Energy storage converter predictive current control method based on data driving model self-learning

Abstract

The invention discloses an energy storage converter predictive current control method based on data driving model self-learning, and belongs to the technical field of power electronic converter control. The energy storage converter is a half-bridge bidirectional DC-DC converter, and the method comprises the following steps: collecting the state quantity of the energy storage converter through a sensor in each control period, forming a data driving prediction model constructed by an input vector input neural network online parameter identification module, and outputting an initial duty ratio prediction value; constructing a value function, performing online optimization on an initial value through a gradient descent method to obtain a final optimal duty ratio, and simplifying a scene to directly use the initial value; comparing the optimal duty ratio with a carrier wave to generate a PWM (Pulse Width Modulation) signal to drive a power switch tube; actual inductive current is collected, an ideal duty ratio enabling the value function to be minimum is reversely solved to serve as a supervision signal, and the model weight is trained and updated online through a sample. The technical problem that traditional continuous set model predictive control depends on an accurate mathematical model is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power electronic converter control, and particularly relates to a predictive current control method for energy storage converters based on data-driven model self-learning. BACKGROUND

[0002] In a battery energy storage system, an energy storage converter (bidirectional DC-DC converter) is a key energy management unit connecting a battery and a DC bus, responsible for implementing the charge and discharge control of the battery. The energy storage converter (bidirectional DC-DC converter) is a core component of the energy storage system, and its performance directly affects the energy conversion efficiency, dynamic performance and battery life of the system. Continuous set model predictive control (CCS-MPC) has great application potential in energy storage converter control due to its fixed switching frequency, good steady-state performance and ability to handle multiple constraints. However, the traditional CCS-MPC relies heavily on an accurate mathematical model. In actual applications, system parameters may change due to magnetic saturation effects, temperature changes, device aging, etc. This model mismatch can cause the prediction accuracy of the traditional CCS-MPC to decrease, leading to steady-state errors, current waveform distortion (THD increase), dynamic performance deterioration, and even system instability in extreme cases.

[0003] To solve the model mismatch problem, there are mainly two types of existing technologies: one is to use robust control design, but usually at the expense of dynamic performance; the other is to introduce a disturbance observer for feedforward compensation, but this is still a patch under a fixed model framework and cannot fully adapt to long-term gradual changes and nonlinear changes in parameters. Therefore, there is an urgent need for a predictive control method that can adapt to system dynamic changes, so that the controller can adjust the control strategy according to the real-time state of the controlled object, thereby maintaining optimal performance throughout its life cycle. SUMMARY

[0004] The purpose of the present application is to provide a predictive current control method for energy storage converters based on data-driven model self-learning, which solves the technical problem of traditional continuous set model predictive control relying on accurate mathematical models.

[0005] The technical solution adopted by the present application is a predictive current control method for energy storage converters based on data-driven model self-learning, wherein the energy storage converter is a half-bridge bidirectional DC-DC converter, comprising the following steps: Step 1: In each control cycle, the state quantities of the energy storage converter are collected by a sensor; Step 2: The state quantities collected in Step 1 are combined into an input vector, which is input into a data-driven predictive model constructed based on an online parameter identification module of a neural network, and the output vector is an initial duty cycle prediction value; Step 3: Construct a value function, using the initial duty cycle prediction value as the initial value, and use gradient descent to optimize the value function online to obtain the final optimal duty cycle; or, in a simplified scenario, use the initial duty cycle prediction value as the final optimal duty cycle. Step 4: Compare the final optimal duty cycle with the carrier wave to generate the corresponding PWM switching signal, which drives the power switching transistors of the half-bridge bidirectional DC-DC converter; Step 5: Collect the actual inductor current, reverse the calculation to find the ideal duty cycle that minimizes the value function as the supervision signal, and use the sample composed of the input vector and the ideal duty cycle to train and update the network weights of the data-driven prediction model online.

[0006] The invention is further characterized by: The state variables in step 1 include: the current inductor current of the energy storage converter. Battery side voltage and DC bus voltage It also includes the duty cycle of the previous cycle. and the inductor current command for the next cycle ; Input vector in step 2 for:

[0007] Output vector Initial duty cycle prediction value for: , .

[0008] The specific structure of the online parameter identification module based on the neural network in step 2 is as follows: the input layer has 5 nodes, the hidden layer has 1 layer with 8 neurons, and the output layer has 1 node; the activation function of the hidden layer is the non-linear Tanh function, and the activation function of the output layer is the linear Purelin function.

