Power converter control method and equipment based on physical information machine learning, and storage medium

Abstract

The invention relates to the technical field of power converters, in particular to a power converter control method and device based on physical information machine learning and a storage medium. Comprising the following steps: setting up a control system; collecting, storing and preprocessing data; building and evaluating a model; model deployment; applying the model; and selecting a prediction model of the operation mode of the power converter. According to the design, physical constraint information is embedded into a loss function, a machine learning algorithm is used for mining the relation between the duty ratio, the frequency and the number of hopping pulses of a control signal and input voltage, input current, output voltage, output current and temperature, and a machine learning control model is deployed on an FPGA. The duty ratio, the frequency and the pulse hopping number of the control signal are predicted and controlled in real time; accurate modeling is performed through the nonlinear capability of a machine learning model, so that a large amount of cost can be saved, and the accuracy of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power converters, and in particular to a power converter control method, device and storage medium based on physical information machine learning. Background Art

[0002] Current power converter control technologies are mainly divided into two categories: analog control technology and digital PID control technology. Analog control technology includes voltage mode control and current mode control, while digital PID control technology can support more complex algorithms, such as nonlinear control and adaptive control. Traditional switching power supplies usually use analog control technology, using components such as comparators, error amplifiers, and analog power management chips to adjust the output voltage. However, this control method has many shortcomings, such as complex control circuits, a large number of components required, and difficulty in modifying the control circuit once it is designed.

[0003] With the rapid development of microelectronics technology, power control technology has evolved from pure analog control to analog-digital hybrid control, and then to the current full-digital PID control. Digital PID control not only simplifies the control circuit and reduces the number of components, but also provides higher flexibility, which is convenient for later adjustment and optimization, and greatly promotes the miniaturization and integration of switching power supplies. The nonlinear characteristics of switching power supply control are caused by multiple nonlinear characteristics such as magnetic components (such as transformers and inductors), load changes, parasitic capacitance and parasitic diodes of switching devices, nonlinear components in the feedback loop (such as the saturation characteristics of the error amplifier), and temperature changes (such as changes in the temperature of the heat sink, transformer, and the internal temperature of the device). In addition, the switching action of the switching device (such as MOSFET or IGBT) at the core of the switching power supply is also nonlinear. These nonlinear characteristics work together to make the power converter a highly nonlinear system, so a nonlinear control strategy is needed to ensure the stability of the output voltage.

[0004] The technical difficulties in the design of the feedback loop of the power converter mainly include the design of the compensation network, the suppression of noise, the response to temperature and environmental influences, the processing of multi-loop control, the stability of the system, the optimization of response speed and dynamic performance, and the adaptation to load changes and nonlinear characteristics. These difficulties require designers to have deep professional knowledge and rich practical experience. Traditional linear control methods, such as adjusting the duty cycle of PWM, the frequency of PFM or the number of pulse jumps of PSM according to the change of output voltage through analog control technology and digital PID control technology, may not fully meet the requirements of high precision and stability. In addition, the high nonlinearity of the power converter makes it very difficult to establish an accurate mathematical model. Due to the complexity of system modeling, the dynamic adjustment requirements of parameter setting, the nonlinear compensation of control strategy design, and the difficulty of implementation and debugging, the design of nonlinear PID controller is a challenge. In view of this, we propose a power converter control method, device and storage medium based on physical information machine learning. Summary of the invention

[0005] The purpose of the present invention is to provide a power converter control method, device and storage medium based on physical information machine learning to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, one of the objectives of the present invention is to provide a power converter control method based on physical information machine learning, comprising the following steps:

[0007] S1. Control system construction: Build a control system architecture including power converter, data acquisition and processing module, machine learning control algorithm and control signal generator;

[0008] S2, data collection, storage and preprocessing: collect physical information and control signal data of power converter operation, splice and store the collected data to facilitate the training and application of machine learning algorithms, and clean and normalize the data;

[0009] S3. Model building and evaluation: Use machine learning to build an accurate mathematical control model related to the operation of the power converter, and evaluate and optimize the model based on the loss function;

[0010] S4, model deployment: optimize and compress the model to reduce resource usage and improve inference speed, compile the trained model into a format suitable for MCU, ASIC and FPGA, and deploy it on the embedded system;

[0011] S5, model application: inputting the real-time measured power converter operating parameters into the corresponding prediction model for real-time prediction;

[0012] S6. Selection of a prediction model for the power converter operation mode: intelligently selecting the power converter operation model based on the operation efficiency and prediction efficiency of the power converter during operation.

