A three-phase grid-connected inverter admittance identification method based on complex-valued neural network

Abstract

The application discloses a kind of based on complex-valued neural network three-phase grid-connected inverter admittance identification method, it is related to power electronics field.The method includes: after constructing the small signal flow graph of three-phase grid-connected inverter in synchronous coordinate system, the port admittance matrix under steady-state operating point is constructed, the corresponding characteristic polynomial is determined and factorized, and the unified pole set is obtained;Based on the pole-residual expansion of the port admittance matrix of unified pole set, the transfer function of the elements in the port admittance matrix is determined;Complex-valued neural network and its loss function are constructed;Complex-valued neural network is trained by loss function;The operating condition information and frequency point of grid-connected point are input into complex-valued neural network, and the port admittance under current operating condition is obtained.The method has the technical advantages of low calculation amount, high identification efficiency, supporting multi-condition fast identification and high physical interpretability of fitting result.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and in particular to a method for identifying the admittance of a three-phase grid-connected inverter based on a complex-valued neural network. Background Technology

[0002] With the integration of numerous power electronic devices into power distribution networks, three-phase grid-connected inverters have gradually become the main interface equipment in power systems. The interaction between their small-signal port admittance characteristics and grid impedance directly affects the resonance risk and stability margin of the parallel system. Given the constantly changing operating conditions, accurately obtaining the frequency domain admittance characteristics of three-phase grid-connected inverters under different conditions for resonance analysis and parallel system operation risk assessment has become a key issue in power electronic grid-connected scenarios.

[0003] Traditional methods often rely on field frequency sweep tests, but frequency sweeping under each operating condition is labor-intensive and inefficient, making it difficult to meet the needs of rapid analysis under multiple operating conditions. In recent years, data-driven port admittance modeling methods have been proposed, which directly fit the relationship between port admittance and frequency and operating conditions through models such as neural networks. However, such methods usually treat admittance as a black-box frequency domain curve for fitting, requiring large and complex network structures and a large number of training samples. The model has numerous parameters, high computational cost, and is difficult to reflect the characteristics of the three-phase grid-connected inverter itself, resulting in poor physical interpretability of the fitting results. Summary of the Invention

[0004] To address the aforementioned technical problems in existing technologies, namely the drawbacks of traditional methods for obtaining the frequency domain admittance characteristics of three-phase grid-connected inverters, which involve large workloads, low efficiency, and inability to meet the requirements for rapid identification under multiple operating conditions, and the drawbacks of introducing neural networks to fit the relationship between port admittance and frequency and operating conditions, which involve complex network structures, large computational loads, and poor physical interpretability of the fitting results, this invention provides a method for identifying the admittance of three-phase grid-connected inverters based on complex-valued neural networks.

[0005] Specifically, the technical solution provided by this invention includes the following steps:

[0006] Step S1: Construct the small-signal flow graph of the three-phase grid-connected inverter in the synchronous coordinate system; construct the port admittance matrix of the three-phase grid-connected inverter at the steady-state operating point based on the small-signal flow graph;

[0007] Step S2: Determine the characteristic polynomial corresponding to the port admittance matrix and factor it to obtain the unified pole set of the three-phase grid-connected inverter; construct the pole-residual expansion of the port admittance matrix based on the unified pole set, and determine the transfer function of each element in the port admittance matrix.

[0008] Step S3: After applying the Swish activation function to the real and imaginary channels of the complex-valued neural network, construct the loss function of the complex-valued neural network by combining the transfer function of each element in the port admittance matrix;

[0009] Step S4: Train the complex-valued neural network using the loss function;

[0010] Step S5: Input the operating condition information and frequency point of the grid connection point into the trained complex value neural network to obtain the port admittance under the current operating condition.

[0011] Compared to existing technologies, the technical solution provided by this invention employs a complex-valued neural network as the modeling framework for port admittance. This allows the complex-domain characteristics of the port admittance model and its frequency-related phase change patterns to be preserved within the complex-valued neural network, enabling the complex-valued neural network to have a more direct ability to express amplitude and phase characteristics. Furthermore, by using the operating condition information of the grid connection point under stable operating conditions as input and the pole-residue as the output of the complex-valued neural network, the network can directly learn the patterns of pole and residual changes with operating conditions under different operating conditions. By constructing a physically constrained complex-valued neural network through a loss function and training it using port admittance data, the complex-valued neural network can quickly identify port admittance based solely on the operating condition information of the current operating condition, effectively compressing the complexity of the neural network and improving physical interpretability. In short, the technical solution provided by this invention has the advantages of low computational cost, high identification efficiency, support for rapid identification under multiple operating conditions, and high physical interpretability of the fitting results. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the system structure of the three-phase grid-connected inverter in this invention.

