A method for full-operating impedance identification of black-box grid-connected inverter

Abstract

The present invention discloses a method for full-operating-condition impedance identification of a black-box grid-connected inverter applicable to a variety of control structures and parameters. The method first derives the common features between the operating points and impedance characteristics of different inverters through theoretical formulas, which can form general prior knowledge applicable to any type of control structure or different parameters; then, a neural network with physical knowledge embedded is established based on the prior knowledge to optimize the model structure and reduce the demand for measurement data; at the same time, transfer learning theory is used to improve the flexibility of the neural network, allowing appropriate architecture adjustments to adapt to different inverters. The method of the present invention can reduce the demand for data sample size of the machine learning model, enhance the model interpretability, and improve the accuracy of model impedance identification. It is adapted to the actual situation of confidentiality of internal information of the inverter group at the engineering site, and reduces external interference to the actual system operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy power generation, and in particular relates to a full-operating impedance identification method for a black-box grid-connected inverter applicable to a variety of control structures and parameters. Background Art

[0002] Three-phase grid-connected inverters are increasingly used as interfaces to integrate renewable energy into the power grid, which introduces harmonic instability problems such as sub- / supersynchronous oscillations, posing new challenges to the safe and stable operation of power systems. Impedance-based stability analysis methods have been widely used to analyze the stability margin of interconnected systems between inverters and the grid, so full-operating-point impedance identification of three-phase grid-connected inverters is essential for analyzing the stability of inverter-grid interconnected systems under various operating conditions.

[0003] In recent years, data-driven machine learning models have the advantage of better fitting the nonlinear relationship between input and output variables and have been successfully applied to the field of full-operating impedance identification. M. Zhang analyzed the relationship between the impedance of the grid-connected inverter and its operating point in the literature [Artificial neural networkbased identification of multi-operating-point impedance model,”IEEE Trans. Power Electron., vol. 36, no. 2, pp. 1231–1235, Feb. 2021], and then used artificial neural networks to obtain multi-operating impedances. In addition, J. Lyu proposed a data-driven method for impedance identification and online stability assessment based on back-propagation neural networks in the literature [Data-driven impedance identification and stability online assessment of windfarm connected with MMC-HVDC, IEEE Trans. Ind. Appl., vol. 60, no. 2, pp. 2567-2576, March-April 2024]. The above neural network is trained from millions of measurement data to learn the potential characteristics of the black box impedance model and then predict the impedance characteristics under unknown operating conditions.

[0004] However, impedance measurement is very time-consuming in practice, and the measured data detected is limited. In addition, these trained neural networks can only be applied to fixed impedance models; this means that if the circuit parameters or control structure of the device to be identified changes dynamically due to operational requirements, the existing neural network will become inapplicable and need to be retrained with new re-measured data, which greatly increases the cost of impedance identification. Therefore, for practical applications, there are still several problems to be solved in this research: (1) Existing identification methods require a large amount of measurement data, resulting in a large measurement cost and training burden; (2) There is a lack of effective migration capabilities to other inverters with different parameters or control structures. Summary of the invention

[0005] In view of the above, the present invention provides a full-operating-condition impedance identification method for a black-box grid-connected inverter that is applicable to a variety of control structures and parameters, which can reduce the demand for data sample size of the machine learning model, enhance the model interpretability, and ensure the accuracy of the model output.

[0006] A full-operating impedance identification method for a black-box grid-connected inverter applicable to a variety of control structures and parameters comprises the following steps:

[0007] (1) The common characteristics of the synchronous control link and the auxiliary control link of the grid-connected inverter are derived;

[0008] (2) Based on the above common characteristics, a unified impedance expression for the grid-connected inverter is derived, revealing that the difference between different grid-connected inverters lies in the number of operating point variables and the corresponding polynomial coefficients;

[0009] (3) Based on the unified impedance expression of the grid-connected inverter, a neural network model with embedded physical knowledge is established, and the model structure is optimized using transfer learning theory;

[0010] (4) For different black-box grid-connected inverters, the above neural network model is trained based on offline pre-training and online training migration to obtain a full-operating condition impedance identification model, and the model is used to perform impedance identification on the black-box grid-connected inverter.