[0013] As a further improvement of this technical solution, in the control system architecture constructed by S1:

[0014] Power converters include at least Buck, Boost, Buck-Boost, CUK and resonant converters, as well as flyback converters, forward converters, push-pull converters, half-bridge converters and full-bridge converters, and parallel topology architectures of these types of converters, and power converters that use ZVS and ZCS technologies to reduce switching losses of power devices;

[0015] The data acquisition and processing module includes the acquisition of the control signal duty cycle, frequency and pulse jump number, the acquisition of input voltage, input current, output voltage, output current and temperature, and A / D conversion; at the same time, it is also necessary to store the maximum and minimum data used for the denormalization process to convert the normalized data back to the scale of the original data;

[0016] The machine learning control algorithm includes at least SVR, RF, BP neural network, RBF and LSTM; the prediction model selection algorithm of the power converter can also adopt these machine learning algorithms, and the prediction model selection algorithm selects whether the topology of the power converter is operated in PWM mode or in PFM / PSM mode according to the operating efficiency and predicted efficiency of the power converter;

[0017] The control signal generator generates (pulse width modulation) PWM signals, (pulse frequency modulation) PFM signals and (pulse step modulation) PSM signals according to the duty cycle, frequency, number of pulse jumps, and dead time predicted by the machine learning algorithm, and controls the switching of power MOS tubes or IGBT power devices through the driving circuit.

[0018] As a further improvement of the technical solution, the specific process of S2 includes the following steps:

[0019] S2.1. Data collection: Collect the input and output voltage and current of the power converter and the environmental parameters, including input voltage, input current, output voltage, output current, and temperature data of the heat sink, transformer, and air, as features for machine learning; collect control signal data, including control signal duty cycle, frequency, and pulse jump number data, as label data for machine learning;

[0020] S2.2, data storage: sort the power converter data according to the acquisition time, use PWM duty cycle and PFM frequency as labels, and use input voltage, input current, output voltage, output current and temperature characteristics as model training data to form a data set;

[0021] S2.3. Data preprocessing: fill missing values ​​in the collected power converter data and normalize the data set to accelerate model convergence.

[0022] As a further improvement of the technical solution, in S2.3, the data preprocessing specifically includes:

[0023] If there are some missing values ​​in the data set, we can use the data from the previous time point to supplement them, or use the weighted average of the previous and next data to supplement them;

[0024] In the case of standard dimensions, the input voltage, input current, output voltage, output current and temperature characteristic data of the power converter, as well as the duty cycle, frequency, and pulse jump number label data have large data differences. The data is preprocessed by normalizing the maximum and minimum values, and the features and labels of different scales are scaled to In order to eliminate the influence of large and small data on the model, the normalization formula is:

[0025]

[0026] In the formula, is the data to be normalized, i.e. the original data, is the normalized data, and They represent the maximum and minimum values ​​in the data to be normalized respectively. Finally, the data is mapped between 0 and 1 and input into the model as feature data for training. The processed data is divided into training set and test set in a ratio of 7:3, where the first 70% of the data set is used as the training set and the last 30% as the test set.

[0027] As a further improvement of the present technical solution, the specific process of S3 includes the following steps:

[0028] S3.1. Model building and training: Construct a machine learning control model using SVR, RF, RBF, and BP neural network machine learning algorithms, and use deep learning algorithms such as LSTM for complex situations; use machine learning as a control method, and use the nonlinear mapping ability of machine learning to learn the relationship between input voltage, input current, temperature, output voltage, output current and pulse width, frequency, and number of jump pulses;

[0029] S3.2. Model evaluation: Apply machine learning to the design of power converter control systems, and introduce physical constraint losses into the loss function of the machine learning model to reflect the penalty terms of physical laws or system characteristics; design a loss function that includes label loss, consideration of the error between the output voltage and the expected voltage, and efficiency loss terms to ensure that the output voltage remains within a fixed error range; the design of physical constraint losses aims to ensure that the behavior predicted by the model is consistent with the behavior of the actual physical system.

[0030] As a further improvement of the technical solution, in S3.2, when machine learning predicts the duty cycle, frequency and number of pulse jumps, the labels are the duty cycle, frequency and number of pulse jumps, and the loss function is defined according to the label loss, output voltage loss and efficiency loss;

[0031] The mean square error (MSE) or absolute error (MAE) is used to measure the error between the output label and the actual label. The fluctuation of the output voltage, the error between the output voltage and the expected voltage, and the efficiency are introduced as constraints. The constraints are used as part of the loss function, that is, a penalty term related to the constraints is added to the original loss function. If the model violates the constraints, the penalty term will increase the loss value, thereby encouraging the model to adjust to meet the constraints. The penalty term of the loss function is used to guide the model to learn in the direction of meeting the conditions. Among them:

[0032] The mean square error (MSE) is used to measure the prediction error of the label. :

[0033]

[0034] in, It is The predicted pulse width of samples, It is The actual pulse width of samples, is the sample size;

[0035] In order to ensure the stability of the output voltage, additional constraints are introduced on the basis of the basic loss function, namely, output voltage fluctuation, output voltage deviation and efficiency constraints, and three penalty terms are introduced accordingly, namely, fluctuation penalty term, voltage deviation penalty term and efficiency penalty term;

[0036] Define a penalty term to measure the fluctuation of output voltage:

[0037]

[0038] in, is the fluctuation function of the output voltage, i.e., the fluctuation penalty, It is Output voltage at a time point; It is Output voltage at a time point;

[0039] Define a penalty term to measure the deviation between the output voltage and the target voltage; use the mean square error (MSE) to measure the error between the output voltage and the target voltage , i.e., the deviation penalty, as a constraint on the output voltage:

[0040]

[0041] in, is the target output voltage, is the output voltage;

[0042] Introducing efficiency constraints helps machine learning models better understand and predict the behavior of the system:

[0043]

[0044] in, is the output power efficiency constraint function, i.e., power penalty; , They are The input voltage and input current at a time point, , They are Output voltage and output current at a time point;

[0045] The basic loss function and the stability constraint are combined to form a comprehensive loss function:

[0046]

[0047] in, It is The actual output voltage of a sample is a normalized value; is the expected output voltage, which is a normalized value and a fixed value;

[0048] in, is a hyperparameter that controls the weight of the volatility penalty; is a hyperparameter that controls the weight of the deviation penalty; is a hyperparameter that controls the weight of the power penalty.

[0049] As a further improvement of the technical solution, in S5, the real-time measured input voltage, input current, output voltage, output current and temperature are input into the duty cycle, frequency, and pulse jump number prediction model to perform real-time prediction of the duty cycle, frequency, and pulse jump number, and perform an inverse normalization operation; wherein the inverse normalization formula is:

[0050]

[0051] in, is the original data, is the normalized data, and are the minimum and maximum values ​​in the original data respectively;

[0052] The process of denormalization is to convert the normalized data back to the scale of the original data.

[0053] As a further improvement of the technical solution, in S6, the power converter is intelligently selected to operate in PWM mode, PFM mode, or PSM mode according to the efficiency of the power converter during operation and the predicted efficiency of the other two modes; the input voltage is selected , output voltage , output current , cycle duration Characterized by input current As labels, train the input current prediction model under PWM, PFM and PSM modulation; assume that the power converter is now running in PWM mode, , , , All can be measured, and the efficiency during operation can be measured as:

[0054]

[0055] At this time, the input current prediction model of PFM / PSM is used to measure the , , , , predicting the input current in PFM / PSM mode, the efficiency in this mode can be obtained:

[0056]

[0057] in, is the measured value of the input current, is the predicted value of input current;

[0058] By comparison , , select the operating mode of the power converter.

[0059] The second object of the present invention is to provide a power converter control device based on physical information machine learning, including a processor, a memory, and a computer program stored in the memory and running on the processor, and the processor is used to implement the steps of the power converter control method based on physical information machine learning as described above when executing the computer program.

[0060] The third object of the present invention is to provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the steps of the power converter control method based on physical information machine learning as mentioned above.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. In the power converter control method, device and storage medium based on physical information machine learning, the duty cycle, frequency and pulse jump number of the control signal are collected, and the input voltage, input current, output voltage, output current and temperature are collected and A / D converted, and the information is spliced ​​and stored for training the machine learning control model. In order to make the machine learning control model adapt to the high-frequency control of the signal, the physical constraint information is embedded in the loss function, and the relationship between the duty cycle, frequency and jump pulse number of the control signal and the input voltage, input current, output voltage, output current and temperature is mined by the machine learning algorithm. The machine learning control model is deployed on the FPGA, and the duty cycle, frequency and pulse jump number of the control signal are predicted and controlled in real time. The information and dead time are sent to the PWM signal generator, the PFM signal generator and the PSM signal generator respectively, and the switch tube of the power converter is controlled to achieve stable output of the output voltage;

[0063] 2. In the power converter control method, device and storage medium based on physical information machine learning, since machine learning is based on statistics, probability theory, etc., it has a relatively strong mathematical foundation. Traditional machine learning models such as linear regression, logistic regression and decision trees have a small number of parameters and the meaning of each parameter is clear, easy to understand and explain. This scheme gives priority to SVR, RF, RBF, BP neural network machine learning algorithms, and uses deep learning algorithms such as LSTM for complex situations; through the nonlinear ability of the machine learning model, the nonlinear effects of input voltage, input current, output voltage, output current and temperature on the duty cycle, frequency and number of pulse jumps of the control signal can be accurately modeled. Compared with modeling through mathematical methods, it can save a lot of costs and improve the accuracy of the model; machine learning combines physical laws with machine learning technology, uses known physical laws to guide the design and training process of machine learning models, and integrates physical constraints with data-driven to improve the model's predictive ability, generalization ability and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is an overall schematic diagram of an exemplary power converter control system in the present invention;

[0065] Figure 2 is a general architecture diagram of an exemplary power converter control model in the present invention;

[0066] Figure 3 is an exemplary power converter data acquisition block diagram of the present invention;

[0067] Figure 4 Substituting the loss function into the exemplary model training in the present invention to construct an illustration diagram;

[0068] Figure 5 This is an exemplary model deployment flow chart of the present invention;

[0069] Figure 6 An exemplary pulse width frequency pulse jump quantity prediction model selection algorithm diagram in the present invention;

[0070] Figure 7 This is a diagram of an exemplary electronic computer platform device architecture in the present invention. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] Example 1

[0073] like Figure 1-Figure 6 As shown, this embodiment provides a power converter control method based on physical information machine learning, characterized in that it includes the following steps:

[0074] S1. Control system construction: Build a control system architecture including power converter, data acquisition and processing module, machine learning control algorithm and control signal generator;

[0075] In the control system architecture constructed in this step:

[0076] Power converters include at least Buck, Boost, Buck-Boost, CUK and resonant converters, as well as flyback converters, forward converters, push-pull converters, half-bridge converters and full-bridge converters, and parallel topology architectures of these types of converters, and power converters that use ZVS and ZCS technologies to reduce switching losses of power devices;

[0077] The data acquisition and processing module includes the acquisition of the control signal duty cycle, frequency and pulse jump number, the acquisition of input voltage, input current, output voltage, output current and temperature, and A / D conversion; at the same time, it is also necessary to store the maximum and minimum data used for the denormalization process to convert the normalized data back to the scale of the original data;

[0078] The machine learning control algorithm includes at least SVR, RF, BP neural network, RBF and LSTM; the prediction model selection algorithm of the power converter can also adopt these machine learning algorithms, and the prediction model selection algorithm selects whether the topology of the power converter is operated in PWM mode or in PFM / PSM mode according to the operating efficiency and predicted efficiency of the power converter;

[0079] The control signal generator generates (pulse width modulation) PWM signals, (pulse frequency modulation) PFM signals and (pulse step modulation) PSM signals according to the duty cycle, frequency, number of pulse jumps, and dead time predicted by the machine learning algorithm, and controls the switching of power MOS tubes or IGBT power devices through the driving circuit.

[0080] The core of power converter control implementation includes data collection and preprocessing, construction and training of control signal (pulse width, frequency and pulse jump number) prediction model, and model deployment, such as Figure 2 shown.

[0081] S2, data collection, storage and preprocessing: collect physical information and control signal data of power converter operation, splice and store the collected data to facilitate the training and application of machine learning algorithms, and clean and normalize the data; the specific process includes the following steps:

[0082] S2.1. Data collection: Collect the input and output voltage and current of the power converter and the environmental parameters, including input voltage, input current, output voltage, output current, and temperature data of the heat sink, transformer, and air, as features for machine learning; collect control signal data, including control signal duty cycle, frequency, and pulse jump number data, as label data for machine learning; specifically, Figure 3 As shown;

[0083] When collecting the data on the number of jump pulses, multiple factors need to be considered comprehensively, including the system response time, the frequency of load changes, the requirements of the control algorithm, the stability of the output voltage and the computing power of the system. A suitable measurement cycle length can optimize the performance of the PSM system.

[0084] Since the power converter is driven by a high-frequency signal, a high-frequency signal synchronous data acquisition solution is required to ensure the acquisition of the control signal duty cycle and frequency, the time synchronization of the input voltage, input current, output voltage, output current and temperature acquisition, and the accuracy of data acquisition.

[0085] For sampling of pulse width and frequency, the rising or falling edge of the pulse can be captured by the STM32 timer, and the pulse width and frequency can be calculated by using the timer interrupt, and the number of pulse jumps can be further calculated.

[0086] For sampling tags, if the sampling period For a 500kHz PWM signal, it includes 500 PWM pulse signals, and the machine learning model predicts the width of each pulse in these 500 PWM signals; for a 500kHz PSM (pulse skipping modulation) signal, it also includes 500 potential pulses within 1 millisecond, and the machine learning model predicts the number of pulses that actually jump (i.e., pulses appear) among these 500 potential pulses; for a PFM (pulse frequency modulation) signal, the machine learning model predicts the interval time between pulses within 1 millisecond, and the number of pulses is calculated based on the pulse width and total time.

[0087] In addition, data collection also includes the following constraints:

[0088] First, data collection for modeling:

[0089] The data of the whole process of the power converter from startup to subsequent operation must be covered;

[0090] Special attention is paid to data acquisition when switching between PWM and PFM / PSM to ensure the accuracy of inference under different modulation modes;

[0091] The initial environment at startup should cover the ambient temperature changes throughout the year.

[0092] Second, data collection for reasoning:

[0093] Data from a period before the inference must be included to ensure sufficient contextual information;

[0094] Contains initial environment data so that the model can be initialized accurately;

[0095] Special attention is paid to data acquisition when switching between PWM and PFM / PSM to ensure the accuracy of inference under different modulation modes;

[0096] Data collected for inference can only retain data for a short period of time to speed up inference.

[0097] S2.2, Data storage: Sort the power converter data according to the acquisition time, use PWM duty cycle and PFM frequency as labels, and use input voltage, input current, output voltage, output current and temperature characteristics as model training data to form a data set, as shown in Table 1 below. The data acquisition of the training mode must ensure the integrity of the data acquisition:

[0098] Table 1 Power converter data table structure

[0099]

[0100] S2.3. Data preprocessing: fill missing values ​​in the collected power converter data and normalize the data set to accelerate model convergence.

