LCL type single-phase grid-connected inverter compound control method based on neural network

Abstract

The invention belongs to the related technical field of inverter control, and discloses an LCL type single-phase grid-connected inverter compound control method based on a neural network. Comprising the following steps: establishing a mathematical model according to a circuit structure of the single-phase LCL type grid-connected inverter; setting a given quantity of a control target, and substituting the given quantity into the mathematical model to obtain a control variable in a steady state; a composite neural network controller is constructed, the composite neural network controller comprises a radial basis function neural network (RBFNN) and a full-connection neural network (FCNN), a basic controller is constructed, the input of the basic controller is an error of a control target, and the output of the basic controller is a control variable; and forming a control variable M together; calculating and updating neural network parameters at a preset learning rate by taking the minimum error as an objective function; and controlling the LCL type single-phase grid-connected inverter by using the output control variable M. According to the composite architecture of'training, namely control ', accurate capture of dynamic characteristics and adaptive adjustment of system parameters are realized, and the control precision, robustness and training efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the technical field of inverter control, and more specifically, relates to a composite control method for LCL type single-phase grid-connected inverters based on neural networks. Background Technology

[0002] Inverters are an important component of new energy fields such as photovoltaic power generation and wind power generation. Their control technology is key to obtaining high-quality power and ensuring stable system operation.

[0003] Currently, grid-connected inverters are classified into L-type and LCL-type based on their filters. Compared to L-type filters, LCL-type filters have a stronger ability to suppress harmonic currents at the inverter's switching frequency. Under the same conditions, LCL-type filters are smaller and less expensive than L-type filters. LCL-type grid-connected inverters are currently receiving widespread attention due to their superior high-frequency harmonic suppression capabilities. Excellent grid-connected inverter products are constantly emerging, especially with significant breakthroughs in maximum power point tracking, parallel operation of the inverter stage, and fully digital control methods. Grid-connected inverters with capacities ranging from 1kW to 20MW are widely used in various fields such as communications, lighting, agricultural irrigation, marine and aviation, and public power. Companies from Germany, the United States, and Japan, such as Siemens, ASMA, Sun Power, and Fronius, dominate the international market.

[0004] In academia, both domestic and international research has long focused on improving the efficiency, stability, and reliability of grid-connected inverters. For example, some studies combine Lyapunov control theory with artificial neural networks (ANNs), using the artificial bee colony (ABC) algorithm to train the ANN offline to optimize the positive definite weighting factor, thereby improving the system's dynamic response and reducing harmonic distortion (THD < 2.5%). However, this method relies on complex offline optimization loops, and the ANN only plays the role of a parameter estimator in control, without directly participating in the generation of control signals, thus limiting its adaptive capability and real-time performance. Other researchers have proposed a bypass capacitor-based suppression scheme to address the inherent leakage current problem in transformerless photovoltaic systems, effectively reducing the peak leakage current from 5A to 22.5mA. This scheme is a passive hardware optimization; while it improves electromagnetic compatibility, it does not address the core of the control algorithm and cannot solve the control performance bottleneck of inverters under complex grid conditions. Furthermore, to overcome the performance degradation of PR controllers during grid frequency fluctuations, some studies use dq coordinate transformation to convert AC quantities to DC quantities, and then utilize a PI controller to achieve theoretically zero steady-state error control. However, this method introduces modules such as SOGI to construct orthogonal signals, which increases the system complexity and the difficulty of phase control. Furthermore, the PI controller's ability to suppress sudden disturbances and harmonics on the AC side in the dq coordinate system remains limited.

[0005] The above studies reflect a common limitation of current control strategies: they either rely on complex offline optimization and manual intervention, or compensate for the inherent defects of the controller by increasing system complexity, failing to achieve a simplified intelligent control core that can learn online and comprehensively take into account both dynamic and steady-state performance.

