Power electronics equipment mechanism data fusion modeling method, device, medium and product

Abstract

The application relates to a power electronic equipment mechanism data fusion modeling method, device, medium and product. The method comprises the following steps: acquiring a node voltage equation group of preset nodes of components of a power electronic equipment nonlinear system, solving the node voltage equation group according to preset conditions to obtain the value of a nonlinear state variable of the nonlinear system; constructing a neural network based on a Koopman operator as a first neural network; based on the value of the nonlinear state variable, the first neural network is iteratively trained for a plurality of times within a preset period, wherein the loss value of the first neural network is determined after each training; after the last training within the preset period is completed, if the determined loss value of the first neural network is less than a preset target loss value, a mechanism data fusion model of the nonlinear system is obtained according to the first neural network that has completed the plurality of iterative trainings within the preset period; and efficient and accurate modeling of the power electronic equipment is realized.

Description

Technical Field

[0001] This application relates to the field of power electronic equipment modeling technology, and in particular to a method, device, medium and product for power electronic equipment mechanism data fusion modeling. Background Technology

[0002] In the modeling of power electronic equipment, traditional techniques typically employ mechanistic modeling methods such as analytical formulas, equivalent circuits, and finite element calculations. However, these mechanistic modeling methods involve computationally intensive methods that are sometimes too large to converge. Therefore, there is an urgent need to propose a new modeling method for power electronic equipment. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, device, medium, and product for fusion modeling of power electronic equipment mechanism data that can achieve efficient and accurate modeling of power electronic equipment, addressing the aforementioned technical problems.

[0004] In a first aspect, this application provides a method for modeling the mechanism data of power electronic equipment, wherein the power electronic equipment includes at least a nonlinear system;

[0005] The modeling method includes:

[0006] Obtain the set of node voltage equations for the preset nodes of the components of the nonlinear system, and solve the set of node voltage equations according to preset conditions to obtain the values ​​of the nonlinear state variables of the nonlinear system.

[0007] Construct a neural network based on the Koopman operator as the first neural network;

[0008] Based on the values ​​of the nonlinear state variables, the first neural network is trained iteratively multiple times within a preset time period. These multiple iterations are performed sequentially, and the loss value of the first neural network is determined after each training iteration.

[0009] If the loss value of the first neural network is less than the preset target loss value after the last training within the preset time period, the mechanism data fusion model of the nonlinear system is obtained based on the first neural network that has completed multiple iterations of training within the preset time period.

[0010] In one embodiment, the first neural network is trained iteratively multiple times within a preset time period based on the value of the nonlinear state variable. The multiple iterative training sessions within the preset time period are performed sequentially, and a loss value for the first neural network is determined after each training iteration, including:

[0011] Based on the value of the nonlinear state variable, the first neural network, which has completed the (i-1)th parameter update, is trained for the i-th time, where i is a positive integer and 1 < i;

[0012] Based on a preset loss function and according to the value of the nonlinear state variable, the i-th loss value of the first neural network after completing the i-th training is determined;

[0013] Based on the i-th loss value, control the first neural network, which has completed the i-th training, to perform the i-th parameter update;

[0014] Based on the value of the nonlinear state variable, the first neural network, which has completed the i-th parameter update, is trained for the (i+1)th time.

[0015] In one embodiment, the modeling method further includes:

[0016] Based on the value of the nonlinear state variable, the first neural network is trained for the first time, wherein the parameters of the first neural network trained for the first time are preset initial values.

[0017] Based on a preset loss function and according to the value of the nonlinear state variable, the first loss value of the first neural network after the first training is determined;

[0018] Based on the first loss value, the first neural network, which has completed the first training, is controlled to perform the first parameter update.

[0019] In one embodiment, the preset time period includes the first to the Mth step sub-time periods performed sequentially, where M is a positive integer, 2≤M, at least two step sub-time periods are the same or different, the sum of all step sub-time periods is equal to the preset time period, each step sub-time period is configured with a preset expected loss value, the preset expected loss value of the jth step sub-time period is less than the preset expected loss value of the (j-1)th step sub-time period, where j is a positive integer, 1<j≤M, and the preset expected loss value of the Mth step sub-time period is equal to the preset target loss value;

[0020] The modeling method also includes:

[0021] If the loss value of the first neural network is less than the preset expected loss value of the (j-1)th step sub-time period after the last training is completed, the first neural network that has completed multiple iterations of training in the (j-1)th step sub-time period will continue to undergo multiple iterations of training in the jth step sub-time period, wherein the multiple iterations of training in the step sub-time period are performed sequentially.

[0022] If the loss value of the first neural network is greater than or equal to the preset expected loss value of the (j-1)th step sub-time period after the last training is completed, the training is terminated.

[0023] In one embodiment, the step of performing the i-th training on the first neural network, which has completed the (i-1)-th parameter update, based on the value of the nonlinear state variable includes:

[0024] The values ​​of the nonlinear state variables are mapped from the nonlinear space to the linear space by the encoder to obtain the values ​​of the linear state variables;

[0025] Based on the general solution of the nodal voltage equations, and according to the Koopman eigenvalues ​​and the values ​​of the linear state variables, the predicted values ​​of the linear state variables are obtained by dynamically evolving in the linear space.

[0026] The predicted values ​​of the linear state variables are mapped from the linear space back to the nonlinear space by the decoder.

[0027] In one embodiment, the first neural network further includes an auxiliary neural network;

[0028] The modeling method also includes:

[0029] Based on the values ​​of the linear state variables, the Koopman feature values ​​are determined by the auxiliary neural network.

[0030] In one embodiment, the first neural network includes an encoder and a decoder, wherein the encoder is a Koopman feature function-based encoder, and the decoder corresponds to the encoder;

[0031] Based on a preset loss function and according to the value of the nonlinear state variable, the i-th loss value of the first neural network after the i-th training is determined, including:

[0032] Based on the preset autoregressive error function and according to the value of the nonlinear state variable, the i-th autoregressive error of the first neural network after the i-th training is determined;

[0033] Based on a preset prediction error function and according to the value of the nonlinear state variable, the i-th prediction error of the first neural network after completing the i-th training is determined;

[0034] Based on a preset linear error function and according to the value of the nonlinear state variable, the i-th linear error of the first neural network after the i-th training is determined;

[0035] Based on the i-th autoregressive error and its corresponding preset first weighting coefficient, the i-th prediction error and its corresponding preset second weighting coefficient, and the i-th linear error and its corresponding preset third weighting coefficient, the i-th loss value of the first neural network after completing the i-th training is determined.

