CLLC resonant converter soft start method based on edge neural network prediction

Abstract

The invention discloses a CLLC resonant converter soft start method based on edge neural network prediction. The method comprises the following steps: firstly, acquiring soft start process data of a circuit under different conditions by using a voltage and current sensor; secondly, constructing an LSTM network to obtain a CLLC circuit model predictor; and taking the process data at the current moment as input, and outputting an output voltage value and a resonance current peak value of the next period to obtain a real optimal switching frequency value. And then a real-time lightweight network is established as an optimal switching frequency controller, a load coefficient, a voltage gain and a resonant current peak value are taken as lightweight network input, and a predicted optimal switching frequency value is output. And finally, combining the optimal switching frequency controller with a PI controller to realize smooth soft start of the CLLC resonant converter. According to the invention, under the conditions of dynamic load and low storage cost, a fast soft starting method with low peak current is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of power electronics, and in particular relates to a CLLC resonant converter soft starting method based on edge neural network prediction. Background Art

[0002] The CLLC resonant converter is an advanced power electronic converter that combines the advantages of the traditional LLC converter in design and further optimizes its performance, especially in the application scenario of bidirectional energy transmission. This converter is particularly suitable for renewable energy power generation, energy storage systems, electric vehicle charging infrastructure, and high-efficiency power conversion.

[0003] However, traditional analog controllers are prone to high inrush currents due to the difference between their linear characteristics and the nonlinear behavior of resonant converters. This not only increases system losses, but may also cause damage to system components. The soft-start process of the CLLC resonant converter faces the challenges of high inrush current and high surge voltage due to its resonant network.

[0004] Although some methods such as state plane analysis and state trajectory analysis can control the current during the startup process, the startup time is long due to the difficulty in measuring high-frequency current and the time delay in the control loop. Typically, the PWM signal is refreshed every 6-12 switching cycles, further extending the startup time. Some proposed digital strategies use a lookup table (LUT) to manage the startup process, but this method relies on the resonant parameters, output capacitance, and load resistance, requiring a high-dimensional LUT, increasing the complexity and storage requirements of the system. Other studies have used additional transformers and diodes to clamp the voltage on the resonant capacitor, but this increases the complexity and cost of the system. Summary of the invention

[0005] In order to overcome the deficiencies of the prior art, the present invention provides a soft starting method for a CLLC resonant converter based on neural network prediction with dynamic control, good stability and fast response speed.

[0006] Based on the CNN model, the present invention proposes a CLLC resonant converter soft start control method, which can effectively suppress the surge current and surge voltage in the resonant cavity. At the same time, based on the LSTM model, a CLLC resonant converter voltage and current predictor is proposed to assist CNN iterative optimization.

[0007] Step (1): Use voltage and current sensors to obtain soft start process data under different circuit conditions.

[0008] The data of the CLLC resonant converter startup cycle under m groups of different load conditions are collected through high-precision op amps. The load of each group differs by 2%, from no-load to full-load in the entire working range. At the same time, each group contains n control cycles, and the switching frequency, output voltage, and resonant current peak value in each cycle are recorded in the SD memory card. Then the data is imported into the computer.

[0009] Step (2): Construct an LSTM network on the computer side to obtain a CLLC circuit model predictor.

[0010] The switching frequency, output voltage and resonant current peak value at the current moment in step (1) are used as network inputs, and the network outputs are the output voltage value and resonant current peak value of the next cycle, thereby obtaining the required CLLC circuit model predictor, and then obtaining the true optimal switching frequency value. The processing network for constructing the CLLC circuit model predictor involves multiple key components, including sequentially connecting the long short-term memory (LSTM) recursive neural network, the transposed matrix, the one-dimensional data expansion, the Softmax activation function, the regularization technology, and the fully connected output layer. The transposed matrix operation is performed on the second and third dimensions of the data. In order to improve the generalization ability of the model, the regularization technology is introduced. Finally, the fully connected layer maps the features to two output nodes to complete the prediction task of the model.

[0011] The fully connected layer maps the LSTM output to the final prediction result to obtain the output voltage and resonant current peak of the next cycle of the CLLC resonant converter. Constrain the resonant current peak and output voltage to obtain a reliable soft start process, and the formula is:

[0012] V o (k+1)<V ref

[0013] I pk (k+1)=I max

[0014] Where V o (k+1) represents the voltage value of the next moment output by the predictor, V ref Represents the target voltage value.

[0015] I pk (k+1) represents the next moment resonant current peak value of the predictor output, I max Represents the maximum value of the resonant current peak. Through the constraints, the optimal switching frequency value is obtained and the result is saved.