[0006] The performance of an inverter fundamentally depends on its control strategy. Traditional methods each have their bottlenecks: hysteresis control has an unfixed switching frequency, which makes filter design difficult; although PI control can track DC signals without steady-state error in a synchronous rotating coordinate system, there is an inherent contradiction between its dynamic response and stability margin, and its ability to suppress AC side harmonics is insufficient; deadbeat control heavily relies on an accurate system model. The inductor and capacitor parameters in the LCL filter can be perturbed by factors such as aging and temperature rise during actual operation, leading to model mismatch and causing system stability problems. In fact, simulations have shown that when the filter capacitor C increases from 10μF to 12μF, the system exhibits significant oscillation and instability; although the PR controller can achieve zero steady-state error tracking of a specific frequency signal, its performance is highly dependent on the stability of the grid frequency. Literature clearly points out that when the grid frequency deviates, the gain of the PR controller will decrease sharply, leading to a rapid decline in control performance.

[0007] More importantly, these traditional controllers are designed based on linear and steady-state assumptions, and their parameters remain fixed once tuned. However, the actual operating environment of grid-connected systems is complex and variable, facing challenges such as sudden changes in the amplitude and phase of grid voltage (caused by large-capacity load switching or faults), background harmonic pollution, and time-varying parameters (such as LCL filter parameter drift). Under these nonlinear and strongly coupled dynamic conditions, traditional controllers, lacking online self-learning and adaptive adjustment capabilities, often exhibit slow response, decreased regulation accuracy, and even the risk of system instability, severely restricting the advanced application of grid-connected inverters in new power systems.

[0008] Currently, researchers both domestically and internationally have proposed several control methods for single-phase grid-connected inverters based on neural networks. Neural networks are functions with powerful fitting capabilities; if properly designed and trained, they can fit any existing controller. Reinforcement learning is gaining attention in neural network training. It's a learning method that guides actions based on the perceived state of the environment to achieve better results and maximize benefits, and can be used to automate complex tasks. In recent years, reinforcement learning has shown outstanding performance in complex decision-making problems such as games and autonomous driving, reaching or even surpassing human decision-making levels. Meanwhile, some scholars have proposed power system safety correction control methods based on reinforcement learning, demonstrating its feasibility in circuit control. However, the trial-and-error process in single reinforcement learning is highly uncertain, potentially damaging power electronic converters and posing excessive risks. Furthermore, traditional controller parameters are fixed, and model-free reinforcement learning incurs high trial-and-error costs, making it difficult to cope with complex grid conditions. Compared to traditional control methods, neural network-based control methods offer greater flexibility, stronger adaptability, and can meet the needs of inverters in different application scenarios. They also possess a certain degree of self-learning capability, reducing the workload of designers. Radial basis function neural networks (RBFNNs) are a superior method. However, neural network controllers still have the following problems: the output of the neural network lacks labels. Traditional methods often directly define output labels. On the one hand, the selection of labels is often complex and numerous circuit information or intermediate variables with unclear physical meaning. Collecting these labels from simulation or experiment is time-consuming and laborious, leading to a sharp increase in training costs and difficulty in convergence. On the other hand, these selected labels cannot be adapted to working conditions outside the collected dataset, which reduces the adaptability of the controller and fails to give full play to the self-learning ability of the neural network controller. Summary of the Invention

[0009] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a composite control method for LCL-type single-phase grid-connected inverters based on neural networks. Its purpose is to solve the technical problems of traditional PR controllers, which struggle to balance dynamic response and steady-state accuracy, and the need for multiple resonant units in parallel to suppress grid harmonics, resulting in complex controller structures and cumbersome parameter tuning. Furthermore, directly applying neural networks to inverter control faces convergence difficulties due to a lack of clear training labels. Additionally, the complex network structure and high computational load make real-time driving difficult in embedded power chips, while model-less trial-and-error learning inherently suffers from stability risks and poor real-time performance.

[0010] To achieve the above objectives, this invention provides a composite control method for an LCL-type single-phase grid-connected inverter based on a neural network, comprising the following steps: (1) A mathematical model is established based on the basic characteristics of a single-phase LCL grid-connected inverter, where X is the state variable in the circuit, M is the circuit control variable, U is the grid voltage, Z is the control target, i.e., the grid-connected current, and C is the parameter matrix. (1) in: (2) This is the grid voltage. The magnitude of the grid voltage. This is the angular frequency of the grid voltage.

[0011] (2) Design the setpoint Z of the control objective according to the control requirements. * As shown in equation (3), when the LCL inverter is used as a power regulator, the reference current needs to include the fundamental component and several odd harmonic components.