[0009] Value function in step 3 The expression is:

[0010] in: It is a traditional physical model based on fixed parameters, according to the candidate duty cycle. The calculated predicted value of the inductor current for the next control cycle.

[0011] In step 3, when using the gradient descent method to optimize the value function online, only 1-2 iterations are performed to obtain the final optimal duty cycle. .

[0012] The simplified scenario in step 3 is as follows: In operating conditions where system parameters change gradually and the current control accuracy requirement is lower than a preset threshold, the gradient descent optimization step is omitted, and the current is directly... As the final optimal duty cycle.

[0013] Step 4 involves comparing the final optimal duty cycle with the carrier wave as follows: Using a triangular wave as the carrier, the DC modulation wave corresponding to the final optimal duty cycle is compared with the triangular wave carrier. When the amplitude of the modulation wave is greater than the amplitude of the carrier wave, a high level is output, and vice versa. The generated PWM switching signal drives the corresponding power switch according to the converter's operating mode.

[0014] Step 5 involves collecting the actual inductor current as follows: The actual inductor current of the system is collected at all times. ; The ideal duty cycle that minimizes the value function is specifically determined by: using the value function Minimize as the objective, solve in reverse The ideal duty cycle to be applied at any given time .

[0015] The specific steps for training and updating the network weights of the data-driven prediction model online in step 5 are as follows: The input vector in step 2 Compared with the ideal duty cycle obtained by the solution Composition of training samples Substituting the values ​​into the neural network of the data-driven prediction model, the weights and biases of the network are adjusted through the backpropagation algorithm to make the model output... Approaching step by step This enables adaptive compensation for system parameter drift and nonlinear dynamics.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention inherits the advantages of Continuous Set Model Predictive Control (CCS-MPC) with its fixed switching frequency and low current ripple, while overcoming its sensitivity to model parameters through a data-driven model. It can still achieve zero steady-state error current tracking even under conditions of inductor saturation and battery voltage variations. Furthermore, the data-driven model can learn online and compensate for system parameter drift and unmodeled dynamics, making the controller insensitive to disturbances such as component aging and temperature changes, significantly improving the long-term stability of the system. Simultaneously, using the output of the data-driven model as high-quality initial values ​​for numerical optimization greatly accelerates the convergence speed of the optimization algorithm, effectively solving the problem of high online computational load in CCS-MPC, and enabling its real-time application and industrialization on low-cost processors. Moreover, it eliminates the need for complex offline measurements and modeling of inductor saturation curves and battery internal resistance, allowing the system to self-optimize during operation, significantly reducing development and maintenance costs. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the predictive current control method for energy storage converters based on data-driven model self-learning, as described in this invention. Figure 2 This is a schematic diagram of the working principle of the online parameter identification module based on neural network in the data-driven model self-learning predictive current control method for energy storage converters of the present invention. Figure 3 This is a schematic diagram of the model update process in the data-driven model self-learning-based predictive current control method for energy storage converters of the present invention. Figure 4 This is a schematic diagram of the simulated inductor current waveform when the inductor parameters of the energy storage converter change in Boost mode, which is part of the predictive current control method for energy storage converter based on data-driven model self-learning of the present invention. Figure 5 This is a schematic diagram of the simulated inductor current waveform when the inductor parameters of the energy storage converter change in Buck mode, which is part of the predictive current control method for energy storage converters based on data-driven model self-learning of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] like Figure 1 As shown, the energy storage converter of this invention adopts a bidirectional half-bridge topology. Wherein, U dc and U b These represent the DC bus voltage and the battery terminal voltage, respectively. C Capacitors supporting the DC bus. L S1 represents the converter inductor; S2 represents the IGBT switching transistors. The bidirectional DC-DC converter has two main operating modes: Boost mode and Buck mode. When the bidirectional DC-DC converter operates in Boost mode, transistor S1 is always off, and transistor S2 is on; when the bidirectional DC-DC converter operates in Buck mode, transistor S1 is on, and transistor S2 is always off.