[0036] Secondly, this application also provides a power electronic equipment mechanism data fusion modeling device, wherein the power electronic equipment includes at least a nonlinear system;

[0037] The modeling device includes:

[0038] The nonlinearization module is used to obtain the set of node voltage equations for the preset nodes of the components of the nonlinear system, and solve the set of node voltage equations according to preset conditions to obtain the values ​​of the nonlinear state variables of the nonlinear system.

[0039] The neural network module is used to construct a neural network based on the Koopman operator, which serves as the first neural network.

[0040] The network training module is used to perform multiple iterative trainings on the first neural network within a preset time period based on the value of the nonlinear state variable. The multiple iterative trainings within the preset time period are performed sequentially, and the loss value of the first neural network is determined after each training.

[0041] The model building module is used to obtain the mechanism data fusion model of the nonlinear system based on the first neural network that has completed multiple iterations of training within the preset time period, provided that the loss value of the first neural network is less than the preset target loss value after the last training within the preset time period.

[0042] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0043] Fourthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0044] The aforementioned method, device, medium, and product for mechanistic data fusion modeling of power electronic equipment, where the power electronic equipment includes at least a nonlinear system, involves obtaining the nodal voltage equations of preset nodes of the components in the nonlinear system and solving these equations according to preset conditions to obtain the values ​​of the nonlinear state variables of the nonlinear system. Then, a neural network based on the Koopman operator is constructed as the first neural network. Furthermore, based on the values ​​of the nonlinear state variables, the first neural network undergoes multiple iterative training sessions within a preset time period. These iterative training sessions are performed sequentially, and the loss value of the first neural network is determined after each training session. Finally, if the loss value of the first neural network determined after the last training session within the preset time period is less than a preset target loss value, a mechanistic data fusion model of the nonlinear system is obtained based on the first neural network that has undergone multiple iterative training sessions within the preset time period, thereby realizing the mechanistic data fusion modeling of power electronic equipment.

[0045] Thus, this embodiment of the application integrates the mechanistic relationship of the nonlinear system of power electronic equipment into the data modeling of the nonlinear system through a deep learning method based on the Koopman operator, thereby establishing a mechanistic data fusion model of the nonlinear system. That is, the mechanistic data fusion model is used to replace the pure mechanistic model of the nonlinear system. This not only makes full use of the efficiency and accuracy of data modeling, achieving efficient and accurate modeling of the nonlinear system, but also achieves efficient and accurate modeling of power electronic equipment. At the same time, compared with the black box model established by data modeling in traditional technology, since the mechanistic relationship of the nonlinear system is integrated into the data modeling of the nonlinear system, the interpretability of the mechanistic data fusion model is also stronger, thus making the mechanistic data fusion model of power electronic equipment more interpretable. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating a method for modeling the mechanism data of power electronic equipment in one embodiment;

[0048] Figure 2A This is a schematic diagram of the topology of a single IGBT and a single FWD in one embodiment;

[0049] Figure 2B for Figure 2A grayscale image;

[0050] Figure 3 This is a schematic diagram of the circuit structure involved in discretizing the hardware into an equivalent admittance and an equivalent current source using the reverse Euler method in one embodiment.

[0051] Figure 4A This illustrates the computational principle of a linearized nonlinear system using the Koopman operator in one embodiment.

[0052] Figure 4B for Figure 4A grayscale image;

[0053] Figure 5 This is one of the flowcharts illustrating step S106 in one embodiment;

[0054] Figure 6 This is a second flowchart illustrating step S106 in one embodiment;

[0055] Figure 7 This is a flowchart illustrating the power electronic equipment mechanism data fusion modeling method in one embodiment;

[0056] Figure 8A This is one of the training diagrams of the first neural network in one embodiment;

[0057] Figure 8B for Figure 8A grayscale image;

[0058] Figure 9 This is a flowchart illustrating step S502 in one embodiment;

[0059] Figure 10A This is a second schematic diagram of the training of the first neural network in one embodiment;

[0060] Figure 10B for Figure 10A grayscale image;

[0061] Figure 11 This is a flowchart illustrating step S504 in one embodiment;

[0062] Figure 12 This is a schematic diagram of a modular multilevel converter type high-voltage direct current transmission topology in one embodiment;

[0063] Figure 13 This is one of the simulation diagrams for a data fusion modeling method for the mechanism of power electronic equipment in one embodiment;

[0064] Figure 14 This is the second simulation diagram of a method for modeling the mechanism data fusion of power electronic equipment in one embodiment;

[0065] Figure 15 This is the third simulation diagram of a method for modeling the mechanism data fusion of power electronic equipment in one embodiment;

[0066] Figure 16 This is a schematic diagram of the structure of a power electronic equipment mechanism data fusion modeling device in one embodiment;

[0067] Figure 17 This is a schematic diagram of the structure of a computer device in one embodiment. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0069] (i) This application provides a method for modeling the mechanism data of power electronic equipment, which can be applied to the modeling of large-scale, modular, high-voltage, and high-capacity power electronic equipment, such as, but not limited to, the modeling of flexible DC converter valve towers, power electronic transformers, etc. Power electronic equipment includes at least a nonlinear system, for example, power electronic equipment includes a nonlinear system and other systems besides the nonlinear system.

[0070] In one exemplary embodiment, reference is made to Figure 1 A method for modeling the mechanism data of power electronic equipment is provided, which includes the following steps S102 to S108.

[0071] S102, obtain the set of node voltage equations for the preset nodes of the components of the nonlinear system of the power electronic equipment, and solve the set of node voltage equations according to the preset conditions to obtain the values ​​of the nonlinear state variables of the nonlinear system.

[0072] In this context, the nonlinear system of power electronic equipment can be understood as a system composed of parts of power electronic equipment whose mechanistic relationships are unclear, whose various electrical parameters are difficult to determine, and / or whose nonlinear solutions are too complex.