[0011] Furthermore, in the step (1), there are multiple synchronous control links of the grid-connected inverter (including phase-locked loop control, virtual synchronous control and droop control, etc.), and the common characteristic expressions of these synchronous control links are as follows:

[0012]

[0013] Where: Δθ is the small disturbance signal of coordinate change, f din 、f qin 、f did 、f qid、f dvn 、f dvd 、f qvn 、f qvd are all linear functions about the working point, and are the voltage signals of the d-axis and q-axis respectively, and They are the small current signals of the d-axis and q-axis respectively.

[0014] Furthermore, in the step (1), there are multiple auxiliary control links of the grid-connected inverter (including current control loop, AC / DC voltage control loop, power control loop and main circuit, etc.), and the common characteristic expressions of these auxiliary control links are as follows:

[0015]

[0016] Where: Δθ is the small disturbance signal of coordinate change, f i1 、f i2 、f i3 、f i4 、f u1 、f u2 、f u3 、f u4 is a linear function about the working point, and are the voltage signals of the d-axis and q-axis respectively, and They are the small current signals of the d-axis and q-axis respectively.

[0017] Furthermore, the unified impedance expression of different grid-connected inverters in step (2) is as follows:

[0018]

[0019] Where: Y dd , Y dq , Y qd , Y qq is the impedance element in the dq coordinate system, x represents the polynomial composed of the operating point variables, A k and B k represents the vector consisting of polynomial coefficients, a k1 ~a k4 and b k1 ~b k4 are all polynomial coefficients, k = 1, 2, 3, 4, T represents transpose, U d and U q are the steady-state voltage values ​​of the d-axis and q-axis, I d and I q are the steady-state current values ​​of the d-axis and q-axis respectively.

[0020] Furthermore, the neural network model in step (3) is composed of an input layer, a hidden layer, and an output layer connected in sequence, wherein the input layer is the operating point variable and frequency, and the output layer is the amplitude and phase of the impedance. The model has 2 branches and 5 hidden layers H1 to H5, specifically:

[0021] The first branch uses hidden layer H1 to simulate the polynomial x composed of working point variables, and the second branch is implemented by hidden layers H2 and H3 to simulate the vector composed of polynomial coefficients;

[0022] The input of hidden layer H1 is the operating point variable, the output is the polynomial x, and the number of neurons is set to 20;

[0023] The input of hidden layer H2 is the operating point frequency, which is used to simulate the output polynomial coefficients, and the number of neurons is set to 20;

[0024] The input of hidden layer H3 is the output of H2, which is used to simulate the output vector A k and B k , the number of neurons is set to 8;

[0025] The input of hidden layer H4 is the output of H1 and H3, which is used for impedance calculation to obtain the impedance element Y dd , Y dq , Y qd , Y qq , the number of neurons is set to 4;

[0026] The input of hidden layer H5 is the output of H4, which is used to calculate the amplitude and phase of each impedance element. The output is The number of neurons is set to 8, where and Y dd The magnitude and phase of and Y dq The magnitude and phase of and Y qd The magnitude and phase of and Y qq amplitude and phase.

[0027] Furthermore, in step (3), for different grid-connected inverters, transfer learning is guided according to the unified impedance expression of the grid-connected inverter to further compress the degree of freedom of the neural network model to reduce the amount of training data.

[0028] Furthermore, the specific implementation of step (4) is as follows:

[0029] 4.1 For a white-box grid-connected inverter with known control structure and parameters, a simulation model is built in Matlab / Simulink and impedance data at different operating points are obtained by frequency sweeping;

[0030] 4.2 Send the impedance data to the neural network model for training, so that the model can achieve the expected recognition performance and obtain an offline pre-training model;

[0031] 4.3 For black-box grid-connected inverters with unknown control structures and parameters in practice, frequency sweep measurement is performed on them through the disturbance injection method to obtain online measurement impedance data at different working points;

[0032] 4.4 The online measured impedance data is sent to the offline pre-training model for further training, and the model structure is optimized through transfer learning theory so that the model can achieve the expected identification performance and obtain the full-operating impedance identification model of the black-box grid-connected inverter.

[0033] Furthermore, the white box grid-connected inverter in step 4.1 has a phase-locked loop synchronization link, an L-type filter and a current controller; during the frequency sweep process, the frequency f range is set to 1 to 200 Hz, the interval is 2 Hz, and the operating point variable I is set d The value range is 0.1~0.9pu and the interval is 0.2pu, I q The value range is 0~0.9pu with an interval of 0.3pu, U d The value range of is 0.85~1.15pu with an interval of 0.1pu, and there are 80 working point variables with corresponding impedance data.

