Ideal transformer interface stability enhancement method and system based on pinn-nis

Abstract

The application discloses an ideal transformer interface stability enhancement method and system based on PINN-NIS, and belongs to the technical field of power hardware-in-the-loop simulation.The ideal transformer interface model with an inductor element existing at a physical side interface is established as an original circuit model, a numerical integral substitution is used to convert the inductor element into a substitution circuit model in which an equivalent resistor and a current source are connected in parallel, a physical information neural network is constructed, and a total loss function containing a voltage difference matching loss and a physical constraint loss is defined, a two-stage optimization strategy is used to train the physical information neural network, and a static perception network model is obtained, and finally, real-time circuit state quantities are input into the static perception network model, and the equivalent resistance value at the time when the stability is the strongest is obtained.The application has advantages in improving the stability of an ideal transformer interface device, and provides an effective technical solution for high-precision and high-stability operation of a high-power controllable grid-connected interface device in grid fault simulation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power hardware-in-the-loop simulation, and particularly relates to an ideal transformer interface stability enhancement method and system based on PINN-NIS (physical information neural network-numerical integral substitution). BACKGROUND

[0002] The degree of power electronicization of power systems is increasing, and the dynamic response speed is faster, and the nonlinear coupling is stronger, which puts forward higher requirements for the accuracy, efficiency and stability of real-time simulation technology. As an advanced technology that connects actual physical devices to digital real-time simulation systems, power hardware-in-the-loop simulation can effectively verify the dynamic performance of controllers, protection devices and other hardware, and has become an important means for new power system research, testing and verification. In this technical framework, the ideal transformer interface model, as a key interface module connecting the digital simulation side and the physical hardware side, undertakes functions such as voltage and current transmission, impedance matching and electrical isolation, and its performance directly affects the accuracy, stability and reliability of the entire simulation system.

[0003] Traditional ideal transformer interface models have many limitations: they are prone to instability due to mismatched subsystem impedances, and parameter tuning relies on customized design; network-based equivalent interfaces require frequent calculation of equivalent impedance, which is computationally intensive; and element equivalent interfaces based on transmission lines have strict application conditions and insufficient adaptability. These problems make it difficult for existing interfaces to meet the needs of complex simulation scenarios involving a large number of power electronic switches, limiting the stability and accuracy of the simulation. Therefore, a numerical integral substitution method can be used to handle energy storage elements such as inductors and capacitors in the circuit, which are discretized into parallel (or series) models of equivalent resistors and historical current (or voltage) sources. This method is based on fixed-step numerical integration formulas (such as the trapezoidal integration method) to convert differential equations into algebraic equations, making it easy to solve within discrete time steps.

[0004] However, in practical applications, especially in the face of systems with wide frequency variation, strong nonlinearity or time-varying parameters (such as distributed energy access scenarios involving a large number of power electronic converters), traditional numerical integral substitution methods have obvious limitations: first, the equivalent resistance value is heavily dependent on the choice of simulation step size, and a large step size can introduce truncation errors, while a small step size increases computational burden, making it difficult to achieve an optimal balance between precision and efficiency; second, when the system operating point or parameters change rapidly, the fixed substitution model is difficult to adaptively adjust, leading to distortion of voltage and current waveforms at the interface, numerical oscillation and even simulation instability; third, it relies on accurate component parameters and measurement data, and the model accuracy further decreases under parameter uncertainty or measurement noise interference.

[0005] In recent years, as a new modeling method combining physical laws and data-driven, Physics-Informed Neural Networks (PINN) has shown great potential in solving scientific computing and engineering inverse problems. PINN can train a proxy model that meets the physical laws by embedding the physical constraints such as control equations and boundary conditions into the neural network loss function in the form of soft penalty terms. This provides a new idea for improving the traditional numerical integral substitution method: the implicit nonlinear mapping between the equivalent resistance and the system state (such as voltage, current and its derivative) can be learned by PINN, so as to realize adaptive and high-precision parameter substitution. However, there are still challenges in applying PINN to the optimization of PHIL interface model, including how to construct a loss function that can fit the measured data and strictly meet the physical constraints of the circuit, how to design an efficient training strategy to achieve fast convergence and high-precision solution, and how to ensure the computational efficiency of the optimized model in real-time simulation.