[0012] (3) When the circuit enters steady state, the grid current equals the reference current. Substituting equations (2) and (3) into equation (1), we can derive the control variable m under steady state. ss The expression is shown in equation (4). In order to control the harmonics of each different frequency, m ss It needs to include in-phase and quadrature components of the corresponding frequency.

[0013] (4) This represents the amplitude of the in-phase component. The amplitude of the orthogonal component.

[0014] (3) From the analysis of equation (4) in step (2), it can be seen that the circuit control variable M under steady state needs to contain 2N components, of which N are in-phase components of each frequency component and the other N are quadrature components of each frequency component. The elements containing these 2N components are selected as the input quantities of the composite neural network controller, as shown in equations (5) and (6): (5) (6) Among them, in equation (6) g is and g ic It represents basic operations such as normalization.

[0015] (4) Select a suitable neural network based on the characteristics of the network input variables. The neural network input variables for the LCL single-phase grid-connected inverter targeted in this invention are all AC quantities. Therefore, a radial basis function neural network (RBFNN) with local characteristics, capable of handling AC quantities well, is selected. As shown in equation (7), each input variable... h is and h ic For a corresponding RBFNN, y is and y ic This is the output of the corresponding RBFNN. π α α represents the inherent structure of RBFNN, and α is the set of trainable parameters in RBFNN.

[0016] (7) (5) Further design a fully connected neural network (FCNN). Represents the inherent structure of RBFNN. β This is the trainable parameter set in FCNN, used to adjust and fuse the output of the preceding RBFNN. y is and y ic The input to FCNN is the result calculated in step (4). y is and y ic The output is the control quantity output by the composite neural network controller. .

[0017] In practice, to ensure the system can operate stably before the composite neural network controller is properly trained, a base controller is designed. Its input is the error e of the control target, and its output, together with the output of the composite neural network controller, constitutes the control variable M. Based on the basic controller function, Its parameter set; (8) (9) (10) (6) Design the objective function of the composite neural network controller according to the principle of minimizing error e. J ; (11) (7) Combining equations (8)(9)(10) in step (5) and equation (11) in step (6), the objective function is... J Find the partial derivatives of the neural network parameters; (12) in yes The collection; (8) Discretizing the partial derivatives involved in equation (12) facilitates subsequent updates of network parameters, enabling neural network parameter training within a single control cycle and achieving "training equals adjustment". However, due to the presence of delay components in the inverter, therefore... k Grid current at any time Z It is by k The output of the neural network at time -Q is determined by the value of the neural network. (13) (9) Calculate and update the neural network parameters according to equation (13). θ ,in η The learning rate; (14) The present invention also provides an electronic device, comprising: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the above-described method.

[0018] The present invention also provides a computer-readable storage medium storing computer instructions for causing a processor to perform the above-described method.

[0019] The present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the above-described method.

[0020] In summary, compared with the prior art, the technical solutions conceived in this invention have the following beneficial effects: (1) Based on the steady-state model of the LCL inverter, this invention constructs a composite neural network consisting of a radial basis function network (RBFNN) and a fully connected network. This design utilizes prior knowledge of the model to explicitly introduce the grid voltage and its orthogonal components as input, fundamentally avoiding the problems of intermediate variable labeling and trial-and-error training. The front-end RBFNN, with its local memory characteristics, accurately processes orthogonal signals and effectively suppresses periodic harmonics; the back-end FCNN fine-tunes and fuses the output of the RBFNN to generate the final modulated wave. This composite architecture of "training as control" achieves accurate capture of dynamic characteristics and adaptive adjustment of system parameters, significantly improving control accuracy, robustness, and training efficiency.