[0016] Step (3): Build a real-time lightweight network as the optimal switching frequency controller.

[0017] The load factor, voltage gain and resonant current peak are used as lightweight network inputs, and the output is the predicted optimal switching frequency value, thereby obtaining the required CLLC circuit optimal switching frequency controller. The load factor and voltage gain need to be calculated using the data collected in step 1, and the calculation formula is as follows:

[0018]

[0019]

[0020] Where k represents the current moment, n represents the transformer turns ratio, g(k) represents the load factor at the current moment, and I o (k) represents the output current value at the current moment, V o (k) represents the output voltage value at the current moment, M(k) represents the voltage gain value at the current moment, V i (k) represents the input voltage value at the current moment.

[0021] The lightweight network architecture consists of several residual convolutional units and fully connected layers, where features are first extracted and learned by several residual convolutional units, and then passed to the fully connected layer to generate output, which is the predicted optimal switching frequency value. Each residual convolutional unit consists of two branches in parallel, one branch consists of three layers of dilated convolutional layers with different dilation rates in sequence, and the second branch consists of one layer of one-dimensional convolutional layer. The outputs of the two branches are added together to obtain the output of the residual convolutional unit. Specifically, the step size of the first layer of dilated convolution is set to k, while the step size of the other two layers of dilated convolution is 1, and the dilation rates are 1, n, and n, respectively. 2 .

[0022] The fully connected layer maps the output features of the residual convolution unit to the output result, obtains the optimal switching frequency, predicts the optimal switching frequency value, and performs reverse training based on the actual optimal switching frequency value; and obtains the optimal switching frequency controller of the CLLC circuit.

[0023] Step (4): Combine the optimal switching frequency controller with the PI controller to achieve smooth soft start of the CLLC resonant converter.

[0024] First, determine the initial pulse width to avoid causing a large resonant current peak in the resonant cavity. There is no energy in the resonant tank at the beginning of soft start. Therefore, assuming that the output voltage remains unchanged within a cycle, the equation can be derived as follows:

[0025] Ir1c(θ)=0.5(sin(θ)+k1 sin(k1θ))

[0026]

[0027] Where Lr represents the resonant inductance, Lm represents the transformer excitation inductance, I r1c (θ) represents the first cycle of the resonant cavity current waveform.

[0028] It can be seen from the equation that the maximum resonant current peak is obtained at one quarter of the resonant period, so the initial pulse width should be greater than twice the resonant frequency. Here, the initial switching frequency is selected as three times the resonant frequency to suppress the resonant current peak. Then collect the output voltage data, output current data, and calculate the load factor and voltage gain. The load factor, voltage gain, resonant current peak and other data are input into the optimal switching frequency controller of step (3), and its output is the optimal control frequency to achieve the increase of the output voltage of the CLLC resonant converter. The output voltage is increased to within 5% of the target voltage through the optimal switching frequency controller, and the increased output voltage is switched to the PI controller for control, thereby achieving a smooth voltage transition and completing the entire soft start process.

[0029] Beneficial effects of the present invention:

[0030] The present invention utilizes a complex neural network to implement a CLLC resonant converter output voltage and resonant current peak predictor, and then trains a lightweight neural network through the output values ​​of the output voltage and resonant current peak predictor to implement an optimal switching frequency controller, and combines a PI controller to implement a smooth soft-start voltage transition, which can implement a fast low resonant current peak soft-start method under dynamic load and low storage cost conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the present invention;

[0032] Figure 2 It is a schematic diagram of the result of modulating the output voltage of the present invention. DETAILED DESCRIPTION

[0033] The present invention provides a CLLC resonant converter soft starting method based on edge neural network prediction, such as Figure 1 As shown, the following steps are included:

[0034] Step (1): Use voltage and current sensors to obtain soft start process data under different circuit conditions.

[0035] The data of the startup cycle of the CLLC resonant converter under 50 groups of different load conditions are collected through high-precision op amps. The load of each group differs by 2%, covering the entire working range from no-load to full-load. At the same time, each group contains 200 control cycles, and the switching frequency, output voltage, and resonant current peak value in each cycle are recorded in the SD memory card. Then the data is imported into the computer.

[0036] Step (2): Construct an LSTM network on the computer side to obtain a CLLC circuit model predictor.