[0006] Therefore, there is a lack of a method in the prior art that can adaptively and accurately calculate and optimize the key parameters (such as equivalent resistance) in numerical integral substitution while ensuring physical consistency, which is difficult to meet the stringent requirements of simulation accuracy and stability of power systems with high proportion of new energy access. SUMMARY

[0007] The purpose of the present application is to improve the ideal transformer interface of numerical integral substitution, to solve the problem of insufficient simulation accuracy caused by fixed step size and rigid model parameters in numerical integral substitution, and to improve the stability of hardware-in-the-loop simulation interface devices.

[0008] The present application is implemented by the following technical solutions.

[0009] An ideal transformer interface stability enhancement method based on PINN-NIS, comprising the following steps:

[0010] S1: establishing an ideal transformer interface model of the original circuit model with an inductor element at the physical side interface, and converting the inductor element into a substitution circuit model of parallel equivalent resistance and equivalent current source through numerical integral substitution;

[0011] S2: constructing a physics-informed neural network composed of a feedforward neural network module and a physical constraint module, wherein the input of the physics-informed neural network is time information and the voltage difference of the load resistance of the original circuit model and the substitution circuit model, and the output is the predicted equivalent resistance value;

[0012] S3: defining a total loss function including voltage difference matching loss and physical constraint loss;

[0013] S4: training the physical information neural network by using a two-stage optimization strategy to obtain a static perception network model;

[0014] S5: inputting real-time circuit state quantities into the static perception network model to predict an equivalent resistance value at the time of the strongest stability.

[0015] Further, based on the Norton equivalent theorem, an inductive element can be equivalent to a substitution circuit model of a parallel connection of an equivalent resistance and an equivalent current source by using numerical integration substitution, satisfying the following relationship:

[0016] ;

[0017] wherein, is the size of the main circuit current of the substitution circuit model, t is time, is the current size at a time on the main circuit of the substitution circuit model, is the voltage across the equivalent resistance, is the equivalent resistance, , is the simulation step size, and L is the inductance.

[0018] Further, the original circuit model satisfies the following relationship:

[0019] ;

[0020] wherein, represents the current size in the original circuit model, represents the size of the load resistance of the original circuit model, represents the power supply voltage;

[0021] The substitution circuit model satisfies the following relationship:

[0022] ;

[0023] wherein, , represents the current size of the equivalent resistance in the substitution circuit model, represents the current size at a time on the load resistance of the substitution circuit model, represents the size of the load resistance of the substitution circuit model;

[0024] The expression of the voltage difference between the original circuit model and the load resistance of the substitution circuit model is:

[0025]

[0026] The equivalent resistance and the voltage difference between the original circuit model and the load resistance of the substitution circuit model The relationship is expressed as:

[0027] .

[0028] Further, the feedforward neural network module in S2 includes an input layer, multiple hidden layers and an output layer; the input variables of the input layer include time information and voltage difference values of the original circuit model and the load resistance of the substitution circuit model; the output variable of the output layer is a predicted equivalent resistance value; in the forward propagation of the feedforward neural network; the linear transformation process of each hidden layer of the feedforward neural network is expressed as weighted summation of the output of the previous layer, the weight matrix and the bias vector of the current layer, and each layer output is obtained through an activation function; the activation function is a hyperbolic tangent function.

[0029] Further, the physical constraint module in S2 includes a circuit core physical law, and the state variables of the original circuit model and the substitution circuit model are substituted into the physical law formula to calculate a physical constraint loss, and the physical constraint loss is transmitted to the loss function of the feedforward neural network to enhance the physical rationality of the output equivalent resistance.