[0034] Furthermore, in the process of optimizing the model structure using transfer learning theory in step 4.4, the hidden layers H1 and H2 of the offline pre-trained model are copied and fine-tuned, the hidden layer H5 is copied and frozen, and then the full-operating impedance identification model of the black-box grid-connected inverter is obtained through data training.

[0035] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to execute the computer program to implement the above-mentioned black box grid-connected inverter full-operating condition impedance identification method applicable to a variety of control structures and parameters.

[0036] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned full-operating-condition impedance identification method for a black-box grid-connected inverter applicable to a variety of control structures and parameters.

[0037] Based on the above technical solution, the present invention has the following beneficial technical effects:

[0038] 1. The present invention improves the physical significance and generalization ability of the model by deriving the common characteristics of different control structures and parameters and using them to guide the design of the neural network model.

[0039] 2. The present invention establishes a full-operating-condition impedance identification model through data drive and improves the accuracy of the model through data training. It can reduce the amount of model training data while improving the model's migration and identification capabilities for different control structures and parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The present invention is a flow chart of the derivation of common characteristics of different grid-connected inverters and the establishment of an impedance identification model for all operating conditions.

[0041] Figure 2 Schematic diagram of the neural network model structure based on embedded physical knowledge.

[0042] Figure 3 Schematic diagram of the optimized structure of the neural network model based on transfer learning theory.

[0043] Figure 4 Schematic diagram of the model training process based on offline pre-training and online training migration. DETAILED DESCRIPTION

[0044] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0045] The present invention is applicable to the full-operating impedance identification method of black-box grid-connected inverters with various control structures and parameters. First, the common characteristics of different grid-connected inverters are derived, and a full-operating impedance identification model of the grid-connected inverter is established. The specific process is as follows: Figure 1 As shown:

[0046] (1) The common characteristics of the synchronous control link and the auxiliary control link of the grid-connected inverter are derived.

[0047] The control structure of the grid-connected inverter consists of two parts: the synchronous control link and the auxiliary control link. The inverter circuit model is usually converted to the dq coordinate system to achieve decoupled control of active and reactive currents; the synchronous control link is used to generate a reference angle θ for coordinate transformation from the three-phase coordinate system to the system dq coordinate system. Typical control schemes include phase-locked loop (PLL), droop control, and virtual synchronous generator (VSG) control. Due to the dynamic performance of the synchronous link, harmonics will cause corresponding phase disturbances, resulting in a phase difference Δθ between the controller dq coordinate system and the system dq coordinate system.

[0048] For phase-locked loop synchronous control:

[0049]

[0050] Where: k p and k i is the phase-locked loop control parameter; the above formula can be rewritten as:

[0051]

[0052] in: and and They are the voltage and current small signals of the dq axis respectively.

[0053] Similarly, in droop control, Δθ can be expressed as:

[0054]

[0055] Where: D p is the droop coefficient, U d , U q and I d ,I q It is the steady-state value of voltage and current of dq axis, that is, the operating point.

[0056] In the virtual synchronous control link, Δθ can be expressed as:

[0057]

[0058] Where: J is the integration constant.

[0059] Therefore, for all these typical synchronization links, Δθ can always be expressed in a unified form:

[0060]

[0061] Where: f din 、f qin 、f did 、f qid 、f dvn 、f dvd 、f qvn 、f qvd is a linear function of the operating point.

[0062] The auxiliary control link is used to ensure the effective and stable operation of the entire system, which is mainly composed of the main circuit, current control, AC / DC voltage control and PQ control. In the main circuit, the flow direction of small signals in the system dq coordinate system can be derived:

[0063]

[0064] Where: Y out is the transfer function of the filter, which consists of the system parameter matrix.

[0065] Taking into account the small disturbances in the system coordinate system transformation process, as follows:

[0066]

[0067] in: and It is the small signal of the duty cycle in the system dq coordinate system and the controller dq coordinate system.

[0068] In other control loops, the small signal flow in the controller dq coordinate system can be derived:

[0069]

[0070] Where: H 1 is from arrive The transfer function, H 2 is from arrive The transfer function of .

[0071] Due to the introduction of the control loop, additional signal flows will be added, which can be expressed as:

[0072]

[0073] Where: H 3 is from arrive The transfer function of .