[0037] The switching frequency, output voltage and resonant current peak value at the current moment in step (1) are used as network inputs, and the network outputs are the output voltage value and resonant current peak value of the next cycle, so as to obtain the required CLLC circuit model predictor, and then obtain the true optimal switching frequency value. The processing network for constructing the CLLC circuit model predictor involves multiple key components, including sequentially connecting the long short-term memory (LSTM) recursive neural network, the transposed matrix, the one-dimensional data expansion, the Softmax activation function, the regularization technology and the fully connected output layer. The transposed matrix operation is performed on the second and third dimensions of the data. In order to improve the generalization ability of the model, the regularization technology is introduced, and its coefficient is set to 0.2. Finally, the fully connected layer maps the features to 2 output nodes to complete the prediction task of the model.

[0038] Long short-term memory recurrent neural networks effectively establish connections between data at different time points by enhancing the storage capacity of short-term data:

[0039]

[0040] where Y input Represents the input parameter value, F LSTM [] represents the long short-term memory neural network function.

[0041] By applying the transposed matrix (Permute), the network is able to extract data features more effectively.

[0042]

[0043] in Represents the output parameter of the long short-term memory neural network function, F Permute [] represents the matrix transpose function.

[0044] The data is reorganized through a one-dimensional flatten operation so that it can be passed as input to the fully connected layer.

[0045]

[0046] in Represents the output parameter of the matrix device function, F Flatten [] represents a one-dimensional expansion function.

[0047] The Dropout regularization technique is used to reduce the risk of overfitting of the network by randomly discarding a part of neurons.

[0048]

[0049] in Represents the output parameter of the one-dimensional expansion function, F Dropout [] represents the regularization function.

[0050] The output of the regularization function is sent to the fully connected layer, which is processed using the Softmax activation function.

[0051]

[0052] in represents the output parameter of the regularization function, represents the output parameter of the regularization function,

[0053] F c [] represents the fully connected layer function.

[0054] The fully connected layer maps the LSTM output to the final prediction result to obtain the output voltage and resonant current peak of the next cycle of the CLLC resonant converter. Constrain the resonant current peak and output voltage to obtain a reliable soft start process, and the formula is:

[0055] V o (k+1)<V ref

[0056] I pk (k+1)=I max

[0057] Where V o (k+1) represents the voltage value of the next moment output by the predictor, V ref Represents the target voltage value.

[0058] I pk (k+1) represents the next moment resonant current peak value of the predictor output, I max Represents the maximum value of the resonant current peak. Through the constraints, the optimal switching frequency value is obtained and the result is saved.

[0059] Step (3): Build a real-time lightweight network as the optimal switching frequency controller.

[0060] The load factor, voltage gain and resonant current peak are used as lightweight network inputs, and the output is the predicted optimal switching frequency value, thereby obtaining the required CLLC circuit optimal switching frequency controller. The load factor and voltage gain need to be calculated using the data collected in step 1, and the calculation formula is as follows:

[0061]

[0062] Where k represents the current moment, n represents the transformer turns ratio, g(k) represents the load factor at the current moment, and I o (k) represents the output current value at the current moment, V o (k) represents the output voltage value at the current moment, M(k) represents the voltage gain value at the current moment, V i (k) represents the input voltage value at the current moment.

[0063] The lightweight network architecture consists of several residual convolutional units and fully connected layers, where features are first extracted and learned by several residual convolutional units, and then passed to the fully connected layer to generate output, which is the predicted optimal switching frequency value. Each residual convolutional unit consists of two branches in parallel, one branch consists of three layers of dilated convolutional layers with different dilation rates in sequence, and the second branch consists of one layer of one-dimensional convolutional layer. The outputs of the two branches are added together to obtain the output of the residual convolutional unit. Specifically, the step size of the first layer of dilated convolution is set to k, while the step size of the other two layers of dilated convolution is 1, and the dilation rates are 1, n, and n, respectively. 2 These dilated convolutional layers adopt the ReLU6 activation function and are sequentially connected to form residual units.

[0064] Y i L =F cnn_1 [F cnn_2 [F cnn_3 [Y i ]]]

[0065] where Y i is the input parameter of the lightweight network, F cnn_1 [], F cnn_2 [], F cnn_3 [] represent the first, second, and third dilated convolution functions, respectively, Y i L Output of three dilated convolutional layers.

[0066] The step size of a one-dimensional convolution layer is k, and the activation function uses the ReLU6 function:

[0067] Y i H =F ReLU6 [Y i ]

[0068] where F ReLU6 [] represents the ReLU6 function, Y i H is the output of the one-dimensional convolutional layer.

[0069] One-dimensional convolution and dilated convolution are connected in parallel to form a residual block:

[0070] YiR =Yi L +Yi H

[0071] Y i R Represents the output features of the residual block. Residual connections help accelerate network convergence, shorten training time, and improve efficiency. After being processed by multiple layers of residual convolution blocks, the data features are sent to the fully connected layer, which also uses the ReLU6 activation function:

[0072] Y o =F ReLU6 [Y i R ]

[0073] Among them, F ReLU6 [] represents the ReLU6 function, Y o is the output of the fully connected layer.