[0030] Further, the total loss function in S3 is obtained by weighted summation of the voltage difference matching loss and the physical constraint loss according to weight coefficients.

[0031] Further, the two-stage optimization strategy in S4 includes: in the first stage, an adaptive matrix estimation algorithm is used to train the parameters of the physical information neural network; and in the second stage, a quasi-Newton method is used to finely train the parameters of the feedforward neural network to obtain a static perception network model.

[0032] Further, the ideal transformer interface stability enhancement method based on the PINN-NIS is suitable for an ideal transformer interface model in a power hardware-in-the-loop interface device.

[0033] The application also provides an ideal transformer interface stability enhancement system based on the PINN-NIS, which comprises:

[0034] A circuit modeling module is configured to establish an ideal transformer interface model with an inductive element at a physical side interface as an original circuit model, and convert the inductive element into a substitution circuit model with an equivalent resistance and an equivalent current source in parallel through numerical integral substitution;

[0035] A physical information neural network construction module is configured to construct a physical information neural network composed of a feedforward neural network module and a physical constraint module;

[0036] A loss function definition module is configured to define a total loss function including a voltage difference matching loss and a physical constraint loss;

[0037] The training optimization module is used for training the physical information neural network by adopting a two-stage optimization strategy, so as to obtain a high-precision static perception network model.

[0038] The real-time prediction module is used for inputting real-time circuit state quantities into the static perception network model, and predicting the equivalent resistance value in the most stable state.

[0039] The application further provides an electronic device, including at least one processor and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to make the at least one processor execute the ideal transformer interface stability enhancement method.

[0040] The application has the following beneficial effects:

[0041] The application combines the physical information neural network with the numerical integral substitution technology, embeds the circuit physical law in the physical constraint module, makes the equivalent resistance output conform to the physical law, and from the root source, suppresses the numerical oscillation and simulation instability problem caused by the accumulation of the truncation error generated by the traditional method due to the fixed step discretization, and dynamically optimizes the equivalent resistance value to adapt to the signal transmission timing, impedance characteristics and dynamic response matching requirements between the digital simulation side and the physical power side in the power hardware-in-the-loop simulation, so that the dynamic stability and robustness of the power hardware-in-the-loop simulation system under complex working conditions can be improved, and the simulation process can be continuous and reliable, and the result can be accurate and reliable. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The flowchart of the method of the application is shown in the figure;

[0043] Figure 2 The structure diagram of the physical information neural network of the application is shown in the figure;

[0044] Figure 3 The loss function convergence curve of the physical information neural network trained by the application is shown in the figure;

[0045] Figure 4 The interface stability comparison diagram of the application and the traditional numerical integral substitution method is shown in the figure. DETAILED DESCRIPTION

[0046] The method of the application will be further described below with reference to the accompanying drawings.

[0047] As shown in the figure, an ideal transformer interface stability enhancement method based on PINN-NIS includes the following steps: Figure 1

[0048] ​S1: establishing an ideal transformer interface model of the existence of inductive elements at the physical side interface as an original circuit model, converting the inductive elements into a substitution circuit model of equivalent resistance and equivalent current source in parallel through numerical integral substitution;

[0049] S2: constructing a physical information neural network composed of a feedforward neural network module and a physical constraint module, the input of the physical information neural network being time information and a voltage difference value of the original circuit model and the load resistance of the substitution circuit model, and the output being a predicted equivalent resistance value;

[0050] S3: defining a total loss function including a voltage difference matching loss and a physical constraint loss;

[0051] S4: training the physical information neural network by adopting a two-stage optimization strategy to obtain a static perception network model;

[0052] S5: inputting real-time circuit state quantities into the static perception network model to predict an equivalent resistance value at the strongest stability.