[0101] In this step, data preprocessing specifically includes:

[0102] If there are some missing values ​​in the data set, we can use the data from the previous time point to supplement them, or use the weighted average of the previous and next data to supplement them;

[0103] In the case of standard dimensions, the input voltage, input current, output voltage, output current and temperature characteristic data of the power converter, as well as the duty cycle, frequency, and pulse jump number label data have large data differences. The data is preprocessed by normalizing the maximum and minimum values, and the features and labels of different scales are scaled to In order to eliminate the influence of large and small data on the model, the normalization formula is:

[0104]

[0105] In the formula, is the data to be normalized, that is, the original data, is the normalized data, and They represent the maximum and minimum values ​​in the data to be normalized respectively. Finally, the data is mapped between 0 and 1 and input into the model as feature data for training. The processed data is divided into training set and test set in a ratio of 7:3, where the first 70% of the data set is used as the training set and the last 30% as the test set.

[0106] S3. Model building and evaluation: Use machine learning to build an accurate mathematical control model related to the operation of the power converter, and evaluate and optimize the model based on the loss function; the specific process includes the following steps:

[0107] S3.1. Model building and training: Construct a machine learning control model using SVR, RF, RBF, and BP neural network machine learning algorithms, and use deep learning algorithms such as LSTM for complex situations; In the design of power converters, the high nonlinearity of power converters makes it very difficult to establish an accurate mathematical model, which makes it difficult to design and debug nonlinear PID. Machine learning is used as a control method, and the nonlinear mapping ability of machine learning is used to learn the relationship between input voltage, input current, temperature, output voltage, output current and pulse width, frequency, and number of jump pulses;

[0108] S3.2. Model evaluation: Apply machine learning to the design of power converter control systems, and introduce physical constraint losses into the loss function of the machine learning model to reflect the penalty terms of physical laws or system characteristics; design a loss function that includes label loss, consideration of the error between the output voltage and the expected voltage, and efficiency loss terms to ensure that the output voltage remains within a fixed error range; the design of physical constraint losses aims to ensure that the behavior predicted by the model is consistent with the behavior of the actual physical system.

[0109] In this step, when machine learning predicts duty cycle, frequency, and the number of pulse jumps, the labels are duty cycle, frequency, and the number of pulse jumps, and the loss function is defined based on label loss, output voltage loss, and efficiency loss; the reason for adding output voltage and output power efficiency to the loss function is that the adjustment of duty cycle directly affects the quality of output voltage, and output voltage and output power efficiency are the most concerned performance indicators in actual application scenarios. Therefore, even if the model directly predicts the duty cycle, frequency, and the number of pulse jumps, by adding the fluctuation of output voltage, the deviation of output voltage, and the change of input and output power to the loss function, the trained model can take into account the fluctuation, deviation, and efficiency of output voltage, and select parameters that can better reflect the final performance of the system to add to the loss function, which helps to optimize the actual application effect of the model, such as Figure 4 As shown;

[0110] The mean square error (MSE) or absolute error (MAE) is used to measure the error between the output label and the actual label. The fluctuation of the output voltage, the error between the output voltage and the expected voltage, and the efficiency are introduced as constraints. The constraints are used as part of the loss function, that is, a penalty term related to the constraints is added to the original loss function. If the model violates the constraints, the penalty term will increase the loss value, thereby encouraging the model to adjust to meet the constraints. The penalty term of the loss function is used to guide the model to learn in the direction of meeting the conditions. Among them:

[0111] The mean square error (MSE) is used to measure the prediction error of the label. :

[0112]

[0113] in, It is The predicted pulse width of samples, It is The actual pulse width of samples, is the sample size;

[0114] In order to ensure the stability of the output voltage, additional constraints are introduced on the basis of the basic loss function, namely, output voltage fluctuation, output voltage deviation and efficiency constraints, and three penalty terms are introduced accordingly, namely, fluctuation penalty term, voltage deviation penalty term and efficiency penalty term;

[0115] Define a penalty term to measure the fluctuation of the output voltage: for example, the change in the output voltage between consecutive time points can be calculated and included in the loss function.

[0116]

[0117] in, is the fluctuation function of the output voltage, i.e., the fluctuation penalty, It is Output voltage at a time point; It is Output voltage at a time point;

[0118] Define a penalty term to measure the deviation between the output voltage and the target voltage; use the mean square error (MSE) to measure the error between the output voltage and the target voltage , i.e., the deviation penalty, as a constraint on the output voltage:

[0119]

[0120] in, is the target output voltage, is the output voltage;

[0121] Introducing efficiency constraints helps machine learning models better understand and predict the behavior of the system:

[0122]

[0123] in, is the output power efficiency constraint function, i.e., power penalty; , They are The input voltage and input current at a time point, , They are Output voltage and output current at a time point;

[0124] The basic loss function and the stability constraint are combined to form a comprehensive loss function:

[0125]

[0126] in, It is The actual output voltage of a sample is a normalized value; is the expected output voltage, which is a normalized value and a fixed value;

[0127] in, is a hyperparameter that controls the weight of the volatility penalty; is a hyperparameter that controls the weight of the deviation penalty; is a hyperparameter that controls the weight of the power penalty.