[0074] The fully connected layer maps the output features of the residual convolution unit to the output result, obtains the optimal switching frequency, predicts the optimal switching frequency value, and performs reverse training based on the actual optimal switching frequency value; and obtains the optimal switching frequency controller of the CLLC circuit.

[0075] Step (4): Combine the optimal switching frequency controller with the PI controller to achieve smooth soft start of the CLLC resonant converter.

[0076] First, determine the initial pulse width to avoid causing a large resonant current peak in the resonant cavity. There is no energy in the resonant tank at the beginning of soft start. Therefore, assuming that the output voltage remains unchanged within a cycle, the equation can be derived as follows:

[0077] Ir1c(θ)=0.5(sin(θ)+k1sin(k1θ))

[0078]

[0079] Where Lr represents the resonant inductance, Lm represents the transformer excitation inductance, I r1c (θ) represents the first cycle of the resonant cavity current waveform.

[0080] It can be seen from the equation that the maximum resonant current peak is obtained at one quarter of the resonant period, so the initial pulse width should be greater than twice the resonant frequency. Here, the initial switching frequency is selected as three times the resonant frequency to suppress the resonant current peak. Then collect the output voltage data, output current data, and calculate the load factor and voltage gain. The load factor, voltage gain, resonant current peak and other data are input into the optimal switching frequency controller of step (3), and its output is the optimal control frequency to achieve the increase of the output voltage of the CLLC resonant converter. The output voltage is increased to within 5% of the target voltage through the optimal switching frequency controller, and the increased output voltage is switched to the PI controller for control, thereby achieving a smooth voltage transition and completing the entire soft start process.

[0081] The target voltage range formula is:

[0082]

[0083] Where μ(k) is the percentage of the target voltage range, V ref To set the target voltage, V o (k) is the output voltage value at time k. The PI controller formula is as follows:

[0084] u(k+1)=u(k)+Kp(e(k)-e(k-1))+Kie(t)

[0085] Where k represents the current moment, u(k) represents the switching frequency at the current moment, u(k-1) represents the switching frequency of the next cycle, e(k) represents the voltage error at the current moment, e(k-1) represents the voltage error of the previous cycle, K p is the proportionality coefficient, K i is the integration coefficient.

[0086] In this embodiment, based on the test system load, the start-up time and the resonant current peak of the two soft start methods are tested respectively, the output voltage is set to 380V, and the load resistance is 75Ω. The results are shown in the following Table 1:

[0087] Table 1

[0088]

[0089] As shown in Table 1, the collaborative strategy of the lightweight network controller and the proportional-integral (PI) controller proposed in this study can effectively ensure the fast startup of the CLLC resonant converter with low resonant current peak. Under the premise of maintaining the maximum value of the resonant current peak unchanged, this strategy shows a shorter startup time compared with the optimal PI controller. In addition, the performance advantage of the optimal PI controller tends to weaken when the load state changes.

[0090] Figure 2 The results of the modulated output voltage of the lightweight network controller and the PI controller are shown. The first 0.8 milliseconds in the figure show the voltage control stage of the lightweight network controller, while the following 0.1 milliseconds are the PI control stage. The results show that the switching process between the lightweight network controller and the PI controller is smooth and does not cause any adverse oscillations. At the same time, the voltage rise process is relatively uniform and the rise time is short.

Claims

1. A CLLC resonant converter soft starting method based on edge neural network prediction, characterized in that: The following steps are involved: Step 1: Use voltage and current sensors to obtain soft start process data under different circuit conditions, including switching frequency, output voltage and resonant current peak value in each cycle; Step 2: Construct an LSTM network to obtain a CLLC circuit model predictor; take the current switching frequency, output voltage and resonant current peak as input, and the CLLC circuit model predictor outputs the output voltage value and resonant current peak value of the next cycle to obtain the true optimal switching frequency value; Step 3: Build a real-time lightweight network as the optimal switching frequency controller, use the load factor, voltage gain and resonant current peak as the input of the lightweight network, and output the predicted optimal switching frequency value; Step 4: Combine the optimal switching frequency controller with the PI controller to achieve smooth soft start of the CLLC resonant converter.