[0053] In the embodiment, based on the Norton equivalent principle, the inductive element of the original circuit model in the time domain simulation is approximated as a parallel combination of an equivalent resistance at the current step and a current source with the size of the main road current at the last time of the substitution circuit model by discretizing and integrating the differential equation of the inductive element, thereby providing an explicit physical parameter target for optimization of the physical information neural network.

[0054] Specifically, it is assumed that the voltage at the left end of the inductive element is v a , the voltage at the right end is v b , the current flowing through the inductor is i ab , and the differential equation of the inductor can be expressed as:

[0055] (1) ;

[0056] wherein v is the inductor voltage, and L is the inductance.

[0057] Integrating formula (1) at time to t can obtain:

[0058] (2) ;

[0059] Since:

[0060] (3) ;

[0061] (4) ;

[0062] wherein i is the inductor current at the last time, is the simulation step.

[0063] The original circuit model inductance is equivalent to an equivalent resistance and an equivalent current source by Norton equivalent, and the integral expression of formula (2) can be obtained by substituting formula (3) and formula (4) into formula (2):

[0064] ;

[0065] wherein, is the size of the main circuit current of the substitution circuit model, t is time, is the current size at a time on the main circuit of the substitution circuit model, is the voltage across the equivalent resistance, is the equivalent resistance, , is the simulation step, and L is the inductance.

[0066] The original circuit model satisfies the following relationship:

[0067] ;

[0068] wherein, represents the current size in the original circuit model, represents the size of the load resistance of the original circuit model, represents the power supply voltage;

[0069] The substitution circuit model satisfies the following relationship:

[0070] ;

[0071] wherein, , represents the current size of the equivalent resistance in the substitution circuit model, represents the current size at a time on the load resistance of the substitution circuit model, represents the size of the load resistance of the substitution circuit model;

[0072] The expression of the voltage difference between the original circuit model and the load resistance of the substitution circuit model is:

[0073] ;

[0074] The relationship between the equivalent resistance and the voltage difference between the original circuit model and the load resistance of the substitution circuit model is expressed as:

[0075] .

[0076] As Figure 2As shown, the physical information neural network mainly includes a feedforward neural network module and a physical law constraint module, wherein the feedforward neural network module is used to establish a nonlinear mapping relationship between the equivalent parameters and the circuit state variables of the numerical integral substitution.

[0077] The feedforward neural network module is composed of an input layer, a plurality of hidden layers and an output layer. The input layer is used to receive characteristic variables related to the ideal transformer interface operating state, and the input variables of the input layer include time information t and the voltage difference of the load resistance of the original circuit model and the substitution circuit model ; the hidden layer is used to perform layer-by-layer nonlinear mapping and feature extraction on the input features; and the output variable of the output layer is the predicted equivalent resistance value .

[0078] In the embodiment, each hidden layer adopts a full connection structure, and the linear mapping between two adjacent hidden layers is performed through a weight matrix and a bias vector , and a nonlinear activation function is introduced to enhance the approximation ability of the feedforward neural network to complex nonlinear relationships.

[0079] The output layer does not adopt a nonlinear activation function, but directly outputs the prediction result of the feedforward neural network through linear mapping, so as to avoid introducing additional non-physical restrictions on the equivalent parameters or physical quantities. Through the above structure design, the feedforward neural network module can realize high-precision function approximation of the numerical integral substitution parameters while ensuring the stability of the calculation.

[0080] Specifically, for the lth layer of the feedforward neural network, the linear transformation can be batch calculated through the following formula:

[0081] ;

[0082] In the formula, is an input variable, , denotes a set of weight and bias parameters of the feedforward neural network, denotes a weight matrix of the lth layer of the feedforward neural network, denotes a bias vector of the lth layer of the feedforward neural network, l denotes a level of the feedforward neural network, denotes a total number of levels of the feedforward neural network, denotes a composite function composed of a plurality of linear transformations and a nonlinear activation.

