new output of the neural network is calculated, and the next iteration cycle begins. When the loss function is less than the allowable value Tol, the iteration of weights and biases stops, indicating that the online training of the neural network is complete. This also means that the controlled converter has reached the target output voltage, and the online control of the system has been achieved.

[0072] To verify the actual control effect of the present invention, this embodiment uses a Buck converter as an example for simulation experiments. See [link to relevant documentation]. Figure 4 The diagram shows the Simulink simulation schematic of the online neural network control system for a Buck converter. The solid-lined box represents the main circuit of the Buck converter, and the dashed-lined box represents the control circuit. The area within the right box is used to adjust the output load R of the single-switch DC-DC converter, and the area within the left box is used to adjust the input voltage v of the single-switch DC-DC converter. in .

[0073] In the main circuit design, the initial input voltage v was established. in A Buck converter simulation circuit with a voltage of 10V, an initial load R of 5Ω, an inductor L of 470uH, a capacitor C of 680uF, and a switching frequency of 50kHz is used. To verify the dynamic characteristics of the control strategy of this invention, v ref The controller was input with a square wave pulse of 10Hz, a high level of 8V, a low level of 5V, and a frequency of 10V. The simulation experiment time was set to 0.5s, and all step changes were set at 0.25s.

[0074] In the control section design, a three-input, single-output neural network model subsystem was built to generate PWM modulation signals. Initial weights were set. The initial bias values ​​are 2, -1, and -1, respectively, with initial bias value b. 0 The value is 1, and the learning rate is set to 0.003.

[0075] Referring to Figure 5 The Buck converter neural network online control system shown in the Simulink simulation output voltage waveform diagram under the input voltage or load jump condition. Among them, the square wave is the reference voltage waveform, and the other part is the output voltage waveform. The output voltage waveform in the figure can better track the change of the reference voltage. Figure 5 (a) is the output voltage waveform for input voltage step change, input voltage v in From 10v to 20v; Figure 5 (b) is the output voltage waveform for load value step change, and the output load R changes from 5Ω to 2.5Ω. It can be seen that no matter which step change, the output voltage waveform at this moment can recover to the original reference voltage value after a short fluctuation.

[0076] In summary, the online control method of the single-switch DC-DC converter based on the neural network does not need to model the single-switch DC-DC converter. At the same time, the proposed control method does not need a large number of data samples to train the neural network offline, but realizes online acquisition of the control law (including the weight and bias) in the converter running process, so as to achieve real-time control. In addition, the proposed neural network controller can cope with the influence of input voltage or output load fluctuation, has good robustness and dynamic response performance.

[0077] Referring to Figure 6 As shown in the figure, according to another aspect of the present application, a single-switch DC-DC converter online control system based on neural network includes:

[0078] The neural network model construction module 601 is used to construct the neural network model including the neural network training structure, the neural network training method and the neural network training algorithm;

[0079] The neural network controller construction module 602 is used to construct the neural network controller taking the single-switch DC-DC converter as the control object and taking the converter output voltage as the control target; the neural network controller includes the neural network model and the loss function;

[0080] The online training module 603 is used to collect the running state data and the control data of the single-switch DC-DC converter in real time, to acquire the gradient online, to realize the online training of the neural network and the online control of the converter.

[0081] The specific implementation of the single-switch DC-DC converter online control system based on neural network is the same as the single-switch DC-DC converter online control method based on neural network. The embodiment will not be repeated.

[0082] The above examples are only for illustrating the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made according to the spirit and essence of the present application shall be covered within the protection scope of the present application.