[0013] Figure 2 This is a schematic diagram of the phase angle disturbance of the synchronous coordinate system of the three-phase grid-connected inverter and the synchronous coordinate system of the controller in this invention.

[0014] Figure 3 This is a schematic diagram of the small-signal flow graph of the three-phase grid-connected inverter in this invention.

[0015] Figure 4 This is a schematic diagram of the overall structure of the complex-valued neural network in this invention. Detailed Implementation

[0016] The technical solution provided by the present invention will be further described in detail below with reference to the accompanying drawings.

[0017] In existing technologies, how to effectively utilize existing small-signal physics knowledge to constrain and compress data-driven models, reduce network complexity, and improve the physical interpretation of results is a problem that remains to be solved.

[0018] The method for identifying the admittance of a three-phase grid-connected inverter based on a complex-valued neural network according to the present invention includes the following steps:

[0019] Step S1: Establish the small-signal admittance model of the three-phase grid-connected inverter in the synchronous coordinate system, and derive the port admittance matrix at the steady-state operating point.

[0020] Step S2: Based on the characteristics of the port admittance matrix, determine the applicable pole / residual expansion form and its order. Specifically, determine the characteristic polynomial corresponding to the port admittance matrix and factorize it to obtain the unified pole set of the three-phase grid-connected inverter; construct the pole-residual expansion of the port admittance matrix based on the unified pole set, and determine the transfer function of each element in the port admittance matrix.

[0021] Step S3: Construct a complex-valued physical constraint neural network based on the pole / residual characteristics at the stable operating point, integrating complex numbers and their algebraic properties into the neural network to map the complex number operation process. Specifically, construct a complex-valued neural network and build the loss function of the complex-valued neural network based on the transfer function of the elements in the port admittance matrix.

[0022] Step S4: Train and validate the complex-valued neural network using electromagnetic simulation data under different operating conditions to learn the mapping relationship between operating conditions and pole-residue, and obtain the small-signal admittance model parameters of the three-phase grid-connected inverter. Specifically, collect electromagnetic data of the three-phase grid-connected inverter under different operating conditions to construct a sample set; based on the sample set, train the complex-valued neural network through a loss function, and adjust the network parameters and hyperparameters; input the electromagnetic data into the complex-valued neural network to identify the reconstructed admittance under the corresponding operating conditions.

[0023] Step S1 includes the following sub-steps:

[0024] Step S11: Construct a small-signal flow graph based on the grid-connected inverter structure and small-signal control mechanism.

[0025] like Figure 1 As shown, the system structure of the three-phase grid-connected inverter is first determined. The system structure of the three-phase grid-connected inverter consists of a DC-side voltage source. Three-phase bridge inverter, filters (including inductors) and resistance The AC power supply forms the main power path. The control loop includes a sampling unit for the grid-connected signal, a coordinate transformation unit in a synchronous rotating coordinate system (dq coordinate system), a phase-locked loop (PLL), a current inner-loop PI controller, and a PWM (Pulse-Width Modulation) modulator control element. The control loop samples the current signal from the PCC (Point of Common Coupling). and voltage signal Introduce a q-axis reference current. Subtract the current signal after transformation to the dq coordinate system Then, PI control is performed. Voltage signal The phase is obtained after inputting into the PLL. Then transform to the dq coordinate system and introduce the d-axis reference current. Subtract the signal output from the dq coordinate system before performing PI control.

[0026] The aforementioned electrical components and control modules together constitute the signal and energy transmission link between the inverter and the power grid.