[0021] (2) Design of a composite neural network based on a system model: Under the framework of the composite neural network controller, the neural network part is further designed as a two-stage structure, with the front stage being a radial basis function neural network (RBFNN) and the back stage being a fully connected neural network (FCNN). The design process is based on the steady-state model of the system. According to the system model, under the condition of no error in steady state, the modulation amount of the inverter consists of a pair of orthogonal AC components. Therefore, the grid voltage and its orthogonal signal are selected as the input of the neural network, so that the neural network controller only needs to fine-tune these two inputs. Unlike the traditional "trial and error" input selection, this model-based input determination method significantly reduces the uncertainty and trial and error cost in the neural network training process. For the in-phase and orthogonal components of the input, two independent RBFNNs are constructed for processing. RBFNN has good locality, and its different nodes can correspond to different segments within a power frequency cycle, which is equivalent to having a "memory" function and will not have the problem of forgetting that ordinary neural networks easily do. With the help of this characteristic, RBFNN can eliminate periodic errors segment by segment in steady state, thereby playing a role similar to repetitive control and achieving effective suppression of interference signals such as harmonic components. After the RBFNN performs fine-tuning and shaping on the input orthogonal signals, it is known from the system model derivation that the amplitudes of the orthogonal components are not exactly equal. Therefore, a simple FCNN is used to fine-tune and fuse the amplitudes of the two signals processed by the RBFNN to generate the final modulated signal. The composite neural network is designed based on the system model, thus eliminating the need for trial and error and exploration, resulting in easy convergence, small network size, and simple computation.

[0022] (3) This invention proposes a system model-guided neural network training strategy. First, the entire control system, excluding the neural network, is modeled, allowing it to act as a bridge in backpropagation during training. In this way, the gradient of the reference signal can be transmitted to the neural network via the system model, thereby achieving weight updates under label-free conditions. During training, the neural network no longer relies on manually set target outputs but automatically obtains training signals through the comparison between the system model and the reference signal. The system model is embedded in the training path of the neural network, making each training iteration equivalent to an error feedback adjustment, essentially equivalent to embedding the idea of ​​Model Predictive Control (MPC). However, unlike MPC, this method does not require explicitly constructing complex optimization problems but achieves "training as adjustment" through online parameter updates of the neural network. Since the training process is based on the system model, it has the advantages of high training accuracy and fast convergence speed.

[0023] (4) Since the structure design of the composite neural network is based on the system model, its scale is small and its computational complexity is low, enabling real-time computation on embedded hardware such as digital signal processors (DSPs). Meanwhile, the training algorithm is based on a discretized system model, which is simple in form and has a light computational burden. Therefore, within a single switching cycle, not only can the calculation and generation of a control signal be completed, but also the training of the neural network can be completed. This high-frequency iterative cycle of control-training-adjustment ensures that the neural network maintains a rapid response and adaptive correction to the system state during operation, possessing rapid dynamic adjustment capabilities. Thus, it can maintain high robustness and stability when facing external disturbances, changes in system parameters, or grid harmonic interference. This invention adopts a "training as adjustment" strategy, discretizing the continuous-time system model and using the latest sampled data to efficiently update the gradient of the neural network weights at the end of each switching cycle (time k). This process is computationally lightweight, requires no trial and error, and can refresh all network parameters in real time, thereby precisely matching the adaptive adjustment rate of the controller with the high-frequency switching characteristics of the converter. While maintaining a dynamic response level comparable to traditional controllers, it achieves rapid tracking of real-time changes in the grid state. Through online learning of the switching cycle frequency, the neural network continuously accumulates experience and optimizes internal parameters during the control process. This not only gradually enhances its adaptability to complex operating conditions but also achieves a unity of dynamic adjustment and steady-state optimization. Therefore, this invention combines the fast dynamic response capability characteristic of traditional controllers with the unique global self-learning capability of neural networks. It fundamentally solves the dilemma of offline training optimizing only steady-state performance at the expense of dynamic quality, or online learning failing to achieve real-time control due to computational latency, thus achieving closed-loop output adjustment equivalent to model prediction results. Attached Figure Description

[0024] Figure 1 This is a circuit diagram for a single-phase grid-connected inverter. Figure 2 This is a diagram of the composite neural network controller structure. Figure 3 This is a block diagram of a neural network gradient propagation system. Figure 4a The waveforms of the grid-connected voltage and current under the control of this method are shown. Figure 4b The waveform diagram shows the dynamic adjustment process of grid-connected voltage and current under the control of this method. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] This invention provides a composite control method for an LCL-type single-phase grid-connected inverter based on a neural network, comprising: A mathematical model is established based on the circuit structure of a single-phase LCL grid-connected inverter. Set the given quantity of the control objective and substitute it into the mathematical model to obtain the control variables under steady state; A composite neural network controller is constructed, comprising an RBFNN and an FCNN. The input to the RBFNN is 2N components of the control variable under steady-state conditions, where N are in-phase components of each frequency and the other N are quadrature components of each frequency. The input to the FCNN is the output of the RBFNN, and the output is the control variable. Construct a basic controller, whose input is the error of the control target and whose output is the control variable. ; and Together they constitute the control variable M; the neural network parameters are calculated and updated under a preset learning rate with the objective function of minimizing the error; The LCL-type single-phase grid-connected inverter is controlled using the output control variable M.