[0073] Nonlinear systems in power electronic equipment typically contain a large number of components. These components can be nonlinear or linear. Examples include, but are not limited to, nonlinear capacitors, nonlinear switching transistors, and linear resistors; among them, nonlinear switching transistors include, but are not limited to, nonlinear diodes, nonlinear bipolar junction transistors (BJTs), nonlinear metal-oxide-semiconductor field-effect transistors (MOSFETs), and nonlinear insulated-gate bipolar transistors (IGBTs).

[0074] Each component may include at least two preset nodes. Preset nodes of a component can be understood as nodes pre-defined in the component's equivalent circuit or topology. For example, refer to... Figure 2A , Figure 2A The illustrative diagrams illustrate the topologies of a single IGBT and a single FWD (Free Wheeling Diode). In the IGBT topology, the static characteristics are characterized by MOSFETs and BJTs, while the dynamic characteristics are characterized by the junction capacitances Ceg, Cdg, Cgs, Cds, and Cec, and the diffusion capacitances Ceb and Cdif. The preset nodes of the IGBT are, for example, nodes f1, f2, f3, f4, f5, f6, and f8, meaning a single IGBT may have seven preset nodes. Nodes f7, f6, and f2 can be preset nodes of the FWD, meaning a single FWD may have three preset nodes. Furthermore, Figure 2A middle , , Both represent resistance. , , All of these represent inductance; Gate, Emitter, and Collector represent the gate, emitter, and collector, respectively.

[0075] For each preset node, at least one node voltage equation is written based on its mechanistic relationship in the equivalent circuit or topology. The node voltage equation can be a nonlinear differential-algebraic equation. The node voltage equations of all preset nodes of a component are combined to obtain the node voltage equation system of that component. When solving the node voltage equation system of the component, the preset conditions may include preset initial conditions and preset boundary conditions for the node voltage equation system of the component. Preset initial conditions may include, for example, preset initial node voltage vectors. Solving the node voltage equation system of the component yields a time-domain numerical solution, which is the specific value of the nonlinear state variable. Multiple different numerical solutions at a certain time can constitute the numerical solution vector at that time, and multiple numerical solution vectors at different times can constitute a dataset. Finally, the input to the first neural network can be the dataset, inputting the values ​​of the nonlinear state variable into the first neural network in the form of a dataset.

[0076] For example, when writing the node voltage equations of a component, the electrical parameters of hardware such as MOSFETs, BJTs, and capacitors in the equivalent circuit or topology of the component are related to the node voltages of the preset nodes at both ends. Based on the electrical parameters of the hardware, the functional relationship (iv) between the current flowing through the hardware and the node voltages at both ends of the hardware can be written. Then, based on Kirchhoff's current law, the functional relationships (iv) of all hardware of the component can be linked together to form the nonlinear differential algebraic equations of the component, that is, the node voltage equations of the component.

[0077] Since the analytical solutions to the listed nodal voltage equations cannot be obtained directly, their time-domain numerical solutions must be obtained through numerical methods. Numerical methods include using the backward Euler method to discretize a hardware component into an equivalent admittance and an equivalent current source.

[0078] For example, taking a capacitor as an example, such as Figure 3 The capacitor shown in (a) has the following dynamic current-voltage (iv) relationship:

[0079] (1);

[0080] in, and Let C be the voltage and current across the capacitor, respectively. In equations (1) to (5), C represents the capacitance. For equation (1), from... arrive Integrating, we get:

[0081] (2);

[0082] Approximating the integral part of equation (2) using the inverse Euler method, we get:

[0083] ,final (3);

[0084] Equation (3) can be viewed as a... Figure 3 (b) Equivalent admittance and equivalent current source A parallel circuit, wherein:

[0085] (4);

[0086] time Equivalent current source With an interval of one time step The previous moment It is related to the capacitor voltage:

[0087] (5);

[0088] As mentioned above, each piece of hardware is discretized using the inverse Euler method into an equivalent admittance and an equivalent current source. For each piece of hardware, the value of the equivalent admittance and the value of the equivalent current source at time k are both related to the electrical parameters at time k. The electrical parameters at time k are related to the node voltages at both ends of the hardware at time k-1. Therefore, the values ​​of the equivalent admittance and the equivalent current source are only related to the node voltages at both ends of the hardware. The equivalent admittance and equivalent current source of all hardware in the equivalent circuit or topology of the components are concatenated into an equivalent admittance matrix through the topological relationships between the hardware. and equivalent current source vector Ohm's Law Formula , Given the node voltage vector, the node voltage vector at time k can be calculated using an iterative method based on the node voltage vector at time k-1.

[0089] For example, let's take solving for the node voltage vector of a component at time k (the node voltage vector represents the node voltage of each preset node of the component) as an example:

[0090] Based on the node voltage vector at time k-1, the equivalent admittance matrix at time k is obtained for the first time. and equivalent current source vector The equivalent admittance matrix obtained at time k for the first time. and equivalent current source vector Substitute into the formula The first inverse solution yields the node voltage vector. .

[0091] If the node voltage vector at time k-1 is the same as the node voltage vector obtained from the first inverse solution... If the absolute value of the difference is less than a preset threshold, then the node voltage vector will be obtained from the first inverse solution. As the node voltage vector at time k.

[0092] If the node voltage vector at time k-1 is the same as the node voltage vector obtained from the first inverse solution... If the absolute value of the difference is greater than or equal to a preset threshold, then the node voltage vector is obtained based on the first inverse solution. The equivalent admittance matrix at time k is obtained for the second time. and equivalent current source vector The equivalent admittance matrix obtained at time k in the second step. and equivalent current source vector Substitute into the formula The second inverse solution yields the node voltage vector. .

[0093] If the first inverse solution yields the node voltage vector The node voltage vector is obtained from the second inverse solution. If the absolute value of the difference is less than a preset threshold, then the node voltage vector will be obtained by the second inverse solution. As the node voltage vector at time k.

[0094] If the first inverse solution yields the node voltage vector The node voltage vector is obtained from the second inverse solution. If the absolute value of the difference is greater than or equal to a preset threshold, then the node voltage vector is obtained according to the second inverse solution. The equivalent admittance matrix at time k is obtained for the third time. and equivalent current source vector The equivalent admittance matrix obtained at time k in the third iteration. and equivalent current source vector Substitute into the formula The third inverse solution yields the node voltage vector. .