[0027] After clarifying the system architecture of the three-phase grid-connected inverter, a small-signal flow graph of the inverter is constructed. Small-signal modeling essentially obtains the linearized characteristics of the three-phase grid-connected inverter at its stable operating point. Small-signal analysis uses small disturbances in the grid-side voltage as the excitation source, which act on both the power stage and the control stage. In the power stage, voltage disturbances directly affect the filter network and thus the grid-connected current. In the control stage, considering the dynamic characteristics of the PLL, the control system simultaneously exists in two rotating coordinate systems: the system synchronization coordinate system of the three-phase grid-connected inverter and the controller synchronization coordinate system. For example... Figure 2 As shown, voltage disturbances cause phase angle disturbances in both coordinate systems, which are further coupled to the controller reference frame, forming an indirect effect of voltage disturbances on current control and voltage control. Figure 2 middle, The q-axis of the system's synchronization coordinate system. The d-axis of the system's synchronous coordinate system; The q-axis of the controller's synchronization coordinate system. The d-axis of the controller's synchronization coordinate system; The phase angle difference caused by the disturbance.

[0028] The small signal flow graph formed based on the above two propagation paths is as follows: Figure 3 As shown. Among them, The disturbance voltage signal injected into the grid connection point, This is the current response signal at the grid connection point. This is the feedback current for the control loop after the disturbance. To control the introduced reference current.

[0029] Step S12: Derive the port admittance matrix of the three-phase grid-connected inverter from the small-signal flow graph. First, the key modules in the small-signal flow graph are mathematically described and converted into a unified frequency domain matrix form. Then, the modules are combined and calculated according to the circuit topology and the connection order of the control loops to obtain the overall transmission relationship between port voltage disturbances and output current disturbances.

[0030] The filter admittance matrix is ​​determined based on the small-signal flow graph, using the following formula:

[0031] ;

[0032] ;

[0033] In the formula, For the Laplace operator, Here is the filter admittance matrix. For the filter circuit inductor, For the filter circuit resistor, This refers to the fundamental angular frequency of a three-phase grid-connected inverter. The filter admittance matrix The product of the identity negative matrix.

[0034] Based on the small-signal flow graph, the transfer functions are determined as follows:

[0035] ;

[0036] ;

[0037] ;

[0038] In the formula, This represents the time delay transfer function matrix caused by digital control and PWM. The equivalent time delay of a three-phase grid-connected inverter. The transfer function matrix for decoupling the current loop controller. The proportional gain of the current loop. This is the integral gain of the current loop. The transfer function for the output angle and q-axis voltage of the phase-locked loop. The proportional gain of the current loop. This is the integral gain of the current loop. This is the fundamental voltage at the steady-state operating point.

[0039] Considering the dynamic characteristics of PLLs, small signals from the mains voltage can cause disturbances in the PLL output phase angle, which in turn affects the control circuit. The formula is as follows:

[0040] ;

[0041] ;

[0042] In the formula, Let be the transfer function of the small-signal disturbance path from the grid voltage to the control current in the synchronous coordinate system. The q-axis component of the grid current at the steady-state operating point. The d-axis component of the grid current at the steady-state operating point. Let be the transfer function of the small-signal disturbance path from the grid voltage to the control voltage in the synchronous coordinate system. This represents the q-axis component of the inverter modulation coefficient at the steady-state operating point. This represents the d-axis component of the inverter modulation coefficient at the steady-state operating point.

[0043] Finally, the port admittance matrix of the three-phase grid-connected inverter is derived based on the small-signal flow graph. The formula is as follows:

[0044] ;

[0045] ;

[0046] In the formula, It is the identity matrix. , , and These are the four elements of the port admittance matrix. This represents the gain of the PWM modulator control circuit.

[0047] Step S2 includes the following sub-steps:

[0048] Step S21: Determine the characteristic polynomial of the admittance matrix After obtaining the small-signal port admittance matrix of the three-phase grid-connected inverter in the synchronous coordinate system, the transfer function structure of each element in the matrix is ​​first analyzed. Sub-modules such as the filter circuit, current PI controller, time delay element, and PLL dynamic characteristics are written as frequency domain transfer functions, and cascaded and superimposed according to the connection relationship of the small-signal flow graph. It can be seen that the closed-loop system corresponding to the port admittance matrix consists of only a finite number of the aforementioned dynamic elements, and its poles are determined by a unified characteristic polynomial. Considering the second-order electromagnetic dynamics introduced by the filter circuit, the second-order phase angle dynamics of the PLL, the integral element of the PI controller, and the first-order approximation of the time delay, after merging and simplifying each element, it can be seen that the highest order of the denominator of the port admittance does not exceed the eighth order.

[0049] Determine the characteristic polynomial The formula is as follows:

[0050] ;

[0051] In the formula, to All of these are characteristic polynomial coefficients of a three-phase grid-connected inverter.