[0020] This invention employs a typical cascaded control structure where the inner current loop is nested below the voltage loop. Targeting the operating characteristics of bidirectional DC-DC converters, continuous-set predictive current controllers are designed for both Boost and Buck operating modes.Figure 1 This is a dual-closed-loop overall control block diagram of the bidirectional DC-DC converter system of the present invention, and its specific process is as follows: 1. Determine the desired DC bus voltage and the current feedback DC bus voltage of the system The difference is calculated, and the resulting voltage error signal is adjusted by a voltage loop PI controller to obtain the inductor current reference signal. ; 2. A data-driven neural network model is used to implement continuous-set model predictive control of the bidirectional DC-DC converter. The implementation process is as follows: Figure 2 As shown.

[0021] Example 1 The present invention discloses a predictive current control method for energy storage converters based on data-driven model self-learning, wherein the energy storage converter is a half-bridge bidirectional DC-DC converter, and includes the following steps: Step 1: In each control cycle, the state variables of the energy storage converter under different operating conditions are collected by the sensors. The obtained experimental data are organized into a data table, which is used as the dataset for neural network training. Step 2: The state variables collected in Step 1 are combined into an input vector and input into the data-driven prediction model built by the online parameter identification module based on the neural network (dynamically identifying the input-output relationship that best reflects the current dynamic behavior of the system). The neural network can then calculate and predict the initial duty cycle prediction value (output vector) based on the current weights and biases. Step 3: Construct a value function with current tracking error as the objective. Using the initial duty cycle prediction value as the initial value, use the gradient descent method to quickly optimize the value function online and find the final optimal duty cycle that minimizes the value function; or, in a simplified scenario, the output of the data-driven model (initial duty cycle prediction value) can be directly used as the final optimal duty cycle, skipping the numerical optimization process. Step 4: Compare the final optimal duty cycle with the carrier wave to generate the corresponding PWM switching signal, which drives the power switching transistors of the half-bridge bidirectional DC-DC converter; Step 5: Collect the actual inductor current, reverse the calculation to find the ideal duty cycle that minimizes the value function as the supervision signal, and use the sample composed of the input vector and the ideal duty cycle to train and update the network weights of the data-driven prediction model online.

[0022] In practical applications, a Hall current sensor (model: ACS712) is used to collect inductor current, and a voltage sensor (model: LV25-P) is used to collect battery-side voltage and DC bus voltage. The control cycle is set to 50μs (which meets the conventional switching frequency requirements of energy storage converters). The initial weights of the neural network are initialized using the Xavier initialization method, and the initial bias value is set to 0. Before model training, 1000 sets of typical operating condition data collected offline (covering Boost / Buck mode, different inductance / voltage values) are used for pre-training to ensure that the initial duty cycle prediction error is ≤5% and to solve the problem of ambiguity in the initial state of the model. The learning rate of the gradient descent method is set to 0.001, and the iteration step size is fixed at 0.005. The reason for choosing one iteration is that within the 50μs control cycle, one iteration can control the calculation time within 10μs, which meets the real-time requirements, and the current tracking error is ≤0.2A, which meets the accuracy requirements of conventional energy storage operating conditions.

[0023] Example 2 Based on Example 1, the state variables in step 1 include: the current inductor current of the energy storage converter. Battery side voltage and DC bus voltage It also includes the duty cycle of the previous cycle. and the inductor current command for the next cycle Among them, the inductor current command for the next cycle The output of the outer loop controller is used to obtain the data.

[0024] The inductor current command for the next cycle is output by the voltage loop PI controller, and the PI controller parameter is the proportional coefficient. K p =0.8, integral coefficient K i =0.05; when the DC bus voltage error u dcref - u dc When the voltage is greater than 2V, the PI controller output is limited to 1.2 times the rated current to avoid overcurrent; the inductor current acquisition accuracy is ±0.1A, and the battery side voltage and DC bus voltage acquisition accuracy is ±0.5V; the acquired signal needs to pass through a second-order Butterworth low-pass filter (cutoff frequency 1kHz) to filter out switching noise interference.

[0025] Example 3 Based on Example 2, the input vector in step 2 for:

[0026] Output vector Initial duty cycle prediction value for: ,

[0027] In the above formula, The duty cycle of the previous cycle is used as the model input to reflect system dynamics.

[0028] Furthermore, the specific structure of the online parameter identification module based on the neural network in step 2 is as follows: the input layer has 5 nodes, the hidden layer has 1 layer with 8 neurons, and the output layer has 1 node; the activation function of the hidden layer adopts the nonlinear Tanh function, and the activation function of the output layer adopts the linear Purelin function.