[0027] Example 1 The following example uses a composite neural network LCL inverter, combined with... Figure 1 ~4 further elaborates on the implementation process of the present invention. The implementation process includes the following steps: (1) such as Figure 1 The grid-connected inverter shown can be analyzed to obtain its mathematical model, where... i 2 represents the inverter output current. L 1 、L 2 is the inverter inductor. C For inverter capacitors, V g This is the grid voltage. V dc This is the DC bus voltage. m This is an SPWM modulated wave. Referring to equation (1), in this example... i 1 、i 2 、v c Let X be the circuit state variable. mFor the control variable M, V g Let U be the input variable and Z be the control target, representing the output current. i 2: (15) (2) Grid-connected current i 2 and grid voltage V g When the inverter is used only as a generator and is in phase, the grid-connected current contains only the fundamental current, and the amplitude of the odd harmonic components is 0. The ideal modulation wave is derived based on equation (15). m ss The time-domain expression is: (16) in: (17) (3) Analysis of equation (16) shows that M contains a pair of orthogonal electrical signals. Referring to the forms of equations (2) to (4), the grid-connected voltage is selected. and its time derivative Design input quantity H, where k 1~ k 2 represents the coefficients for each term, which are designed according to the actual circuit parameters to normalize the input: (18) (4) Select RBFNN as the intelligent control function, according to Figure 2 The structure shown indicates that the neural network consists of three layers: an input layer, a hidden layer, and an output layer. The input layer contains H, which is determined in step (2). The hidden layer contains three neurons with activation functions as shown in equation (19). The output layer contains the input values ​​of the FCNN network. ,in c It is the center value of the radial basis functions, defined for The set of these adjustable neural network parameters: (19) (20) (twenty one) (5) Further design FCNN as a composite neural network, according to Figure 2 The FCNN has a structure with 5 nodes, taking the RBFNN output as input and the modulated wave compensation value as output. ,in It is the function gain, defined for The set of these adjustable neural network parameters: (twenty two) (twenty three) (twenty four) (6) Selecting a proportional controller as the basic stability control of the system, the expression for the control variable m is obtained. The proportional coefficient is the output of the proportional controller. , and neural network compensation value Together they form the modulated wave m, resulting in k Time-modulated wave m [ k ]; (25) (26) As derivation shows, under zero-error steady-state conditions, the control output of an LCL single-phase grid-connected inverter should consist of a pair of orthogonal sinusoidal signals, where the in-phase component is in phase with the grid voltage, and the quadrature component is orthogonal to the grid voltage. If the controller can explicitly generate this pair of orthogonal signals, the inverter can achieve accurate tracking of the ideal grid-connected current. Therefore, guided by the system model, this invention designs a proportional-neural network composite controller. The proportional controller provides basic control to maintain system stability, while the neural network controller generates the orthogonal signals required for steady-state operation. In the initial control phase, the proportional controller provides the main regulation, ensuring rapid current convergence. As the neural network learns and adjusts, the current error gradually decreases, and the neural network controller gradually dominates the output, approximating the ideal in-phase and quadrature component structure. At this point, the role of the proportional controller naturally weakens, only regaining its compensatory function when the system is disturbed or the neural network output deviates.

[0028] (7) Reference Figure 3 The system block diagram is presented, and the output of the composite neural network is derived and established. With control variables The gradient relationship.

[0029] First, we obtain the continuous-time system function: (27) Select sampling time T Discretize it to obtain the z-domain expression: (28) The parameters are as follows: (29) It is not difficult to find In response to itself, For the inverter input response, we can derive their respective time-domain difference equations by discussing them separately: (30) Part One Response Second part response Difference equation: (31) (32) The input sequence for: (33) Therefore, for : (34) hour: (35) As can be seen from equation (11), here we only care about and The partial derivative relationship between them is as follows: (36) Thus, the gradient relationship between the two is obtained: (37) Among the parameters K and d o It has been given in equation (29).