[0052] During the merging and simplification process, due to the unknown filtering and control parameters, zero-pole cancellation may occur in practice, resulting in the effective denominator order of some channels being lower than the eighth order. However, this phenomenon only stems from the propagation path characteristics of specific channels and does not affect the intrinsic dynamics of the entire closed-loop system. All poles of the three-phase grid-connected inverter are determined by a unified closed-loop state matrix, independent of the propagation path of specific channels. Therefore, each element of the port admittance matrix can be described within the framework of this unified eighth-order characteristic polynomial. Factoring the characteristic polynomial yields the unified pole set of the three-phase grid-connected inverter, as shown in the following formula:

[0053] ;

[0054] In the formula, For the first A unified pole.

[0055] Step S22: Construct the pole-residual expansion of the port admittance matrix based on the unified pole set. After determining the common characteristic polynomial and its pole set, the four elements of the port admittance matrix are uniformly expressed as the ratio of the numerator polynomial to the denominator polynomial. The transfer function of each element can then be constructed, as shown in the following formula:

[0056] ;

[0057] ; ;

[0058] Theoretical modeling shows that, For a numerator polynomial of order no higher than seven, expand the above fractional part as follows:

[0059] ;

[0060] In the formula, For elements Regarding the first A unified pole The corresponding residual coefficients.

[0061] Through the aforementioned pole-residual expansion, all frequency domain characteristics of the port admittance matrix can be uniformly expressed as a combination of a set of fixed poles and corresponding residuals. This combination retains the electromagnetic dynamics, control dynamics, and PLL dynamics information in the small-signal model of a three-phase grid-connected inverter, giving the port admittance a clear structural characteristic in its mathematical description.

[0062] Step S3 includes the following sub-steps:

[0063] Step S31: Construct the network structure of the complex-valued neural network. Unlike traditional real-valued neural networks, the neurons, connection weights, and biases in the complex-valued neural network are all represented in complex form. This allows for direct processing of amplitude and phase information within a unified computational framework, providing a more natural expressive ability for dynamic systems in the complex domain.

[0064] The small-signal model of the port admittance of a three-phase grid-connected inverter consists of a set of complex-valued poles and complex-valued residuals. Its mathematical form is essentially in the complex domain, and the calculation process involves complex multiplication, complex addition, and the resulting amplitude-phase changes. It shares the same algebraic foundation as complex-valued neural networks. Therefore, using a complex-valued neural network as the modeling framework for port admittance can preserve the complex-domain characteristics of the admittance model and its frequency-related phase change patterns within the network, enabling the network to more directly express the amplitude-phase characteristics of the admittance, thereby achieving higher fitting performance in complex frequency domain modeling.

[0065] In complex-valued neural networks, complex-valued neurons are typically implemented using two sets of real-valued parameters. The formulas for setting the complex-valued weights and biases of a complex-valued neural network are as follows:

[0066] ;

[0067] ;

[0068] In the formula, For the weights of the complex neural network, For the bias of the complex-valued neural network, Let the real part of the weight be . For the real part of the bias, The imaginary part of the weights, For the imaginary part of the bias, It is an ordinal unit.

[0069] The multiplication and addition operations between complex-valued weights and biases and the input can be implemented in the real domain through a set of real-valued matrix operations, when the input is complex. Then, its unified matrix calculation form is obtained, and the formula is as follows:

[0070] ;

[0071] ;

[0072] In selecting the activation function, it is necessary to ensure that the loss function and the network output are differentiable with respect to the real and imaginary parts of each complex-valued parameter, so that the network can be trained using the backpropagation algorithm. In engineering implementation, differentiable real-valued activation functions can be applied to the real and imaginary parts of the complex-valued signal respectively, thereby ensuring that both parts can be effectively updated with gradients. This invention applies the Swish activation function to the real and imaginary channels respectively, as shown in the following formula:

[0073] ;

[0074] This form is still based on real-valued operators in its implementation, which facilitates differentiation and numerical computation, while providing complex-valued neural networks with smooth and sufficiently nonlinear response characteristics.

[0075] Step S32: Construct the loss function of the complex-valued neural network by combining the port admittance matrix. This invention divides the forward computation process of the complex-valued neural network into two parts: "encoding" and "decoding".