[0128] S4, Model deployment: Optimize and compress the model to reduce resource usage and increase inference speed, compile the trained model into a format suitable for MCU, ASIC and FPGA, and deploy it on the embedded system; Figure 5 As shown;

[0129] S5, model application: inputting the real-time measured power converter operating parameters into the corresponding prediction model for real-time prediction;

[0130] In this step, the real-time measured input voltage, input current, output voltage, output current and temperature are input into the duty cycle, frequency, pulse jump number prediction model to perform real-time prediction of the duty cycle, frequency, and pulse jump number, and perform an inverse normalization operation; wherein the inverse normalization formula is:

[0131]

[0132] in, is the original data, is the normalized data, and are the minimum and maximum values ​​in the original data respectively;

[0133] The process of denormalization is to convert the normalized data back to the scale of the original data. The specific operation depends on the normalization method you used in the preprocessing stage. Make sure to save the normalization parameters during training and use these parameters to denormalize when predicting to get the prediction results on the original scale.

[0134] When generating the number of jump pulses of PSM, the distribution of jump pulses needs to be considered. The distribution of jump pulses depends on the specific control strategy and system requirements, and there are distribution modes such as uniform distribution, random distribution, and centralized distribution. When the power switch tube implements ZVS, the number of jump pulses needs to be centralized and the number of jump pulses needs to be an even number to ensure the ZVS shutdown of the power switch tube. Therefore, when the number of jump pulses is detected to be an odd number, the number of pulses needs to be increased or decreased by 1.

[0135] S6. Selection of prediction model of power converter operation mode: intelligently selecting the operation model of the power converter according to the operation efficiency and prediction efficiency of the power converter during operation;

[0136] In this step, if Figure 6 As shown, the efficiency of the power converter during operation and the predicted efficiency of the other two modes are used to intelligently select whether the power converter is operated in PWM mode, PFM mode, or PSM mode; the selection is based on the input voltage , output voltage , output current , cycle duration Characterized by input current As labels, train the input current prediction model under PWM, PFM and PSM modulation; assuming that the power converter is now operating in PWM mode, , , , All can be measured, and the efficiency during operation can be measured as:

[0137]

[0138] At this time, the input current prediction model of PFM / PSM is used to measure the , , , , predicting the input current in PFM / PSM mode, the efficiency in this mode can be obtained:

[0139]

[0140] in, is the measured value of the input current, is the predicted value of input current;

[0141] By comparison , , select the operating mode of the power converter.

[0142] This technical solution proposes a power converter control method based on physical information machine learning; it comprehensively considers the factors that affect the duty cycle and frequency of the control signal, namely the input voltage, input current, output voltage, output current and temperature, where the temperature includes the radiator temperature, transformer temperature and the air temperature in the equipment; it proposes that when measuring and generating the number of jump pulses of PSM, the sampling period and the distribution theory of pulses need to be considered; it can be used alone for a PWM-based power converter control method and a PFM / PSM-based power converter control method, and can also be used for a hybrid control method of a power converter based on PWM and PFM / PSM; it proposes a design scheme for a digital feedback loop and a design method for a loss function based on power loss, and uses a machine learning algorithm for high-frequency fine control in high-frequency application scenarios, and the desired output voltage A loss function is introduced.

[0143] Compared with the existing technology, this technical solution has the following advantages:

[0144] (1) Through digitalization, it has revolutionized the research and development of power converter control circuits, especially providing a new idea for the research and development of wide voltage input, wide load output and high power power converters;

[0145] (2) Adaptive control: The parameters of a PID controller are usually fixed or adjusted through a simple adaptive algorithm; this means that the PID controller may not be able to adjust in time to maintain optimal performance when faced with changes in system parameters (such as load changes, temperature changes, etc.); machine learning models can learn and adapt to changes in system parameters online, such as load changes, temperature changes, etc.; this means that the controller can maintain good performance even when system conditions change;

[0146] (3) Nonlinear processing: Power converters often exhibit highly nonlinear behavior, and PID controllers may have difficulty handling nonlinear systems; machine learning-based control methods are better able to capture and handle these nonlinear relationships;

[0147] (4) Dynamic response: The response of a PID controller is usually based on a linear assumption and may not be able to respond quickly and accurately to dynamic changes in the system, especially when system parameters change significantly or when there are external disturbances. Machine learning models, especially deep neural networks, can learn and predict the dynamic behavior of the system, thereby providing faster and more accurate control responses.

[0148] (5) Robustness: PID controllers have poor robustness to model errors, measurement noise, and external disturbances. Once the system deviates from the predetermined operating point, the PID controller may not be able to respond effectively, resulting in performance degradation. Machine learning-based controllers can handle uncertainties such as model errors, measurement noise, and external disturbances through learning, thereby improving the robustness of the system.

[0149] (6) Global optimization: PID controller parameter adjustment is usually based on local optimization methods, which may not be able to find the global optimal solution. In contrast, machine learning can find a better control strategy through global search.

[0150] (7) Accurate modeling: PID controllers require an accurate system model for parameter tuning. For complex systems that are difficult to model or have uncertain models, the effectiveness of PID controllers may be affected. Machine learning-based control methods can learn from large amounts of data and use historical data to improve control strategies. This is particularly important in the era of big data. Machine learning models do not require accurate modeling, but rather learn the input-output relationship of the system through data, which is particularly useful for complex systems that are difficult to model.