[0095] If the second inverse solution yields the node voltage vector The node voltage vector is obtained from the third inverse solution. If the absolute value of the difference is less than a preset threshold, then the node voltage vector will be obtained by the third inverse solution. As the node voltage vector at time k.

[0096] If the second inverse solution yields the node voltage vector The node voltage vector is obtained from the third inverse solution. If the absolute value of the difference is greater than or equal to a preset threshold, then the node voltage vector is obtained according to the third inverse solution. and formula The node voltage vector is obtained by continuing the fourth inverse solution. And determine whether the node voltage vector Unew obtained from the third inverse solution is the same as the node voltage vector obtained from the fourth inverse solution. The relationship between the absolute value of the difference and the preset threshold is used until the node voltage vector is obtained in this inverse solution. The node voltage vector obtained from the previous inverse solution The absolute value of the difference is less than a preset threshold, thus the node voltage vector obtained from this inverse solution is... This serves as the node voltage vector at time k, thus yielding the final node voltage vector at time k. Furthermore, when solving for the node voltage vector of a component at time k=1, the equivalent admittance matrix at time k=1 can be obtained for the first time based on a preset initial node voltage vector. and equivalent current source vector And continue with subsequent calculations.

[0097] The node voltage vector obtained at time k can be used as the vector of the values ​​of the nonlinear state variables at time k, that is, the nonlinear numerical solution vector at time k.

[0098] S104, construct a neural network based on the Koopman operator as the first neural network.

[0099] The Koopman operator is a linear, infinite-dimensional operator. The Koopman operator acts on the measurement function. This is equivalent to measuring the function. Mapping function in nonlinear systems Phase recombination, that is:

[0100] ;

[0101] ;in, This is the Koopman operator.

[0102] refer to Figure 4A , Figure 4A An illustrative diagram illustrates the computational principle of linearizing a nonlinear system using the Koopman operator, including the nonlinear numerical solution vector at time k in the nonlinear system. The linear numerical solution vector at time k in a linear system The mapping relationship; where, and The dimensions may be the same or different; multiple time points are represented as 1…k…; Let be the nonlinear numerical solution vector at time k+1 in the nonlinear system. Let be the predicted linear numerical solution vector at time k+1 in a linear system. Let be the Koopman eigenvalue matrix in the time domain, where t represents time in the time domain. As mentioned earlier, when t takes the time k, It can be obtained from the Koopman eigenvector at time k. Let K be the Koopman eigenvector at time k. Dimensions and The same dimensions for By taking the elements of the Koopman eigenvalue vector and constructing a diagonal matrix from them, and filling the remaining elements with 0, we can obtain the k-th time step. In a linear space, the prediction of the value of a linear state variable is only a prediction from time k to time k+1, where time k and time k+1 are one time step. Time step The specific size can be configured according to actual needs, and this application embodiment does not limit it; that is:

[0103] ;

[0104] ;

[0105] ;

[0106] Therefore, the Koopman operator can be used to constrain the layer structure and activation function of a basic neural network in deep learning in order to construct a neural network based on the Koopman operator, and this neural network based on the Koopman operator can be used as the first neural network.

[0107] S106, Based on the value of the nonlinear state variable, the first neural network is trained iteratively multiple times within a preset time period. The multiple iterative trainings within the preset time period are performed sequentially, and the loss value of the first neural network is determined after each training is completed.

[0108] In each training iteration, the values ​​of nonlinear state variables are input into the first neural network to train it. Based on a pre-defined loss function, a loss value is determined for the first neural network after each training iteration. This loss value reflects the accuracy of the trained first neural network. In iterative training, the loss value determined after the previous training iteration guides the training of the first neural network in the next iteration.

[0109] S108, after completing the last training within the preset time period, if the loss value of the determined first neural network is less than the preset target loss value, the mechanism data fusion model of the nonlinear system is obtained based on the first neural network that has completed multiple iterations of training within the preset time period.

[0110] For example, if the preset time period is configured to be 240 minutes, the first neural network will be iteratively trained multiple times within this 240-minute period based on the values ​​of the nonlinear state variables. When the loss value of the first neural network determined after the last training within this 240-minute period is less than the preset target loss value, it indicates that the first neural network has achieved the target accuracy after completing the iterative training within this 240-minute period. At this point, a mechanism data fusion model of the nonlinear system of power electronic equipment can be constructed based on the parameters of the first neural network that has completed the iterative training within this 240-minute period. The parameters of the first neural network include, but are not limited to, weights and biases.

[0111] It is understood that the specific size of the preset time period is not limited in the embodiments of this application, and can be configured according to actual needs; the total number of training sessions within the preset time period is determined by the duration of completing one iteration of training, and is not limited in this embodiment of the application.

[0112] (ii) The following is an illustrative explanation of how to perform each training iteration within a preset time period. This explanation is also applicable to the case where i=1:

[0113] In one exemplary embodiment, reference is made to Figure 5 Step S106 includes the following steps S502 to S508.

[0114] S502, based on the values ​​of the nonlinear state variables, perform the i-th training on the first neural network that has completed the (i-1)-th parameter update, where i is a positive integer and 1 < i.

[0115] The updating of the parameters of the first neural network can include updating the weights and biases of the first neural network; that is, the parameters of the first neural network can include the weights and biases of the first neural network. The values ​​of the nonlinear state variables can be input into the first neural network that has completed the (i-1)th parameter update to perform the i-th training on the first neural network that has completed the (i-1)th parameter update.

[0116] S504, based on the preset loss function and according to the value of the nonlinear state variable, determine the i-th loss value of the first neural network after completing the i-th training.

[0117] S506, based on the i-th loss value, control the first neural network that has completed the i-th training to perform the i-th parameter update.

[0118] S508, based on the values ​​of nonlinear state variables, performs the (i+1)th training on the first neural network that has completed the i-th parameter update.

[0119] For example, if the preset time period is configured to be 240 minutes, the first neural network will be trained iteratively multiple times within these 240 minutes based on the values ​​of the nonlinear state variables. Specifically, after the i-th training iteration is completed and the i-th loss value is determined, the i-th reference gradient can be determined based on the i-th loss value. This reference gradient is then input into the first neural network that has completed the i-th training iteration, enabling it to perform the i-th parameter update. Subsequently, the values ​​of the nonlinear state variables are input into the first neural network that has completed the i-th parameter update, allowing it to undergo the (i+1)-th training iteration.