[0076] During the encoding phase, the operating condition information OP describing the stable operating point of the three-phase grid-connected inverter is taken as input. After multi-layer complex-valued linear transformation and nonlinear activation in step S31, the network outputs a set of complex-valued poles under this operating condition. and complex residual The encoding stage learns the mapping relationship between the operating conditions and the parameters, and the structure of the parameters is determined by the aforementioned port admittance matrix and pole-residual derivation results. A "decoding" stage is introduced after the network output to convert the complex-valued poles... and complex residual Substituting the pole-residual expansion derived from the physical model, the elements of the port admittance matrix are constructed at any frequency point, as shown in the following formula:

[0077] ;

[0078] In the formula, The fundamental angular frequency at the steady-state operating point (OP) The corresponding admittance fitting value.

[0079] Therefore, under a given operating condition (OP), the reconstructed admittance fitting value can be obtained based on the poles and residual parameters of the network output. To ensure that the network output parameters approximate the reference admittance in the frequency domain and satisfy the physical constraints of the poles, a complex-valued neural network loss function is constructed. The formula is as follows:

[0080] ;

[0081] ;

[0082] ;

[0083] In the formula, The fitting loss function for the port admittance data. These are the weighting coefficients. The extreme point physical constraint loss function, The serial number of the stable operating point. The index of the frequency point. For the first Under the stable operating point OP, the first The actual admittance value corresponding to each frequency point For the first Under the stable operating point OP, the first The admittance fitting value corresponding to each frequency point The penalty coefficient for the pole stability constraint is... For stability margin, physical constraints, by penalizing cases that are close to or greater than a given value, suppress the real parts of poles within the left half-plane, thereby ensuring that the small-signal system corresponding to the reconstructed admittance meets the stability requirements.

[0084] Step S4 includes the following sub-steps:

[0085] Step S41: Construct a multi-condition port admittance sample set and divide it into training, validation and test datasets.

[0086] The overall structure of a complex-valued neural network is as follows: Figure 4 As shown. Within the stable operating range of the three-phase grid-connected inverter, the operating condition information of the grid connection point (including the fundamental voltage) is selected. Active power and reactive power Several typical combinations of these conditions yield a set of steady-state operating conditions. Under each condition, electromagnetic transient simulation is used to... Figure 1 The three-phase grid-connected inverter shown was subjected to a small-signal frequency disturbance test. Port voltage and current responses were collected at a preset set of frequency points, and the elements of the reference port admittance matrix were calculated accordingly, forming a sample set. Subsequently, the operating condition sample set was divided into a training set, a validation set, and a test set according to a certain proportion: the training set was used to learn the parameters of the complex-valued neural network based on the loss function defined in step S32; the validation set was used to monitor the generalization performance of the model and adjust the network hyperparameters during training; and the test set was used to evaluate the accuracy and stability of admittance prediction.

[0087] Step S42: Train and validate the physical constraint-based complex-valued neural network. In each iteration, randomly select several operating condition information and corresponding frequency points from the training set constructed in step S41. This is used as a training batch input to a complex-valued neural network. Since the input at each frequency point is a real number, the imaginary part can be set to zero, meaning that the input at each operating condition and frequency point... Add the imaginary part later To obtain the frequency point Active power reactive power and fundamental voltage Complex-valued neurons in the input layer.

[0088] A gradient descent-based optimization method is employed to train a complex-valued neural network. The complex-valued weights, biases, and outputs of each layer are uniformly represented as two sets of real-valued parameters, one with real parts and one with imaginary parts, and participate in forward and backward computations in real-valued form. The loss function is constructed as a single real-valued function using the real and imaginary errors of the admittance and the physical constraints of the poles. Based on this real-valued representation, the existing automatic differentiation and gradient optimization algorithm Adam can be directly applied to all real and imaginary parameters, thereby achieving joint updates of the complex-valued weights and biases. This training method is numerically compatible with conventional real-valued neural networks and mathematically equivalent to simultaneously optimizing the magnitude and phase characteristics of the port admittance, enabling the complex-valued neural network to converge stably in the complex domain.

[0089] Using the validation set defined in step S41, the training results under different hyperparameter configurations are evaluated. Appropriate hyperparameters, such as the learning rate, training epochs, and network structure, are obtained by monitoring changes in the validation loss. After training, the operating condition information corresponding to the test set is input into the trained complex-valued neural network to obtain the poles, residuals, and reconstructed port admittances for each operating condition. These are compared with the reference port admittance obtained from electromagnetic simulation, and the amplitude and phase errors within different frequency ranges are statistically analyzed to evaluate the prediction accuracy of the trained complex-valued neural network. Through the above training and validation process, parameters suitable for the port admittance matrix of a three-phase grid-connected inverter under multiple operating conditions are obtained.