[0029] The input layer has 5 nodes corresponding to 5 state variables. Simulation results show that the model fitting error is minimized (≤3%) when the number of hidden layer neurons is in the range of 6-10, with 8 neurons, and the computational load is moderate (single cycle computation time ≤8 seconds). μs The output layer has one node with a corresponding duty cycle (range 0-1), which meets the converter control requirements. The input vector needs to be min-max normalized, with a normalization interval of [-1, 1] (e.g., inductor current). The range is 0-50A, and the normalization formula is: The output layer duty cycle needs to be constrained between 0 and 1 by a limiting function to avoid exceeding the driving range of the switching transistor; compared with the sigmoid function, the hidden layer Tanh function reduces the model identification error by 40% in system parameter drift scenarios (such as a 20% change in inductance value); the output layer Purelin function can ensure continuous output of the duty cycle and avoid the increase in current ripple caused by discrete output.

[0030] Example 4 Based on Example 1, the value function in step 3 The expression is:

[0031] in: It is a traditional physical model based on fixed parameters, according to the candidate duty cycle. The calculated predicted value of the inductor current for the next control cycle.

[0032] Furthermore, in step 3, when using the gradient descent method to optimize the value function online, only 1-2 iterations are performed to obtain the final optimal duty cycle. .

[0033] Furthermore, the scenario in step 3 is simplified as follows: In operating conditions where system parameters change gradually and the current control accuracy requirement is lower than a preset threshold, the gradient descent optimization step is omitted, and the current is directly... As the final optimal duty cycle.

[0034] value function The core function is to minimize current tracking error, corresponding to the technical effect of 'zero steady-state error current tracking' in the claims; in practical applications, setting J A value of ≤0.04 (i.e., current tracking error ≤0.2A) is considered a qualified threshold. If the value is exceeded, the gradient descent method will be triggered for optimization.

[0035] Simplified scenario judgment criteria: ① Current control accuracy requirement ≤ ±5% (e.g., during the standby charging phase of an energy storage system, the load current fluctuation is small); ② System parameter change rate ≤ 0.1% / s (e.g., the battery voltage slowly decreases from 48V to 46V in 200s; ambient temperature gradient ≤ 2℃ / min); In this case, after omitting the optimization step, the current ripple is still ≤ 0.5A, which meets the operating conditions.

[0036] Comparative experimental data: In the first iteration, the current tracking error was 0.15A and the calculation time was 8μs; in the second iteration, the error was 0.1A and the calculation time was 15μs. For normal operating conditions (such as energy storage charging and discharging power ≤50kW), one iteration is sufficient, while for high-precision operating conditions (such as grid-side grid connection control, current THD requirement ≤3%), two iterations are required.

[0037] Example 5 Based on Example 1, step 4 specifically compares the final optimal duty cycle with the carrier wave as follows: Using a triangular wave as the carrier, the DC modulation wave corresponding to the final optimal duty cycle is compared with the triangular wave carrier. When the amplitude of the modulation wave is greater than the amplitude of the carrier wave, a high level is output, and vice versa. The generated PWM switching signal drives the corresponding power switch according to the converter's operating mode.

[0038] The triangular wave carrier frequency is set to 20kHz (matching the switching frequency of the energy storage converter), with an amplitude range of 0-1. The DC modulation wave amplitude corresponding to the optimal duty cycle is equal to the duty cycle value (e.g., duty cycle 0.6, modulation wave amplitude 0.6). The drive signal is matched according to the converter's operating mode: ① Boost mode (battery discharging to the DC bus): switch S1 is always off, S2 receives the PWM signal (high level for conduction, low level for off); ② Buck mode (DC bus charging to the battery): switch S2 is always off, S1 receives the PWM signal; a 1μs delay (dead time) is required when it is turned on to avoid shoot-through between the upper and lower transistors. Taking the optimal duty cycle of 0.6 as an example, with a 20kHz carrier period of 50μs and a modulation wave amplitude of 0.6: when the carrier value < 0.6, a high level is output (lasting 30μs); when the carrier value > 0.6, a low level is output (lasting 20μs), and the generated square wave signal directly drives the IGBT switch.