[0151] (8) Manual parameter adjustment: PID controller parameter adjustment usually requires professional knowledge and experience, which increases the workload and complexity of manual parameter adjustment. Machine learning models can automatically adjust parameters through training, reducing the workload and complexity of manual parameter adjustment.

[0152] (9) Continuous optimization: Over time, the model can improve control performance through continuous learning and optimization. As computing power and data volume increase, machine learning models can be continuously improved and expanded to cope with more complex control tasks. When system parameters change or environmental conditions change, PID parameters may need to be adjusted frequently, increasing maintenance costs and time.

[0153] (10) Flexibility: PID controllers are usually designed for specific systems and applications and lack flexibility; if the system topology or application requirements change, the controller may need to be redesigned and adjusted; machine learning-based control methods can easily adapt to different power converter topologies and application requirements.

[0154] like Figure 7 As shown, this embodiment also provides a power converter control device based on physical information machine learning, including a processor, a memory, and a computer program stored in the memory and running on the processor.

[0155] The processor includes one or more processing cores. The processor is connected to a memory via a bus. The memory is used to store program instructions. When the processor executes the program instructions in the memory, the steps of the power converter control method based on physical information machine learning are implemented.

[0156] Optionally, the memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0157] In addition, this embodiment also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned power converter control method based on physical information machine learning are implemented.

[0158] Optionally, the present invention also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the steps of the power converter control method based on physical information machine learning in the above-mentioned aspects.

[0159] A person of ordinary skill in the art can understand that the process of implementing all or part of the steps of the above-mentioned embodiments can be completed by hardware, or can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a disk or an optical disk, etc.

[0160] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A power converter control method based on physical information machine learning, characterized in that: The steps include: S1. Control system construction: Build a control system architecture including power converter, data acquisition and processing module, machine learning control algorithm and control signal generator; S2, data collection, storage and preprocessing: collect physical information and control signal data of power converter operation, splice and store the collected data, and clean and normalize the data; S3. Model building and evaluation: Use machine learning to build an accurate mathematical control model related to the operation of the power converter, and evaluate and optimize the model based on the loss function; S4, model deployment: optimize and compress the model, compile the trained model into a format suitable for MCU, ASIC and FPGA, and deploy it on the embedded system; S5, model application: inputting the real-time measured power converter operating parameters into the corresponding prediction model for real-time prediction; S6. Selection of a prediction model for the power converter operation mode: intelligently selecting the power converter operation model based on the operation efficiency and prediction efficiency of the power converter during operation.

2. The power converter control method based on physical information machine learning according to claim 1, characterized in that: In the control system architecture constructed by S1: Power converters include at least Buck, Boost, Buck-Boost, CUK and resonant converters, as well as flyback converters, forward converters, push-pull converters, half-bridge converters and full-bridge converters, and parallel topology architectures of these types of converters, and power converters that use ZVS and ZCS technologies to reduce switching losses of power devices; The data acquisition and processing module includes the acquisition of the control signal duty cycle, frequency and pulse jump number, the acquisition of input voltage, input current, output voltage, output current and temperature, and A / D conversion; The machine learning control algorithm includes at least SVR, RF, BP neural network, RBF and LSTM; the prediction model selection algorithm of the power converter can also adopt these machine learning algorithms, and the prediction model selection algorithm selects whether the topology of the power converter is operated in PWM mode or in PFM / PSM mode according to the operating efficiency and predicted efficiency of the power converter; The control signal generator generates pulse width modulation (PWM) signals, pulse frequency modulation (PFM) signals and pulse step modulation (PSM) signals according to the duty cycle, frequency, number of pulse jumps, and dead time predicted by the machine learning algorithm, and controls the switching of power MOS tubes or IGBT power devices through the driving circuit.

3. The power converter control method based on physical information machine learning according to claim 2, characterized in that: The specific process of S2 includes the following steps: S2.

1. Data collection: Collect the input and output voltage and current of the power converter and the environmental parameters, including input voltage, input current, output voltage, output current, and temperature data of the heat sink, transformer, and air as features for machine learning; Collect control signal data, including control signal duty cycle, frequency, and pulse jump quantity data, as label data for machine learning; S2.2, data storage: sort the power converter data according to the acquisition time, use PWM duty cycle and PFM frequency as labels, and use input voltage, input current, output voltage, output current and temperature characteristics as model training data to form a data set; S2.

3. Data preprocessing: fill missing values ​​in the collected power converter data and normalize the data set.

4. The power converter control method based on physical information machine learning according to claim 3 is characterized in that: In S2.3, data preprocessing specifically includes: If there are some missing values ​​in the data set, we can use the data from the previous time point to supplement them, or use the weighted average of the previous and next data to supplement them; In the case of standard dimensions, the input voltage, input current, output voltage, output current and temperature characteristic data of the power converter, as well as the duty cycle, frequency, and pulse jump number label data have large data differences. The data is preprocessed by normalizing the maximum and minimum values, and the features and labels of different scales are scaled to In order to eliminate the influence of large and small data on the model, the normalization formula is: In the formula, is the data to be normalized, that is, the original data, is the normalized data, and They represent the maximum and minimum values ​​in the data to be normalized respectively. Finally, the data is mapped between 0 and 1 and input into the model as feature data for training. The processed data is divided into training set and test set in a ratio of 7:3, where the first 70% of the data set is used as the training set and the last 30% as the test set.