[0120] Optionally, step S106 further includes: performing a first training on the first neural network based on the value of the nonlinear state variable, wherein the parameter values ​​of the first neural network in the first training are preset initial values, which may include preset initial weights and preset initial biases, etc.; then, determining the first loss value of the first neural network after the first training based on a preset loss function and according to the value of the nonlinear state variable; then, determining the first reference gradient according to the first loss value, and inputting the first reference gradient into the first neural network that has completed the first training, so that the first neural network that has completed the first training performs the first parameter update; then, inputting the value of the nonlinear state variable into the first neural network that has completed the first parameter update to perform a second training on the first neural network that has completed the first parameter update.

[0121] (III) The following is an example of how to perform multiple iterative training sessions within a preset time period:

[0122] In an exemplary embodiment, the preset time period includes sequentially occurring step sub-time periods from the 1st to the Mth, where M is a positive integer, 2 ≤ M, at least two step sub-time periods are the same or different, the sum of all step sub-time periods equals the preset time period, each step sub-time period is configured with a preset expected loss value, the preset expected loss value of the j-th step sub-time period is less than the preset expected loss value of the (j-1)-th step sub-time period, where j is a positive integer, 1 < j ≤ M, and the preset expected loss value of the M-th step sub-time period equals the preset target loss value. Furthermore, at least two step sub-time periods are either consecutive or spaced apart.

[0123] For example, the preset time period is configured to be 240 minutes, and M is configured to be 5. The preset time period is divided into five sequentially occurring step-time periods: the first step-time period [0~20 minutes], the second step-time period [20 minutes~40 minutes], the third step-time period [40 minutes~60 minutes], the fourth step-time period [60 minutes~120 minutes], and the fifth step-time period [120 minutes~240 minutes]. During iterative training within this 240-minute period, the first to fifth step-time periods are also performed sequentially, with multiple iterative training sessions conducted within each step-time period.

[0124] refer to Figure 6 Step S106 includes the following steps S602 to S604.

[0125] S602, after completing the last training in the (j-1)th step sub-time period, if the determined loss value of the first neural network is less than the preset expected loss value in the (j-1)th step sub-time period, the first neural network that has completed multiple iterations of training in the (j-1)th step sub-time period continues to undergo multiple iterations of training in the jth step sub-time period.

[0126] S604: After completing the last training session in the (j-1)th step time period, if the loss value of the determined first neural network is greater than or equal to the preset expected loss value in the (j-1)th step time period, the training is terminated.

[0127] In this embodiment, the significance of the step is that the preset expected loss value of the j-th sub-time period is less than the preset expected loss value of the (j-1)-th sub-time period, so as to perform step-like iterative training within the preset time period and improve modeling efficiency. In other words, if the loss value of the determined first neural network is greater than or equal to the preset expected loss value of the (j-1)th step-time period after the last training in the (j-1)th step-time period, training is terminated directly. This indicates that the first neural network has not achieved the expected accuracy after completing the iterative training in the (j-1)th step-time period, and therefore the last parameter update in the (j-1)th step-time period and all subsequent step-time periods are no longer iteratively trained. Only if the loss value of the determined first neural network is less than the preset expected loss value of the (j-1)th step-time period after the last training in the (j-1)th step-time period, this indicates that the first neural network has achieved the expected accuracy after completing the iterative training in the (j-1)th step-time period, and therefore the last parameter update in the (j-1)th step-time period and all subsequent step-time periods are continued, thereby improving modeling efficiency.

[0128] Optionally, refer to Figure 7 After exiting the training in step S604, the power electronic equipment mechanism data fusion modeling method also includes the following steps S702~S706.

[0129] S702, construct a neural network based on the Koopman operator as a second neural network, wherein the structure of the second neural network is different from that of the first neural network.

[0130] S704, based on the value of the nonlinear state variable, performs multiple iterative trainings on the second neural network within a preset time period. The multiple iterative trainings are performed sequentially, and the loss value of the second neural network is determined after each training.

[0131] S706, after completing the last training within the preset time period, if the loss value of the determined second neural network is less than the preset target loss value, the mechanism data fusion model of the nonlinear system is obtained based on the second neural network that has completed multiple iterations of training within the preset time period.

[0132] In this embodiment, after exiting training in step S604, the first neural network will no longer be trained iteratively multiple times within a preset time period based on the values ​​of nonlinear state variables. Instead, the first neural network will be discarded, and a new neural network based on the Koopman operator will be reconstructed as the second neural network. The structure of the second neural network is different from that of the first neural network. Then, the second neural network will be trained iteratively multiple times within a preset time period based on the values ​​of nonlinear state variables.

[0133] Similarly, during the multiple iterative training of the second neural network within a preset time period, if the loss value of the second neural network is greater than or equal to the preset expected loss value of the (j-1)th step sub-time period after the last training is completed, the training is terminated. The second neural network is no longer trained iteratively within the preset time period based on the values ​​of the nonlinear state variables. Instead, the second neural network is discarded, and a new neural network based on the Koopman operator is constructed as the third neural network. The structure of the third neural network is different from the structures of the second and first neural networks. Then, based on the values ​​of the nonlinear state variables, the third neural network is trained iteratively within the preset time period… This embodiment continues until a neural network based on the Koopman operator is found. After the last training within the preset time period, the loss value of the determined neural network is less than the preset target loss value. Finally, the mechanism data fusion model of the nonlinear system is obtained based on the neural network that has completed multiple iterative training within the preset time period.

[0134] (iv) The following provides a further illustrative explanation of how to perform each training iteration within a preset time period. This explanation is also applicable to the case where i=1:

[0135] In one exemplary embodiment, reference is made to Figure 8A The first neural network includes an encoder ( ) and decoder ( The encoder is based on the Koopman feature function, and the decoder corresponds to the encoder. The encoder, decoder, and Koopman operator act on the numerical solution vector as follows:

[0136] ;

[0137] ;

[0138] ;

[0139] = ;

[0140] = .