[0039] Example 6 like Figure 3 As shown, based on Example 1, the specific steps for collecting the actual inductor current in step 5 are as follows: The actual inductor current of the system is collected at all times. ; The ideal duty cycle that minimizes the value function is specifically determined by: using the value function Minimize as the objective, solve in reverse The ideal duty cycle to be applied at any given time .

[0040] Furthermore, the online training and updating of the network weights of the data-driven prediction model in step 5 specifically involves: taking the input vector from step 2... Compared with the ideal duty cycle obtained by the solution Composition of training samples Substituting the values ​​into the neural network of the data-driven prediction model, the weights and biases of the network are adjusted through the backpropagation algorithm to make the model output... Approaching step by step This enables adaptive compensation for system parameter drift and nonlinear dynamics.

[0041] Network weight update frequency: Updated once every 10 control cycles (500μs) to avoid frequent updates that could cause model oscillations; Batch gradient descent is used during training, with 10 sets of samples used for parameter adjustment each time.

[0042] The bidirectional DC-DC converter has two operating modes: Boost and Buck. The corresponding mathematical models for the two operating modes are different. Therefore, it is necessary to design predictive current controllers for the two modes separately and calculate the predicted inductor current based on the actual model.

[0043] 1. Boost Mode Assuming that the conduction time of switch S2 is t1 within one switching cycle Ts, from the equation... After a first-order forward difference approximation, we can obtain: (1) in, i L ( k The current is the inductor current of the switching transistor S2 at the start of the current cycle (i.e., measured in this cycle). i Lo ( k ) is the inductor current of the switching transistor S2 when the current conduction ends. U b ( k () is the current battery voltage.

[0044] Assuming that within one control cycle Ts, the turn-off time of switch S2 is t2, and t1 + t2 = Ts, from the equation... After a first-order forward difference approximation, we can obtain: (2) in, It is the inductor current of the switching transistor S2 at the end of the current turn-off (i.e., the current at the next moment). U dc ( k ) is the DC bus voltage at the current moment.

[0045] Within one switching cycle Ts, from equations (1) and (2), we can obtain: (3) Define the duty cycle in Boost operating mode The inductor current at this duty cycle can be calculated using equation (3). .

[0046] (4) This duty cycle is the modulation wave of the predictive current controller in Boost mode. By comparing it with the carrier wave, a corresponding PWM switching signal is generated, which drives the power switching transistors of the half-bridge converter.

[0047] 2. Buck Mode Assuming that the conduction time of switch S1 is t3 within one switching cycle Ts, from the equation... After a first-order forward difference approximation, we can obtain: (5) in, i L ( k The current is the inductor current of the switching transistor S1 at the start of the current cycle (i.e., measured in this cycle). i Lo ( k ) is the inductor current of the switching transistor S1 at the end of the current conduction cycle. U dc ( k () is the DC bus voltage at the current moment. U b ( k () is the current battery voltage.

[0048] Assuming that within one control cycle Ts, the turn-off time of switch S1 is t4, and t3 + t4 = Ts, from the equation... After a first-order forward difference approximation, we can obtain: (6) in, It is the inductor current of the switching transistor S1 at the end of the current turn-off (i.e., the current at the next moment).

[0049] Within one switching cycle Ts, from equations (5) and (6), we can obtain: (7) Define the duty cycle in Buck working mode The inductor current at this duty cycle can be calculated using equation (7). .

[0050] (8) in, d Buck The modulation waveform for the predictive current controller of a bidirectional DC-DC converter operating in Buck mode.

[0051] A simulation model of an independently operating photovoltaic energy storage system using batteries as the energy storage medium was built in the MATLAB / Simulink environment. The simulation parameters were adjusted to achieve satisfactory control results. Figure 4 and Figure 5 The simulation comparison demonstrates the advantages of this method in Boost and Buck modes. Figure 4 The left side shows the actual output inductor current waveform based on NN+CCS-MPC. Figure 4 The right side shows the inductor current waveform based on CCS-MPC; Figure 5 The left side shows the actual output inductor current waveform based on NN+CCS-MPC. Figure 5 The right side shows the inductor current waveform based on CCS-MPC. In the simulation, the controller parameters... L The inductance in the circuit is always 2mH from 0s to 0.1s. L =4mH, 0.1s~0.2s inductance in the circuit L =2mH, 0.2s~0.3s inductance in the circuit L =1mH, 0.3s~0.4s inductance in the circuit L =0.66mH. The current waveform of traditional CCS-MPC (using a fixed 2mH model) gradually exhibits amplitude errors and phase lag, resulting in a significant increase in THD. However, the method proposed in this invention, with its online identification module dynamically updating the prediction model, maintains accurate current tracking and excellent waveform quality throughout, with virtually no change in THD.