Claims

1. An online control method for a single-switch DC-DC converter based on a neural network, characterized in that, include: Construct a neural network model that includes the neural network training structure, neural network training method, and neural network training algorithm; A neural network controller is constructed, which uses a single-switch DC-DC converter as the controlled object and the converter output voltage as the control objective; the neural network controller includes a neural network model and a loss function. Real-time acquisition of operating status and control data of a single-switch DC-DC converter, online gradient acquisition, and online training of neural network and online control of converter; The neural network training structure adopts a feedforward neural network structure; the neural network training method adopts a supervised learning method, continuously adjusting the weights and bias parameters to complete the learning process; the neural network training algorithm adopts the backpropagation algorithm, which uses gradient descent to perform multiple composite operations on the nonlinear function to reproduce the input-output mapping relationship. The neural network model is a single-layer neuron, containing only an input layer and an output layer; the input layer consists of three neurons, and the input signals of the three neurons are the input voltages. Output voltage and converter inductor current The output layer is composed of a single neuron, and the output d of a single neuron serves as the PWM modulation signal for a single-switch DC-DC converter. The input layer and the output layer are interconnected, and the mapping relationship between neurons in the layers is as follows: ； Where I represents the number of neurons in the input layer. This represents the weights that map the i-th neuron in the input layer to the neurons in the output layer. Let represent the i-th input of the input layer, b represent the bias of the input layer neuron mapping to the output layer neuron, f(x) be the weighted sum of all inputs, and S represent the output layer activation function. , express -f(x) to the power of; The backpropagation (BP) algorithm is used for online gradient acquisition; based on gradient descent and the chain rule, it is expressed as follows: ； in, and The weights are respectively The change in bias b, where the superscript k and the k in parentheses represent the iterative sampled value or calculated value at time k. For learning rate; Represents the loss function; This represents the output voltage at time k; This represents the PWM modulation signal at time k; This represents the weighted sum of all inputs at time k; This represents the weights of the i-th neuron in the input layer mapped to the neuron in the output layer at time k-1; This represents the bias that maps the input layer neurons to the output layer neurons at time k-1; The online gradient acquisition method is as follows: at each sampling time, the required transformer parameters are collected and fed into the above formula for calculation, where, , , and The partial derivatives are calculated directly, as follows: ； ； in, This represents the output reference voltage at time k; This represents the output voltage at time k-1; This represents the PWM modulation signal at time k-1; This represents the input signal of the i-th neuron in the input layer at time k.

2. The online control method for a single-switch DC-DC converter based on a neural network according to claim 1, characterized in that, The loss function uses the mean squared error to represent the difference between the predicted and actual values. The neural network model parameters are improved by reducing the value of the loss function, as shown below: ； Where N is the number of samples, For predicted values, This is the actual value; based on the actual control conditions, the predicted value is the output voltage of the single-switch DC-DC converter. The actual value is the output reference voltage. In online training, each sampling time corresponds to an output prediction value, that is, the number of samples N is 1.

3. The online control method for a single-switch DC-DC converter based on a neural network according to claim 1, characterized in that, During online training, a value-based truncation method is used to limit... To control the output range and prevent gradient explosion during iteration, it is expressed as follows: ； In the above formula, let for Given an interval [ , At time k, if Less than Set it to ;if Greater than Set it to .

4. The online control method for a single-switch DC-DC converter based on a neural network according to claim 1, characterized in that, weight The k-th iteration of bias b is represented as follows: ； in, This represents the weights of the i-th neuron in the input layer mapped to the neuron in the output layer at time k; This represents the bias at time k, where the input layer neurons are mapped to the output layer neurons.

5. The online control method for a single-switch DC-DC converter based on a neural network according to claim 4, characterized in that, After each iteration, the weights are... And the bias b is substituted. The new output d of the neural network is calculated, and the next iteration cycle begins. When the loss function is less than the allowable value, the iteration of weights and biases stops, indicating that the online training of the neural network is complete, the controlled single-switch DC-DC converter has reached the target output voltage, and the online control of the system has been realized.

6. An online control system for a single-switch DC-DC converter based on a neural network, characterized in that, The method based on any one of claims 1 to 5 includes: The neural network model building module is used to build neural network models, including neural network training structures, neural network training methods, and neural network training algorithms. A neural network controller construction module is used to construct a neural network controller that uses a single-switch DC-DC converter as the controlled object and the converter's output voltage as the control objective; the neural network controller includes a neural network model and a loss function; The online training module is used to collect real-time operating status and control data of the single-switch DC-DC converter, acquire gradients online, and realize online training of neural networks and online control of the converter.

Images