5. The power converter control method based on physical information machine learning according to claim 4 is characterized in that: The specific process of S3 includes the following steps: S3.

1. Model building and training: Construct a machine learning control model using SVR, RF, RBF, and BP neural network machine learning algorithms, and use LSTM deep learning algorithms for complex situations; use machine learning as a control method, and use the nonlinear mapping ability of machine learning to learn the relationship between input voltage, input current, temperature, output voltage, output current and pulse width, frequency, and number of jump pulses; S3.

2. Model evaluation: Apply machine learning to the design of power converter control systems, and introduce physical constraint losses into the loss function of the machine learning model to reflect the penalty terms of physical laws or system characteristics; design a loss function that includes label loss, consideration of the error between the output voltage and the expected voltage, and efficiency loss terms to ensure that the output voltage remains within a fixed error range.

6. The power converter control method based on physical information machine learning according to claim 5, characterized in that: In S3.2, when machine learning predicts duty cycle, frequency and pulse jump number, the labels are duty cycle, frequency and pulse jump number, and the loss function is defined according to label loss, output voltage loss and efficiency loss; The mean square error (MSE) or absolute error (MAE) is used to measure the error between the output label and the actual label. The fluctuation of the output voltage, the error between the output voltage and the expected voltage, and the efficiency are introduced as constraints. The constraints are used as part of the loss function, that is, a penalty term related to the constraints is added to the original loss function. If the model violates the constraints, the penalty term will increase the loss value, thereby encouraging the model to adjust to meet the constraints. The penalty term of the loss function is used to guide the model to learn in the direction of meeting the conditions. Among them: The mean square error (MSE) is used to measure the prediction error of the label. : in, It is The predicted pulse width of samples, It is The actual pulse width of samples, is the sample size; In order to ensure the stability of the output voltage, additional constraints are introduced on the basis of the basic loss function, namely, output voltage fluctuation, output voltage deviation and efficiency constraints, and three penalty terms are introduced accordingly, namely, fluctuation penalty term, voltage deviation penalty term and efficiency penalty term; Define a penalty term to measure the fluctuation of output voltage: in, is the fluctuation function of the output voltage, i.e., the fluctuation penalty, It is Output voltage at a time point; It is Output voltage at a time point; Define a penalty term to measure the deviation between the output voltage and the target voltage; use the mean square error MSE to measure the error between the output voltage and the target voltage , i.e., the deviation penalty, as a constraint on the output voltage: in, is the target output voltage, is the output voltage; Introducing efficiency constraints helps machine learning models better understand and predict the behavior of the system: in, is the output power efficiency constraint function, i.e., power penalty; , They are The input voltage and input current at a time point, , They are Output voltage and output current at a time point; The basic loss function and the stability constraint are combined to form a comprehensive loss function: in, It is The actual output voltage of a sample is a normalized value; is the expected output voltage, which is a normalized value and a fixed value; in, is a hyperparameter that controls the weight of the volatility penalty; is a hyperparameter that controls the weight of the deviation penalty; is a hyperparameter that controls the weight of the power penalty.

7. The power converter control method based on physical information machine learning according to claim 6, characterized in that: In S5, the input voltage, input current, output voltage, output current and temperature measured in real time are input into the duty cycle, frequency and pulse jump number prediction model to perform real-time prediction of the duty cycle, frequency and pulse jump number, and perform an inverse normalization operation; wherein the inverse normalization formula is: in, is the original data, is the normalized data, and are the minimum and maximum values ​​in the original data respectively; The process of denormalization is to convert the normalized data back to the scale of the original data.

8. The power converter control method based on physical information machine learning according to claim 7, characterized in that: In S6, the power converter is intelligently selected to operate in PWM mode, PFM mode, or PSM mode based on the efficiency of the power converter during operation and the predicted efficiency of the other two modes; the input voltage is selected , output voltage , output current , cycle duration Characterized by input current As labels, train the input current prediction model under PWM, PFM and PSM modulation; assume that the power converter is now running in PWM mode, , , , All can be measured, and the efficiency during operation can be measured as: At this time, the input current prediction model of PFM / PSM is used to measure the , , , , predicting the input current in PFM / PSM mode, the efficiency in this mode can be obtained: in, is the measured value of the input current, is the predicted value of input current; By comparison , , select the operating mode of the power converter.

9. A power converter control device based on physical information machine learning, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and running on the processor, and the processor is used to implement the steps of the power converter control method based on physical information machine learning as described in any one of claims 1-8 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the power converter control method based on physical information machine learning as described in any one of claims 1 to 8 are implemented.

Images