[0052] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting current control of an energy storage converter based on self-learning of a data-driven model, the energy storage converter being a half-bridge bidirectional DC-DC converter, characterized in that, The method comprises the following steps: Step 1: collecting state variables of the energy storage converter through a sensor in each control cycle; Step 2: inputting the state variables collected in step 1 into an input vector, inputting the input vector into a data-driven prediction model constructed based on an online parameter identification module of a neural network, and outputting an initial duty cycle prediction value; Step 3: constructing a value function, taking the initial duty cycle prediction value as an initial value, and using a gradient descent method to perform online optimization on the value function to obtain a final optimal duty cycle; or, in a simplified scenario, taking the initial duty cycle prediction value as the final optimal duty cycle; Step 4: comparing the final optimal duty cycle with a carrier to generate a corresponding PWM switching signal to drive a power switch tube of the half-bridge bidirectional DC-DC converter; Step 5: collecting an actual inductor current, inversely calculating an ideal duty cycle that minimizes the value function as a supervision signal, using a sample composed of the input vector and the ideal duty cycle to perform online training and update network weights of the data-driven prediction model. The specific structure of the online parameter identification module based on the neural network in step 2 is that: the input layer has 5 nodes, the hidden layer has one layer and contains 8 neurons, and the output layer has one node; the hidden layer activation function adopts a nonlinear Tanh function, and the output layer activation function adopts a linear Purelin function.

2. The data-driven model based self-learning energy storage converter predictive current control method of claim 1, wherein, The state quantities described in step 1 include the current inductance current of the energy storage converter , the battery-side voltage , and the DC bus voltage , also including the previous cycle duty ratio and the next cycle inductance current command .

3. The data-driven model based self-learning energy storage converter predictive current control method of claim 2, wherein, the input vector described in step 2 is: Output vector Initial duty cycle prediction value is: , 。 4. The data-driven model based self-learning energy storage converter predictive current control method of claim 1, wherein, The simplified scenario in step 3 is that:

5. The data-driven model based self-learning energy storage converter predictive current control method of claim 1, wherein, The value function described in step 3 The expression for the value function is: wherein: is a next control cycle inductance current prediction value calculated based on a conventional physical model based on fixed parameters according to the candidate duty cycle ​ 6. The data-driven model based self-learning energy storage converter predictive current control method of claim 5, wherein, In step 3, the gradient descent method is used to optimize the value function online, and only 1-2 iterations are performed to obtain the final optimal duty cycle .

7. The data-driven model based self-learning energy storage converter predictive current control method of claim 5, wherein, The comparison of the final optimal duty cycle with the carrier in step 4 is specifically that: The system parameter changes gently, and the current control accuracy requirement is lower than the preset threshold value, at this time, the gradient descent method optimization step is omitted, and the As the final optimal duty cycle.

8. The data-driven model based self-learning energy storage converter predictive current control method of claim 1, wherein, A triangular wave is used as a carrier, a direct current modulation wave corresponding to the final optimal duty cycle is compared with the triangular wave carrier, a high level is output when the modulation wave amplitude is greater than the carrier amplitude, and a low level is output otherwise; the generated PWM switching signal drives the corresponding power switch tube according to the working mode of the converter. The online training and updating of the network weights of the data-driven prediction model in step 5 are specifically that: 9.The data-driven model based self-learning energy transformer predictive current control method of claim 1, wherein, The actual inductor current is collected in Step 5, specifically: At the moment, the actual inductor current of the system is collected ; The ideal duty cycle that minimizes the value function is found by reversing the process and is given by minimizing the value function the ideal duty cycle that should be applied at time .

10. The data-driven model based self-learning energy storage converter predictive current control method of claim 7, wherein, ​ The input vector in step 2 is The ideal duty cycle obtained by solving The training sample is composed of The neural network of the data-driven prediction model is substituted into the data, and the weights and biases of the network are adjusted through the back propagation algorithm, so that the output of the model Step by step Adaptive compensation of system parameter drift and nonlinear dynamics is realized.

Images