[0083] For the lth layer of the feedforward neural network, the calculation formula of the linear transformation is:

[0084] ;

[0085] In the formula, This represents the output of the (l-1)th layer of the feedforward neural network after activation by the activation function. This represents the output of the l-th layer of the feedforward neural network.

[0086] Applying the activation function to the hidden layer of a feedforward neural network yields:

[0087] ;

[0088] The activation function is the hyperbolic tangent function, with the following expression:

[0089] ;

[0090] The linear mapping of the output layer can be represented as:

[0091] ;

[0092] In summary, the output of the feedforward neural network can be expressed as:

[0093] ;

[0094] Where e is the base of the natural logarithm, Let z represent the activation function, and z be the input to the activation function. This is the output of the l-th layer of the feedforward neural network after activation by the activation function. This is the output of the (l-1)th layer of the feedforward neural network after activation by the activation function. This is the output of the l-th layer of the feedforward neural network. For the first generation of the feedforward neural network The layer is output after being activated by the activation function. For the Lth generation of the feedforward neural network NN The weight matrix of the layer, For the feedforward neural network Layer bias vector, For the feedforward neural network The weight matrix of the layer, This is the weight matrix of the first layer of the feedforward neural network. This is the bias vector of the first layer of the feedforward neural network. For the feedforward neural network The bias vector of the layer.

[0095] The physical constraint module is used to incorporate the fundamental physical laws of the circuit into the training process of the feedforward neural network. By constructing a physical constraint loss, soft constraints are applied to the output of the feedforward neural network module, ensuring that the feedforward neural network satisfies the circuit constitutive relations and Kirchhoff's laws while fitting the data. This results in the predicted equivalent resistance value. It can be represented as:

[0096] ;

[0097] Specifically, the physical law constraint module constructs the physical law constraint between the original circuit model and the substitution circuit model based on the substitution circuit model, that is:

[0098] ;

[0099] The above physical law constraint is matched with the voltage difference matching loss in the form of a loss function to jointly participate in the training of the feedforward neural network.

[0100] As shown in Figure 3 , the loss function convergence curve of the physical information neural network training of the application is given. The total loss function is composed of the voltage difference matching loss and the physical constraint loss. The voltage difference matching loss forces the output of the feedforward neural network module to match the measured voltage data; the physical constraint loss ensures that the output of the feedforward neural network module strictly satisfies the underlying circuit physical law by embedding the circuit basic law (such as Kirchhoff's voltage law) into the total loss function, thereby enhancing the generalization ability and physical consistency of the ideal transformer interface model. The expression of the total loss function is:

[0101]

[0102] In the formula, is the total loss, is the voltage difference matching loss, is the physical constraint loss, is the weight coefficient.

[0103] The expression of the voltage difference matching loss is:

[0104]

[0105] In the formula, is the voltage on the load resistor in the original circuit model, is the voltage on the load resistor in the substitution circuit model, both of which are measured data;

[0106] The expression of the physical constraint loss is:

[0107] ;

[0108] In the formula, is the predicted equivalent resistance value, which is predicted by the feedforward neural network, is the equivalent resistance, which satisfies the expression:

[0109] ;

[0110] From Figure 3 It can be seen that with the increase of the number of training iterations, the loss function as a whole presents a gradually decreasing and finally tends to be stable change trend, indicating that the physical information neural network can effectively learn the mapping relationship between the numerical integration substitution parameter and the ideal transformer interface state in the training process, while gradually meeting the circuit physical constraint conditions. In addition, in the training process, two circuit models need to be built to compare the states of the substitution circuit model and the original circuit model, so that the substitution circuit model can be equivalent to the original circuit model. Through the training of the physical information neural network, a more accurate equivalent resistance can be obtained by using the trained physical information neural network to simplify the data calculation process. When the training reaches the preset number of iterations, the loss function converges to a stable interval, indicating that the physical information neural network model constructed has obtained a relatively stable and physically consistent equivalent resistance value, providing a reliable basis for subsequent calculation of numerical integration substitution parameters.