[0090] Simply put, a small-signal flow graph of a three-phase grid-connected inverter is constructed in a synchronous coordinate system, and the port admittance matrix at the steady-state operating point is derived. The port admittance is then simplified to a pole-residual form, and the frequency domain admittance characteristics are characterized by complex-valued poles and complex-valued residual parameters. Using a complex-valued neural network as the modeling framework for port admittance preserves the complex-domain characteristics of the admittance model and its frequency-dependent phase variation within the network, enabling the network to more directly express amplitude and phase characteristics. Based on this, using stable operating conditions at the grid connection point as input and the pole-residual parameters as the output of the complex-valued neural network, the network directly learns the variation of poles and residuals with different operating conditions. By constructing a loss function that combines admittance fitting error with pole stability requirements, a physical constraint complex-valued neural network is built. The network is then trained and validated offline using port admittance data. This allows the trained complex-valued neural network to quickly identify the corresponding port admittance parameters in practical applications, given only the current operating condition. Under the constraints of the physical model, the network achieves rapid identification of port admittance under multiple operating conditions, while effectively compressing the complexity of the neural network and improving physical interpretability.

[0091] As can be seen from the accompanying figures, compared with the prior art, the technical solution provided by this invention uses a complex-valued neural network as the modeling framework for port admittance. This allows the complex-domain characteristics of the port admittance model and its frequency-related phase change law to be preserved within the complex-valued neural network, enabling the complex-valued neural network to have a more direct ability to express amplitude and phase characteristics. Furthermore, by using the operating condition information of the grid connection point's stable operating condition as input and the pole-residue as the output of the complex-valued neural network, the network can directly learn the laws governing the changes of poles and residuals under different operating conditions. By constructing a physically constrained complex-valued neural network through a loss function and training it using port admittance data, the complex-valued neural network can quickly identify port admittance based solely on the operating condition information of the current operating condition, effectively compressing the complexity of the neural network and improving physical interpretability. In short, the technical solution provided by this invention has the advantages of low computational cost, high identification efficiency, support for rapid identification under multiple operating conditions, and high physical interpretability of the fitting results.

Claims

1. A method for admittance identification of a three-phase grid-connected inverter based on a complex-valued neural network, characterized in that, Includes the following steps: Step S1: Construct the small-signal flow graph of the three-phase grid-connected inverter in the synchronous coordinate system; construct the port admittance matrix of the three-phase grid-connected inverter at the steady-state operating point based on the small-signal flow graph; Step S2: Determine the characteristic polynomial corresponding to the port admittance matrix and factor it to obtain the unified pole set of the three-phase grid-connected inverter; construct the pole-residual expansion of the port admittance matrix based on the unified pole set, and determine the transfer function of each element in the port admittance matrix. Step S3: After applying the Swish activation function to the real and imaginary channels of the complex-valued neural network, construct the loss function of the complex-valued neural network by combining the transfer function of each element in the port admittance matrix; Step S4: Train the complex-valued neural network using the loss function; Step S5: Input the operating condition information and frequency point of the grid connection point into the trained complex value neural network to obtain the port admittance under the current operating condition; Step S3 specifically includes: Step S31: Set the complex weights and complex biases of the complex-valued neural network, as shown in the following formula: ； ； In the formula, For the weights of the complex neural network, For the bias of the complex-valued neural network, Let the real part of the weight be . For the real part of the bias, The imaginary part of the weights, For the imaginary part of the bias, It is an ordinal unit; The Swish activation function is applied to both the real and virtual channels to obtain the response output. The formula is as follows: ； ； ； In the formula, Indicates a complex value. For complex values The real part, For complex values The imaginary part; Step S32: Construct the loss function for the complex-valued neural network, including the encoding and decoding stages; During the encoding phase, the operating condition information describing the stable operating point (OP) of the three-phase grid-connected inverter is taken as input, and step S31 is executed, whereby the complex-valued neural network outputs the complex-valued poles under the corresponding operating condition. and complex residual During the decoding stage, the complex-valued poles are... and complex residual Substitute the port admittance matrix to calculate the fundamental angular frequency at the steady-state operating point (OP). Corresponding admittance fitting value The formula is as follows: ； ； ； Loss function for constructing complex valued neural networks The formula is as follows: ； ； ； In the formula, The fitting loss function for the port admittance data. These are the weighting coefficients. The extreme point physical constraint loss function, The serial number of the stable operating point. The index of the frequency point. For the first Under the stable operating point OP, the first The actual admittance value corresponding to each frequency point For the first Under the stable operating point OP, the first The admittance fitting value corresponding to each frequency point The penalty coefficient for the pole stability constraint is... For stability margin.