[0030] (8) According to the error e The principle of minimization is used to design the objective function of a composite neural network controller. J ; (38) (9) Referring to the structure of equation (13), and combining equation (20) in step (4), equation (23) in step (5), equation (26) in step (6), and equation (37) in step (7), the objective function is... J beg The partial derivatives of these adjustable neural network parameters are shown in equations (39) to (45). J right The partial derivative is J right The set of partial derivatives (taking Vg as an example): (39) (40) (41) (42) (43) (44) (45) (10) Update the parameter set of the neural network according to equations (39) to (45). ,by For example: (46) The control of the unidirectional grid-connected inverter at time k has now been completed.

[0031] Figure 4a The waveforms of the grid-connected voltage and current under the control of this method show that the current maintains good phase tracking with respect to the voltage and keeps the amplitude constant. Figure 4b In the first waveform segment (0-0.015s), the composite neural network controller is not connected, and the circuit relies solely on the basic controller to maintain basic stability. The output error is large and shows an increasing trend. In the second waveform segment (0.015-approximately 0.016s), the composite neural network controller is connected when the error is large, and network training begins. The controller parameters are updated and corrected in real time. After a very short learning period, the output error decreases significantly and remains close to zero, reflecting the controller's excellent dynamic response performance. In the third waveform segment (approximately 0.016-0.08s), the controller continues to control the output for stable tracking. As the error decreases, the basic controller naturally weakens, and the composite neural network controller becomes dominant. Through continuous self-learning, it is ready to cope with complex and sudden operating conditions, efficiently optimizing parameters and achieving a balance between dynamic response and steady-state performance, as well as a unity of dynamic adjustment and steady-state learning.

[0032] In summary, the control method of this invention can be used to control power electronic converters simply and reliably. The control method has a self-learning function, good control performance, strong adaptability, and is suitable for various operating conditions.

[0033] Example 2 The present invention also relates to an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0034] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor performs various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0035] Example 3 The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0036] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0037] Example 4 This invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method described in the above embodiments of this invention.

[0038] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" in this invention are intended to illustrate the invention and are not intended to limit the invention.

[0039] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A composite control method for an LCL-type single-phase grid-connected inverter based on neural networks, characterized in that, include: A mathematical model is established based on the circuit structure of a single-phase LCL grid-connected inverter. Set the given quantity of the control objective and substitute it into the mathematical model to obtain the control variables under steady state; A composite neural network controller is constructed, comprising a radial basis function neural network (RBFNN) and a fully connected neural network (FCNN). The input to the RBFNN is 2N components of the control variable in steady state, where N are in-phase components of each frequency and the other N are quadrature components of each frequency. The input to the FCNN is the output of the RBFNN, and the output is the control variable. ; Construct a basic controller, whose input is the error of the control target and whose output is the control variable. ; and Together they constitute the control variable M; the neural network parameters are calculated and updated under a preset learning rate with the objective function of minimizing the error; The LCL-type single-phase grid-connected inverter is controlled using the output control variable M.

2. The composite control method for an LCL-type single-phase grid-connected inverter as described in claim 1, characterized in that, The mathematical model is expressed as follows: Where X is the state variable in the circuit, M is the control variable, U is the grid voltage, Z is the control target, and C is the parameter matrix. This is the grid voltage. The magnitude of the grid voltage. This is the angular frequency of the grid voltage.

3. The composite control method for an LCL-type single-phase grid-connected inverter as described in claim 1, characterized in that, The objective function of the composite neural network controller is: in, e Z represents the error. * The given quantity for the control objective.

4. The composite control method for an LCL-type single-phase grid-connected inverter as described in claim 3, characterized in that, For the objective function J Find the partial derivatives of the neural network parameters: in, yes α is the set of trainable parameters in RBFNN. β This is the set of trainable parameters in FCNN; Discretizing the partial derivatives above allows for neural network parameter training within a single control cycle: in k Grid current at any time Z It is by k The output of the neural network at time -Q is determined; Calculate and update neural network parameters θ : in, η The learning rate is a constant used to adjust the training speed.

5. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 4.

7. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 4.

Images