[0141] Optionally, the first neural network may also include an auxiliary neural network ( Based on the values ​​of linear state variables, Koopman feature values ​​can be determined using an auxiliary neural network. That is, = ,in, for The vector formed by the Koopman eigenvalues ​​corresponding to each element in the vector.

[0142] refer to Figure 9 In step S502, based on the value of the nonlinear state variable, the first neural network that has completed the (i-1)th parameter update is trained for the i-th time, including the following steps S902~S906.

[0143] S902 maps the values ​​of nonlinear state variables from the nonlinear space to the linear space through an encoder, thus obtaining the values ​​of linear state variables.

[0144] S904 is based on the general solution of a system of nonlinear differential-algebraic equations, and obtains the predicted values ​​of linear state variables by dynamically evolving in a linear space according to the Koopman eigenvalues ​​and the values ​​of linear state variables.

[0145] The general solution of the nonlinear differential-algebraic equation system can be based on the format of the general solution of the differential equation system. The nonlinear differential-algebraic equation system can be a set of nodal voltage equations.

[0146] S906 maps the predicted linear state variables back from the linear space to the nonlinear space through a decoder.

[0147] In this embodiment, the Koopman eigenvalues ​​correspond one-to-one with the values ​​of the linear state variables, and the values ​​of the linear state variables change as the values ​​of the nonlinear state variables change. (See reference...) Figure 10A It is understandable that, since the first neural network includes an encoder, a decoder, and an auxiliary neural network, performing multiple iterative trainings on the first neural network within a preset time period is actually performing multiple iterative trainings on the encoder, decoder, and auxiliary neural network simultaneously within the preset time period. Therefore, although the numerical solution vector input to the first neural network is the same in two adjacent training sessions, the Koopman eigenvectors are different in two adjacent training sessions because the auxiliary neural networks in the two adjacent training sessions are different. For the same training session, if the numerical solution vector input to the first neural network is different, that is, the input to the first neural network is different, the corresponding Koopman eigenvectors will also be different.

[0148] in, ; ;

[0149] n is Dimensions This is consistent with the discrete time involved in writing the nodal voltage equations using the trapezoidal rule, for .

[0150] Optionally, refer to Figure 11 In step S504, the i-th loss value of the first neural network after the i-th training is determined based on the preset loss function and the value of the nonlinear state variable, including the following steps S1102~S1108.

[0151] S1102, based on the preset autoregressive error function and according to the nonlinear state variable, determine the i-th autoregressive error of the first neural network after completing the i-th training.

[0152] For example, a pre-defined autoregressive error function for:

[0153]

[0154] MSE is the mean squared error.

[0155] S1104, Based on the preset prediction error function and according to the nonlinear state variable, determine the i-th prediction error of the first neural network after completing the i-th training.

[0156] For example, a preset prediction error function for:

[0157]

[0158] in, It is the number of prediction steps in the linear space, and can be a positive integer greater than 1.

[0159] S1106, Based on a preset linear error function and according to the nonlinear state variable, determine the i-th linear error of the first neural network after completing the i-th training.

[0160] For example, a preset linear error function for:

[0161]

[0162] Where T is the number of time steps.

[0163] S1108, based on the i-th autoregressive error and its corresponding preset first weighting coefficient, the i-th prediction error and its corresponding preset second weighting coefficient, and the i-th linear error and its corresponding preset third weighting coefficient, determine the i-th loss value of the first neural network after completing the i-th training.

[0164] In this embodiment of the application, as an important indicator for evaluating the accuracy of the mechanism data fusion model, three loss functions are introduced in the loss function part: the autoregressive error function, the prediction error function, and the linear error function. These are linearly combined, and the modeling of different nonlinear systems can be adapted by adjusting the magnitude of the first weighting coefficient, the second weighting coefficient, and / or the third weighting coefficient.

[0165] (V) The following is an exemplary simulation illustration of the data fusion modeling method for power electronic equipment mechanisms:

[0166] Figure 12 This is the basic topology of a Modular Multilevel Converter-High Voltage Direct Current (MMC-HVDC) transmission system. This topology includes multiple arms, each consisting of N sub-modules (SM) and an arm inductor (L0). Typically, the sub-modules (SM) have a half-bridge topology, which includes two IGBTs (upper and lower) and two freewheeling diodes (i.e., freewheeling diodes D1 and D2). This refers to the voltage across capacitor C0 in submodule SM. and These represent the current and output voltage of the submodule SM, respectively.

[0167] The main challenge in MMC simulation lies in its complex topology and the large number of IGBTs it contains. This results in an excessive number of circuit nodes and a large matrix solution scale. Direct solution would lead to excessively long simulation times, making it difficult to apply. Furthermore, an excessive number of MMC levels can even cause the simulation to fail to converge. System-level analysis shows that the high-frequency components of the bridge arm injected current ip (ipa, ipb, ipc) decay rapidly. Therefore, the change in bridge arm injected current ip between two adjacent time steps is very small. Thus, in this embodiment, the voltage and current of the submodule SM are decoupled from the bridge arm to perform circuit partitioning. All submodules SM are separated from their respective bridge arms to form a minimal circuit, thereby reducing the computational burden and improving numerical stability.

[0168] The following study focuses on a single MMC submodule SM. Submodule SM consists of two IGBTs (upper and lower), two freewheeling diodes (D1 and D2), and a capacitor C0. As mentioned earlier, submodule SM includes 14 preset nodes (f1~f8 multiplied by 2 = 16). Node f8 of the upper IGBT coincides with node f1 of the lower IGBT, forming a single common node. Node f8 of the lower IGBT serves as a reference node for both IGBTs. Voltage simulations using these 14 nodes yield the following results: Figure 13 As shown, the overall result error is small; among them, Figure 13 The diagram consists of two rows, each with seven columns. In the upper row, from left to right, are the diagrams of nodes f1, f2, f3, f4, f5, f6, and f8 of the upper IGBT. In the lower row, from left to right, are the diagrams of nodes f1, f2, f3, f4, f5, f6, and f8 of the lower IGBT. The line "True" represents the simulation of the numerical solution, and the line "Predicate" represents the simulation of the output of the mechanistic data fusion model. The lines "Predicate" and "True" are very similar, which sufficiently demonstrates the high accuracy of the mechanistic data fusion model's output.