2. The CLLC resonant converter soft starting method based on edge neural network prediction according to claim 1, characterized in that: The step 1 is specifically implemented as follows: using a high-precision operational amplifier to collect data within the startup cycle of the CLLC resonant converter under m groups of different load conditions; each group of loads differs by 2%, covering the entire working range from no-load to full-load; and each group contains n control cycles, recording the switching frequency, output voltage and resonant current in each cycle.

3. The CLLC resonant converter soft starting method based on edge neural network prediction according to claim 2 is characterized in that: The specific implementation process of step 2 is as follows: Constructing a CLLC circuit model predictor, including a sequentially connected long short-term memory (LSTM) recursive neural network, a transposed matrix, a one-dimensional data expansion, a Softmax activation function, a regularization technique, and a fully connected output layer; The transposed matrix operation is performed on the second and third dimensions of the data; the fully connected layer maps the features to two output nodes to complete the prediction task; The fully connected layer obtains the output voltage and resonant current peak of the next cycle of the CLLC resonant converter; constrains its resonant current peak and output voltage to obtain the soft start process, and its formula is: V o (k+1)<V ref I pk (k+1)=I max Where V o (k+1) represents the voltage value of the next moment output by the predictor, V ref Represents the target voltage value; I pk (k+1) represents the next moment resonant current peak value of the predictor output, I max Represents the maximum value of the resonant current peak; through the constraint conditions, the optimal switching frequency value is obtained and the result is saved.

4. The CLLC resonant converter soft starting method based on edge neural network prediction according to claim 3 is characterized in that: The specific implementation process of the real-time lightweight network is as follows: The load factor and voltage gain calculation formula are as follows: Where k represents the current moment, n represents the transformer turns ratio, g(k) represents the load factor at the current moment, and I o (k) represents the output current value at the current moment, V o (k) represents the output voltage value at the current moment, M(k) represents the voltage gain value at the current moment, V i (k) represents the input voltage value at the current moment; Build a real-time lightweight network as the optimal switching frequency controller, and use the load factor, voltage gain and resonant current peak as lightweight network inputs; The lightweight network architecture consists of multiple residual convolutional units and fully connected layers, where features are first extracted and learned by several residual convolutional units, and then passed to the fully connected layer to generate output, which is the predicted optimal switching frequency value; Each residual convolution unit consists of two branches in parallel. One branch consists of three layers of dilated convolution layers with different dilation rates in sequence, and the second branch consists of one layer of one-dimensional convolution layer. The outputs of the two branches are added together to obtain the output of the residual convolution unit. The step size of the first layer of dilated convolution is set to k, while the step size of the other two layers of dilated convolution is 1, and the dilation rates are 1, n, and n 2 ; The fully connected layer maps the output features of the residual convolution unit to the output result, obtains the optimal switching frequency, predicts the optimal switching frequency value, and performs reverse training based on the actual optimal switching frequency value; and obtains the optimal switching frequency controller of the CLLC circuit.

5. The CLLC resonant converter soft starting method based on edge neural network prediction according to claim 4 is characterized in that: The specific implementation process of step 4 is as follows: First determine the initial pulse width, assuming that the output voltage remains unchanged within a cycle, and derive the equation as follows: Ir1c(θ)=0.5(sin(θ)+k1 sin(k1θ)) Where Lr represents the resonant inductance, Lm represents the transformer excitation inductance, I r1c (θ) represents the first cycle of the resonant cavity current waveform; It can be seen from the equation that the maximum resonant current peak is obtained at one-quarter of the resonant period, so the initial pulse width should be greater than twice the resonant frequency. Here, the initial switching frequency is selected as three times the resonant frequency to suppress the resonant current peak; then the output voltage data and output current data are collected, and the load factor and voltage gain are calculated; the load factor, voltage gain, and resonant current peak are input into the optimal switching frequency controller, and its output is the optimal control frequency to achieve the increase in the output voltage of the CLLC resonant converter; the output voltage is increased to within 5% of the target voltage through the optimal switching frequency controller, and the increased output voltage is switched to the PI controller for control to achieve a smooth voltage transition and complete the entire soft start process.

6. The CLLC resonant converter soft starting method based on edge neural network prediction according to claim 5, characterized in that: The target voltage range formula is: Where μ(k) is the percentage of the target voltage range, V ref To set the target voltage, V o (k) is the output voltage value at time k; The PI controller formula is as follows: u(k+1)=u(k)+Kp(e(k)-e(k-1))+Kie(t) Where k represents the current moment, u(k) represents the switching frequency at the current moment, u(k-1) represents the switching frequency of the next cycle, e(k) represents the voltage error at the current moment, e(k-1) represents the voltage error of the previous cycle, K p is the proportionality coefficient, K i is the integration coefficient.

Images