2. The method for identifying the admittance of a three-phase grid-connected inverter based on a complex-valued neural network as described in claim 1, characterized in that, The construction of the port admittance matrix of the three-phase grid-connected inverter at the steady-state operating point based on the small-signal flow graph specifically involves determining the filter admittance matrix based on the small-signal flow graph, as shown in the following formula: ； ； In the formula, For the Laplace operator, Here is the filter admittance matrix. For the filter circuit inductor, For the filter circuit resistor, This refers to the fundamental angular frequency of a three-phase grid-connected inverter. The filter admittance matrix The product of the identity negative matrix; Based on the circuit topology and connection sequence of the control elements in the small-signal flow diagram, the overall transmission relationship between port voltage disturbance and output current disturbance is clarified, as shown in the following formula: ； ； ； ； ； In the formula, This represents the time delay transfer function matrix caused by digital control and PWM. The equivalent time delay of a three-phase grid-connected inverter. The transfer function matrix for decoupling the current loop controller. The proportional gain of the current loop. This is the integral gain of the current loop. The transfer function for the output angle and q-axis voltage of the phase-locked loop. The proportional gain of the current loop. This is the integral gain of the current loop. The fundamental voltage at the steady-state operating point; Let be the transfer function of the small-signal disturbance path from the grid voltage to the control current in the synchronous coordinate system. The q-axis component of the grid current at the steady-state operating point. The d-axis component of the grid current at the steady-state operating point. Let be the transfer function of the small-signal disturbance path from the grid voltage to the control voltage in the synchronous coordinate system. This represents the q-axis component of the inverter modulation coefficient at the steady-state operating point. The d-axis component of the inverter modulation coefficient at the steady-state operating point; The port admittance matrix of the three-phase grid-connected inverter under steady-state operating point is constructed using the following formula: ； In the formula, Here is the port admittance matrix. , , and These are the four elements of the port admittance matrix. It is the identity matrix. This represents the gain of the PWM modulator control circuit.

3. The method for identifying the admittance of a three-phase grid-connected inverter based on a complex-valued neural network as described in claim 2, characterized in that, Step S2 specifically includes: determining the frequency domain transfer function corresponding to the dynamic characteristics of the filter circuit, the current PI controller, the time delay element, and the phase-locked loop; The characteristic polynomial of the port admittance model is determined by superimposing the frequency domain transfer functions through the cascaded small-signal flow graph. The formula is as follows: ； In the formula, to These are all characteristic polynomial coefficients of a three-phase grid-connected inverter; Factorization of characteristic polynomial The unified pole set of the three-phase grid-connected inverter is obtained by the following formula: ； In the formula, For the first A unified pole; The pole-residual expansion of the port admittance matrix is ​​constructed based on a unified pole set, and the transfer function of each element in the port admittance matrix is ​​determined, as shown in the following formula: ； ； ； In the formula, For elements Regarding the first A unified pole The corresponding residual coefficients.

4. The method for identifying the admittance of a three-phase grid-connected inverter based on a complex-valued neural network as described in claim 3, characterized in that, Step S4 specifically includes: Within the stable operating range of the three-phase grid-connected inverter, several combinations of operating condition information collected at the grid connection point are used to obtain a set of steady-state operating conditions. Under each operating condition, the port voltage and current response are collected at a preset set of frequency points, and the elements of the reference port admittance matrix are calculated to form a sample set. The sample set is divided into a training set and a validation set; Based on training set and loss function A complex-valued neural network is trained using the gradient descent algorithm, and the weights of the complex-valued neural network are adjusted. and bias ; The hyperparameters of the complex-valued neural network are adjusted during training using a validation set; The operating condition information and frequency point of the grid connection point are input into the complex value neural network to identify the port admittance under the current operating condition. The operating condition information includes fundamental voltage, active power, and reactive power.

Images