[0111] In this embodiment, a two-stage optimization strategy is used to train the physical information neural network. In the first stage, an adaptive moment estimation algorithm is used for training, and the first-order gradient information of the adaptive moment estimation algorithm is used to realize fast convergence and obtain the global approximate equivalent resistance value. In the second stage, a quasi-Newton method is used for training, and the high-order approximation capability of the quasi-Newton method is used to finely tune the equivalent resistance value, and finally a high-precision and high-stability static perception network model is obtained.

[0112] The trained static perception network model is integrated into the ideal transformer interface of the power hardware-in-the-loop simulation system, and the equivalent resistance value with the strongest stability is dynamically calculated and output according to the real-time circuit state, replacing the traditional fixed parameter numerical integration substitution, so as to realize adaptive and accurate adjustment of the ideal transformer interface model, and significantly improve the accuracy and numerical stability of the power hardware-in-the-loop simulation.

[0113] Figure 4 The performance difference of the method proposed in the embodiment of the present application and the traditional numerical integration substitution method in the stability of the ideal transformer interface is compared. The comparison result reflects the influence of different interface modeling methods on the consistency and numerical stability of the electrical quantity at the interface under the same circuit conditions. Figure 4 It can be seen that compared with the traditional fixed parameter numerical integration substitution method and the traditional interface device, the method proposed in the present application has stronger stability under the Nyquist curve. The result shows that by introducing the physical information neural network to optimize the numerical integration substitution parameter, the stability and consistency of the ideal transformer interface can be improved.

[0114] In conclusion, the ideal transformer interface stability enhancement method based on PINN-NIS has certain advantages in improving the stability of the interface device, and can provide an effective technical solution for high-precision and high-stability operation of high-power controllable grid-connected interface devices in power grid fault simulation.

[0115] Another example of the present application provides an ideal transformer interface stability enhancement system based on PINN-NIS, comprising:

[0116] The circuit modeling module is used to establish an ideal transformer interface model with an inductive element at the physical side interface as an original circuit model, and convert the inductive element into a substitution circuit model with an equivalent resistance and an equivalent current source in parallel through numerical integral substitution;

[0117] The physical information neural network construction module is used to construct a physical information neural network composed of a feedforward neural network module and a physical constraint module;

[0118] The loss function definition module is used to define a total loss function including a voltage difference matching loss and a physical constraint loss;

[0119] The training optimization module is used to train the physical information neural network by adopting a two-stage optimization strategy to obtain a high-precision static perception network model;

[0120] The real-time prediction module is used to input real-time circuit state quantities into the static perception network model to predict the equivalent resistance value at the strongest stability.

[0121] Another embodiment of the present application provides an electronic device, comprising at least one processor and a memory connected in communication with the at least one processor, wherein the memory stores instructions executable by the at least one processor, characterized in that the instructions are executed by the at least one processor to make the at least one processor execute the ideal transformer interface stability enhancement method.

[0122] Although the embodiments of the present application and the drawings are disclosed for the purpose of illustration, those skilled in the art can understand that various alternatives, changes and modifications are possible without departing from the spirit and scope of the present application and the appended claims, therefore, the scope of the present application is not limited to the disclosed content of the embodiments and the drawings.

Claims

1. A PINN-NIS-based ideal transformer interface stability enhancement method, characterized in that, The method comprises the following steps: S1: establishing an ideal transformer interface model with an inductive element existing at a physical side interface as an original circuit model, and converting the inductive element into a substitution circuit model of an equivalent resistor and an equivalent current source in parallel through numerical integral substitution; S2: constructing a physical information neural network composed of a feedforward neural network module and a physical constraint module, wherein the input of the physical information neural network is time information and a voltage difference value of the original circuit model and a load resistor of the substitution circuit model, and the output is a predicted equivalent resistor resistance value; S3: defining a total loss function comprising a voltage difference matching loss and a physical constraint loss; S4: training the physical information neural network by adopting a two-stage optimization strategy to obtain a static perception network model; S5: inputting real-time circuit state quantities into the static perception network model to predict an equivalent resistor value at the strongest stability.

2. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 1, wherein, Based on the Norton equivalent theorem, the inductive element is equivalent to a substitution circuit model of an equivalent resistor and an equivalent current source in parallel through numerical integral substitution, and the following relationship is met: ； wherein, is the size of the DC current of the substitution circuit model, t is time, is the size of the current at a moment on the DC of the substitution circuit model, is the voltage across the equivalent resistance, is the equivalent resistance, , is the simulation step, L is the inductance.

3. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 2, wherein, The original circuit model meets the following relationship: ； wherein, represents the magnitude of the current in the original circuit model, represents the magnitude of the load resistance of the original circuit model, represents the magnitude of the supply voltage; The substitution circuit model meets the following relationship: ； wherein, , represents the current size of the equivalent resistance in the substitution circuit model, represents the current size of the equivalent resistance in the substitution circuit model, represents the load resistance size of the substitution circuit model; Voltage difference between original circuit model and substituted circuit model load resistance The expression is: The relationship between the equivalent resistance and the voltage difference of the original circuit model and the replacement circuit model load resistance is expressed as: R = R0+ R1 。 4. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 1, wherein, The feedforward neural network module in S2 comprises an input layer, multiple hidden layers and an output layer; the input variables of the input layer comprise time information and a voltage difference value of the original circuit model and a load resistor of the substitution circuit model; the output variable of the output layer is a predicted equivalent resistor resistance value, which is obtained in the forward propagation of the feedforward neural network; the linear transformation process of each hidden layer of the feedforward neural network is represented as weighted summation of the output of the previous layer, the weight matrix and the bias vector of the current layer, and each layer output is obtained through an activation function; the activation function is a hyperbolic tangent function.

5. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 1, wherein, The physical constraint module in S2 comprises a circuit core physical law, and the state variables of the original circuit model and the substitution circuit model are substituted into the physical law formula to calculate a physical constraint loss, which is then input into the loss function of the feedforward neural network to enhance the physical rationality of the output equivalent resistor.

6. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 1, wherein, The total loss function in S3 is obtained by weighted summation of the voltage difference matching loss and the physical constraint loss according to a weight coefficient.

7. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 1, wherein, The two-stage optimization strategy in S4 comprises: in the first stage, training the parameters of the physical information neural network by using an adaptive moment estimation algorithm; and in the second stage, finely training the parameters of the feedforward neural network by using a quasi-Newton method to obtain a static perception network model.

8. The PINN-NIS-based ideal transformer interface stability enhancement method of claim 1, wherein, The method is suitable for an ideal transformer interface model in a power hardware-in-the-loop interface device.

9. A PINN-NIS based ideal transformer interface stability enhancement system, characterized in that, The method comprises: a circuit modeling module, which is used to establish an ideal transformer interface model with an inductive element existing at a physical side interface as an original circuit model, and convert the inductive element into a substitution circuit model of an equivalent resistor and an equivalent current source in parallel through numerical integral substitution; a physical information neural network construction module, which is used to construct a physical information neural network composed of a feedforward neural network module and a physical constraint module; a loss function definition module, which is used to define a total loss function comprising a voltage difference matching loss and a physical constraint loss; The training optimization module is configured to train the physical information neural network by using a two-stage optimization strategy, so as to obtain a high-precision static perception network model. The real-time prediction module is configured to input real-time circuit state quantities into the static perception network model, and predict an equivalent resistance value in a most stable state.

10. An electronic device comprising: At least one processor and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to perform the ideal transformer interface stability enhancement method in any one of claims 1-8.

Images