[0074] After linearizing the controller, H 1 ~H 3 All of them are composed of linear functions with working points. Combined with the above derivation, the common characteristics of the auxiliary control link can be expressed as:

[0075]

[0076] Where: f i1 、f i2 、f i3 、f i4 、f u1 、f u2 、f u3 、f u4 is a linear function of the operating point.

[0077] (2) Combined with the impedance definition expression, the unified impedance expression of different inverters is derived, where the definition of dq admittance is as follows:

[0078]

[0079] Combined with the common features derived in step (1), the above formula can be rewritten as:

[0080]

[0081] where g i1 ~g i4 and g u1 ~g u4 The eight variables are all polynomial functions of the operating point and can be expanded as:

[0082]

[0083] Where: x represents the polynomial variable of the working point, A k and B k are the corresponding coefficients of the polynomial variables consisting of the system parameters, and k is the number of polynomial variables.

[0084] Combining the above formula, the dq impedance of different inverters can be uniformly expressed as:

[0085]

[0086] It can be seen that x is a matrix related only to the working point, A k and B k is a matrix related only to system parameters. For different inverters, x, A k and B k The elements in are different; for a specific inverter, the full-operating impedance is obtained under the assumption that the system parameters remain unchanged. When the system parameters change due to different control strategies, the full-operating impedance model should be rebuilt.

[0087] Admittance-based stability analysis is usually implemented in Bode plots, where the admittance is displayed on a logarithmic scale; therefore, the admittance model in the above equation is presented in the form of logarithmic amplitude and phase, that is:

[0088] Y pn-mag =Mag[Y pn (U d ,U q ,I d ,I q ,f p )]

[0089] Y pn-pha =Ang[Y pn (U d ,U q ,I d ,I q ,f p )]

[0090] (3) Based on the inverter unified impedance expression derived in step (2), a neural network model based on physical knowledge embedding is established. The input of the neural network model is the operating point variable and frequency, and the output is the amplitude and phase of the impedance. The model consists of two branches and five hidden layers, such as Figure 2 As shown, specifically:

[0091] Two branches: In order to achieve impedance characteristic decoupling between the operating point and the system parameters, the input should be connected to two branches respectively. Branch 1 is used to simulate the polynomial variables of the operating point in x; Branch 2 has two layers, which is used to simulate the corresponding coefficients of the polynomial variables composed of the system parameters.

[0092] First layer: The input of the first layer is the operating point, which is used to simulate x, the polynomial variable of the inverter operating point. Therefore, the number of neurons in the first layer is consistent with the polynomial variable of the operating point. However, due to the black box characteristics, the specific number of polynomial variables of the inverter is unknown. Since the general inverter order does not exceed 20, the initial number of neurons in the first layer is set to 20.

[0093] Second layer: The second layer is the first layer of branch 2, and the input is frequency. The second layer simulates a k and b k , which are the corresponding coefficients of the polynomial variables. The number of coefficients should be consistent with the polynomial variables. Therefore, the number of neurons in the second layer should be the same as in the first layer, also set to 20.

[0094] The third layer: The third layer is the second layer of branch 2, and the output of the second layer is the input of the third layer. The third layer can simulate A k and B k , there are 8 neurons.

[0095] The fourth layer: The fourth layer represents the dq impedance calculation. k and B k And the working point vector x in the first layer is the input of the fourth layer. Therefore, the fourth layer has 4 neurons, corresponding to Y dd , Y dq , Y qd , Y qq .

[0096] Layer 5: The fifth layer corresponds to the impedance magnitude and phase calculation. The input is Y from the fourth layer dd , Y dq , Y qd , Y qq The fifth layer has 8 neurons, corresponding to

[0097] (4) The unified expression obtained in step (2) is used to further optimize the neural network model structure using transfer learning theory. Transfer learning is a machine learning method that uses the knowledge learned from a basic model (source domain) to improve the learning efficiency and prediction performance of another related model (target domain). The premise of this method is that there should be shared underlying features or knowledge between the basic model and the transfer model.

[0098] From the above derivation, it can be seen that although inverters may have differences in system / control parameters and control structures, they have common characteristics and can be expressed in a unified form. Therefore, the neural network model established for the inverter can also have a similar framework, and transfer learning theory can be used to further optimize the model structure and reduce the amount of data required.