[0169] In practical applications, it is impossible to measure the voltage of the nodes inside the submodule SM. To meet practical application requirements, the following parameters are selected: the input IGBT current, the control signal Vg of the input f4 node, and the output voltage of the submodule SM. The voltage of capacitor C0 in submodule SM The simulation was repeated, and the results are as follows: Figure 14 ;in, Figure 14The diagram consists of two rows of illustrations, each with three columns. In the top row, from left to right, are illustrations of the current flowing through the upper IGBT, the current flowing through the lower IGBT, and the control signal Vg of the upper IGBT. In the bottom row, from left to right, are the control signal Vg of the lower IGBT and the output voltage of the submodule SM. The voltage of capacitor C0 in submodule SM The diagram shows that the True line represents the simulation of the numerical solution, and the Predicate line represents the simulation of the output of the mechanistic data fusion model.

[0170] Besides the prediction accuracy of the mechanistic data fusion model and the simulation acceleration effect it brings, the interpretability of the mechanistic data fusion model is another key aspect of the technical solution implementation in this application. Interpretability refers to explaining the mechanistic relationships satisfied by the data during the calculation process based on theoretical foundations. The examples designed in this paper are all time-dependent nonlinear differential-algebraic equation systems. The input is the numerical solution vector of the nonlinear differential-algebraic equation system at time k, and the output is the numerical solution vector at time k+1.

[0171] Figure 15 The image shows the Koopman eigenvalues ​​found by the submodule SM using an auxiliary neural network. Data of length 128 is randomly selected from numerical solution vectors at multiple time points (e.g., from the numerical solution vector at time k=1 to the numerical solution vector at time k=128). Each numerical solution vector includes 10 numerical solutions selected from 14 numerical solutions. The values ​​are then calculated and mapped by the encoder using a function. The obtained Koopman eigenvalues ​​and the relationship between the input and output of the auxiliary neural network are presented. In this simulation, to fully demonstrate the relationship between the voltages of all 14 nodes, the input... Scaling to the interval [0, 1] corresponds to It remains unchanged.

[0172] Figure 15 The left image shows that each line represents a Koopman eigenvalue of a numerical solution vector at time points 1 to 128. Within this 128-time window, It remains almost unchanged and is distributed around the value 1. This is because during the calculation, the time length for each feed to the first neural network is 128. Within this time length, due to the simulation time step... The values ​​are relatively small, and the magnitudes of the numerical solution vectors at different time points are similar. Even though the Koopman eigenvalue at time k depends on the value of the nonlinear state variable at time k, within the relatively small time interval of 128, the Koopman eigenvalues ​​should be equal or approximately the same, and since the magnitudes of the numerical solution vectors at different time points are similar, the Koopman eigenvalues ​​should be distributed around 1.

[0173] Figure 15 The image on the right shows: The values ​​of the linear state variables of the 14 node voltages and the predicted linear state variables strictly satisfy a linear relationship. This is because the calculation process of the mechanistic data fusion model is as follows: first, the values ​​of the nonlinear state variables are mapped to the linear space through an encoder; then, the Koopman eigenvalues ​​corresponding to the linear state variable values ​​are calculated through an auxiliary neural network; and finally, the calculation... The prediction is completed, which is also the calculation principle required by the general solution format of the nonlinear differential algebraic equation system. To satisfy this principle, the values ​​of the linear state variables and the predicted values ​​of the linear state variables strictly satisfy the linear relationship, and the simulation results are consistent with the theoretical inference.

[0174] This application applies the Koopman operator to the data modeling process, essentially mapping the nonlinear system to a linear system through the data model, transforming the iterative solution process into a simple computational process, thereby simplifying the modeling process. The entire modeling scheme of this application includes at least: solving the nodal voltage equations, designing the loss function, and designing the neural network architecture. In particular, the introduction of the Koopman operator realizes the linearization of the nonlinear system and accelerates the solution of the refined model of power electronic equipment.

[0175] In one exemplary embodiment, reference Figure 16 A data fusion modeling device for the mechanism of power electronic equipment is also provided. The power electronic equipment includes a nonlinear system, and the modeling device includes:

[0176] The nonlinearization module 1610 is used to obtain the set of node voltage equations of the preset nodes of the components of the nonlinear system, and solve the set of node voltage equations according to the preset conditions to obtain the values ​​of the nonlinear state variables of the nonlinear system.

[0177] Neural network module 1620 is used to construct a neural network based on the Koopman operator, serving as the first neural network;

[0178] The network training module 1630 is used to train the first neural network multiple times within a preset time period based on the values ​​of nonlinear state variables. The multiple iterations within the preset time period are performed sequentially, and the loss value of the first neural network is determined after each training.

[0179] The model building module 1640 is used to obtain a mechanism data fusion model of the nonlinear system based on the first neural network that has completed multiple iterations of training within the preset time period, provided that the loss value of the determined first neural network is less than the preset target loss value after the last training within the preset time period.

[0180] In an exemplary embodiment, the network training module 1630 is further configured to perform the i-th training on the first neural network that has completed the (i-1)-th parameter update based on the value of the nonlinear state variable, where i is a positive integer and 1 < i; determine the i-th loss value of the first neural network after completing the i-th training based on a preset loss function and the value of the nonlinear state variable; control the first neural network that has completed the i-th training to perform the i-th parameter update based on the i-th loss value; and perform the i+1-th training on the first neural network that has completed the i-th parameter update based on the value of the nonlinear state variable.

[0181] In an exemplary embodiment, the network training module 1630 is further configured to perform a first training on the first neural network based on the value of a nonlinear state variable, wherein the parameter values ​​of the first neural network for the first training are preset initial values; determine the first loss value of the first neural network after the first training is completed based on a preset loss function and according to the value of the nonlinear state variable; and control the first neural network that has completed the first training to perform a first parameter update according to the first loss value.