[0099] Transfer learning frameworks such as Figure 3 As shown in the figure, the basic model is established based on the training dataset A, which contains a large amount of impedance data about known inverters. For other unknown inverters, the basic model can be adjusted to obtain a migration model for black-box impedance identification; for each inverter, the operating point and system parameters have impedance characteristics decoupled, and the initial number of neurons is set to 20, so the weights and biases of the first and second layers in the basic model can be directly copied to the migration model. Next, since the amplitude and phase calculations are fixed for each inverter, the fifth layer of the basic model can be copied and then frozen in the migration model. Finally, the migration model can be trained based on dataset B, which consists of a relatively small amount of impedance data from unknown inverters; through continuous training, the migration model can learn the potential characteristics of unknown inverters and achieve excellent prediction performance.

[0100] For different black-box grid-connected inverters, this implementation method constructs a full-operating impedance identification model through a training method based on offline pre-training and online training migration to reduce the amount of model training data. The specific process is as follows: Figure 4 As shown:

[0101] (1) In order to realize impedance identification of other black-box inverters, a basic neural network identification model should be first established based on the white-box inverter, which has a phase-locked loop synchronization link, an L-type filter and a current controller; then the inverter is built in MATLAB / Simulink to generate simulation data for training the basic model.

[0102] (2) For the simulation model in step (1), the operating conditions are changed, and the impedance data of the response is first obtained through the frequency sweep module. Specifically, in the impedance data set, the frequency f ranges from 1 Hz to 200 Hz, with an interval of 2 Hz, and I d The range is 0.1~0.9pu, the interval is 0.2pu, I qThe range is 0~0.9pu, the interval is 0.3pu, U d The range is 0.85~1.15pu, with an interval of 0.1pu, and there are 80 working points with corresponding impedance data. The data is then sent to the neural network model for training. Due to the limited amount of data, the data set is randomly divided into a training set and a validation set; the training set is used to train the neural network model to learn the potential characteristics of impedance; the validation set is used to evaluate whether the training process is successful and to improve the prediction performance by optimizing the model parameters. This implementation uses the mean square error as the loss function to train the neural network. The training algorithm is adaptive moment estimation, and the validation algorithm for optimizing the model parameters is the Bayesian optimization algorithm. After the training is completed, an offline impedance identification model is obtained.

[0103] (3) For a practical black box inverter, the acquisition of impedance data requires injecting frequency disturbance signals into the inverter multiple times under stable operating conditions, then sampling the signal from the common coupling point, and extracting the voltage and current signals at the disturbance frequency through fast Fourier transform. At this time, the impedance can be expressed as follows:

[0104]

[0105] Where: V sp 、V sp2 Respectively represent the frequency f p and f p -2f 1 The voltage disturbance at I sp ,I sp2 Indicates that at frequency f p and f p -2f 1 The resulting current response.

[0106] Repeat the above steps until all the specified frequency points are measured. This is the frequency scanning process. After this process, the impedance data of the inverter at a working point will be obtained. By changing the working point of the inverter and continuing the frequency scanning process, the impedance data set required for measurement can be obtained. Specifically: In the impedance measurement data set, the frequency f ranges from 1Hz to 200Hz, the interval is 5Hz, and I d The range is 0.1~0.9pu, the interval is 0.4pu, I q The range is 0~0.9pu, the interval is 0.4pu, U d The range is from 0.9 to 1.1 pu with an interval of 0.1 pu. There are 27 operating point variables with corresponding impedance data.

[0107] (4) The impedance identification model of the black box inverter is established by using the transfer learning theory through the offline impedance identification model of step (2), and then the model is trained by the limited measurement data in step (3). Similarly, the data set is randomly divided into a training set and a validation set, and the mean square error is used as the loss function to train the neural network. The training algorithm is adaptive moment estimation, and the validation algorithm for optimizing the model parameters is the Bayesian optimization algorithm. After the training is completed, the online impedance identification migration model is obtained.

[0108] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art to the present invention based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A method for identifying impedance of a black box grid-connected inverter under all working conditions applicable to a variety of control structures and parameters comprises the following steps: (1) The common characteristics of the synchronous control link and the auxiliary control link of the grid-connected inverter are derived; There are many types of synchronous control links for grid-connected inverters. The common characteristic expressions of these synchronous control links are as follows: in: Δθ is the small disturbance signal of coordinate change, f din 、f qin 、f did 、f qid 、f dvn 、f dvd 、f qvn 、f qvd are all linear functions about the working point, and are the voltage signals of the d-axis and q-axis respectively, and They are the small current signals of d-axis and q-axis respectively; There are many auxiliary control links for grid-connected inverters. The common characteristic expressions of these auxiliary control links are as follows: Where: f i1 、f i2 、f i3 、f i4 、f u1 、f u2 、f u3 、f u4 is a linear function about the working point; (2) Based on the above common characteristics, the unified impedance expression of the grid-connected inverter is derived as follows, revealing that the difference between different grid-connected inverters lies in the number of operating point variables and the corresponding polynomial coefficients; Where: Y dd , Y dq , Y qd , Y qq is the impedance element in the dq coordinate system, x represents the polynomial composed of the operating point variables, A k and B k represents the vector consisting of polynomial coefficients, a k1 ~a k4 and b k1 ~b k4 are all polynomial coefficients, k = 1, 2, 3, 4, T represents transpose, U d and U q are the steady-state voltage values ​​of the d-axis and q-axis, I d and I q are the steady-state values ​​of the current on the d-axis and q-axis respectively; (3) Based on the unified impedance expression of the grid-connected inverter, a neural network model with embedded physical knowledge is established, and the model structure is optimized using transfer learning theory; (4) For different black-box grid-connected inverters, the above neural network model is trained based on offline pre-training and online training migration to obtain a full-operating condition impedance identification model, and the model is used to perform impedance identification on the black-box grid-connected inverter.

2. The method for identifying impedance of a black box grid-connected inverter under all operating conditions according to claim 1, characterized in that: The neural network model in step (3) is composed of an input layer, a hidden layer, and an output layer connected in sequence, wherein the input layer is the operating point variable and frequency, and the output layer is the amplitude and phase of the impedance. The model has 2 branches and 5 hidden layers H1 to H5. Specifically: The first branch uses hidden layer H1 to simulate the polynomial x composed of working point variables, and the second branch is implemented by hidden layers H2 and H3 to simulate the vector composed of polynomial coefficients; The input of hidden layer H1 is the operating point variable, the output is the polynomial x, and the number of neurons is set to 20; The input of hidden layer H2 is the operating point frequency, which is used to simulate the output polynomial coefficients, and the number of neurons is set to 20; The input of hidden layer H3 is the output of H2, which is used to simulate the output vector A k and B k , the number of neurons is set to 8; The input of hidden layer H4 is the output of H1 and H3, which is used for impedance calculation to obtain the impedance element Y dd , Y dq , Y qd , Y qq , the number of neurons is set to 4; The input of hidden layer H5 is the output of H4, which is used to calculate the amplitude and phase of each impedance element. The output is The number of neurons is set to 8, where and Y dd The magnitude and phase of and Y dq The magnitude and phase of and Y qd The magnitude and phase of and Y qq amplitude and phase.

3. The method for identifying impedance of a black box grid-connected inverter under all operating conditions according to claim 1, characterized in that: In the step (3), for different grid-connected inverters, transfer learning is guided according to the unified impedance expression of the grid-connected inverter to further compress the degree of freedom of the neural network model to reduce the amount of training data.

4. The method for identifying impedance of a black box grid-connected inverter under all operating conditions according to claim 2, characterized in that: The specific implementation of step (4) is as follows: 4.1 For a white-box grid-connected inverter with known control structure and parameters, a simulation model is built in Matlab / Simulink and impedance data at different operating points are obtained by frequency sweeping; 4.2 Send the impedance data to the neural network model for training, so that the model can achieve the expected recognition performance and obtain an offline pre-training model; 4.3 For black-box grid-connected inverters with unknown control structures and parameters in practice, frequency sweep measurement is performed on them through the disturbance injection method to obtain online measurement impedance data at different working points; 4.4 The online measured impedance data is sent to the offline pre-training model for further training, and the model structure is optimized through transfer learning theory so that the model can achieve the expected identification performance and obtain the full-operating impedance identification model of the black-box grid-connected inverter.

5. The method for identifying impedance of a black box grid-connected inverter under all operating conditions according to claim 4 is characterized in that: In the process of optimizing the model structure by using transfer learning theory in step 4.4, the hidden layers H1 and H2 of the offline pre-trained model are copied and fine-tuned, the hidden layer H5 is copied and frozen, and then the full-operating impedance identification model of the black-box grid-connected inverter is obtained through data training.

6. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is used to execute the computer program to implement the full-operating-condition impedance identification method of a black-box grid-connected inverter as claimed in any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for identifying impedance of a black-box grid-connected inverter under all operating conditions as claimed in any one of claims 1 to 5 is implemented.

Images