[0182] In an exemplary embodiment, the preset time period includes the first to the Mth step sub-time periods performed sequentially, where M is a positive integer, 2≤M, at least two step sub-time periods are the same or different, the sum of all step sub-time periods equals the preset time period, each step sub-time period is configured with a preset expected loss value, the preset expected loss value of the jth step sub-time period is less than the preset expected loss value of the (j-1)th step sub-time period, where j is a positive integer, 1<j≤M, and the preset expected loss value of the Mth step sub-time period equals the preset target loss value; the network training module 1630 is further configured to, after completing the (j-1)th step sub-time period After the last training session within the time frame, if the loss value of the determined first neural network is less than the preset expected loss value of the (j-1)th step time frame, the first neural network that has completed multiple iterations of training within the (j-1)th step time frame continues to undergo multiple iterations of training within the j-th step time frame. The multiple iterations of training within the step time frame are performed sequentially. The network training module 1630 is also used to exit training if, after the last training session within the (j-1)th step time frame, the loss value of the determined first neural network is greater than or equal to the preset expected loss value of the (j-1)th step time frame.

[0183] In an exemplary embodiment, the first neural network includes an encoder and a decoder, wherein the encoder is a Koopman feature function-based encoder, and the decoder corresponds to the encoder; the network training module 1630 is further configured to map the values ​​of nonlinear state variables from the nonlinear space to the linear space through the encoder to obtain the values ​​of linear state variables; based on the general solution of the nodal voltage equations and according to the Koopman feature values ​​and the values ​​of linear state variables, dynamically evolve in the linear space to obtain the predicted values ​​of linear state variables; and map the predicted values ​​of linear state variables from the linear space back to the nonlinear space through the decoder.

[0184] In one exemplary embodiment, the first neural network further includes an auxiliary neural network; the network training module 1630 is also used to determine Koopman feature values ​​based on the values ​​of linear state variables through the auxiliary neural network.

[0185] In an exemplary embodiment, the network training module 1630 is further configured to: determine the i-th autoregressive error of the first neural network after the i-th training based on a preset autoregressive error function and the value of the nonlinear state variable; determine the i-th prediction error of the first neural network after the i-th training based on a preset prediction error function and the value of the nonlinear state variable; determine the i-th linear error of the first neural network after the i-th training based on a preset linear error function and the value of the nonlinear state variable; and determine the i-th loss value of the first neural network after the i-th training based on the i-th autoregressive error and its corresponding preset first weighting coefficient, the i-th prediction error and its corresponding preset second weighting coefficient, and the i-th linear error and its corresponding preset third weighting coefficient.

[0186] Each module in the aforementioned power electronic equipment mechanism data fusion modeling device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0187] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 17As shown. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores relevant data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements a data fusion modeling method for the mechanism of power electronic equipment. Those skilled in the art will understand that... Figure 17 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0188] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in any of the above method embodiments.

[0189] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0190] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one of relational databases and non-relational databases. Non-relational databases may include blockchain-based distributed databases, etc., and are not limited thereto. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited thereto. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this application. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A method for fusion modeling of power electronic equipment mechanism data, characterized in that, The power electronic equipment includes at least a nonlinear system, and the nonlinear system includes multiple components; The modeling method includes: For each of the aforementioned components, based on the equivalent circuit of the component, multiple hardware components and multiple preset nodes in the equivalent circuit are obtained. The multiple preset nodes include multiple internal nodes and multiple external terminal nodes of the component. For each piece of hardware, based on the electrical parameters of the hardware, a functional relationship between the current flowing through the hardware and the node voltage of the preset nodes at both ends of the hardware is written. Based on Kirchhoff's current law, the functional relationships corresponding to the multiple hardware components are correlated to obtain the node voltage equation set of the component. Wherein, the hardware is a discrete electronic device, at least two of the multiple hardware components are of different types, and the node voltage equation set is a nonlinear differential-algebraic equation set. Solve the set of node voltage equations corresponding to the multiple components at the current time according to the preset conditions, and obtain the multiple internal node voltages and multiple external terminal node voltages of each component at the current time. Use the multiple internal node voltages and multiple external terminal node voltages of each component at multiple consecutive times as the values ​​of the nonlinear state variables of the nonlinear system at multiple consecutive times. Construct a neural network based on the Koopman operator as the first neural network; The first neural network is trained iteratively over a preset time period using the values ​​of the nonlinear state variables of the nonlinear system at multiple consecutive time points. The values ​​of the nonlinear state variables are time-domain numerical solutions; multiple different numerical solutions at the current time point constitute the numerical solution vector at the current time point; and the numerical solution vectors at multiple different time points constitute a dataset. This dataset is input into the first neural network for training. The multiple iterative training sessions within the preset time period are performed sequentially, and the loss value of the first neural network is determined after each training iteration. If the loss value of the first neural network is less than the preset target loss value after the last training within the preset time period, the mechanism data fusion model of the nonlinear system is obtained based on the first neural network that has completed multiple iterations of training within the preset time period.

2. The modeling method according to claim 1, characterized in that, Based on the value of the nonlinear state variable, the first neural network is trained iteratively multiple times within a preset time period. The multiple iterative training sessions within the preset time period are performed sequentially, and after each training iteration, a loss value for the first neural network is determined, including: Based on the value of the nonlinear state variable, the first neural network, which has completed the (i-1)th parameter update, is trained for the i-th time, where i is a positive integer and 1 < i; Based on a preset loss function and according to the value of the nonlinear state variable, the i-th loss value of the first neural network after completing the i-th training is determined; Based on the i-th loss value, control the first neural network, which has completed the i-th training, to perform the i-th parameter update; Based on the value of the nonlinear state variable, the first neural network, which has completed the i-th parameter update, is trained for the (i+1)th time.

3. The modeling method according to claim 2, characterized in that, The modeling method also includes: Based on the value of the nonlinear state variable, the first neural network is trained for the first time, wherein the parameters of the first neural network trained for the first time are preset initial values. Based on a preset loss function and according to the value of the nonlinear state variable, the first loss value of the first neural network after the first training is determined; Based on the first loss value, the first neural network, which has completed the first training, is controlled to perform the first parameter update.

4. The method according to claim 2, characterized in that, The preset time period includes the 1st to the Mth step sub-time periods, where M is a positive integer, 2≤M, at least two step sub-time periods are the same or different, the sum of all step sub-time periods equals the preset time period, each step sub-time period is configured with a preset expected loss value, the preset expected loss value of the jth step sub-time period is less than the preset expected loss value of the (j-1)th step sub-time period, where j is a positive integer, 1<j≤M, the preset expected loss value of the Mth step sub-time period equals the preset target loss value; The modeling method also includes: If the loss value of the first neural network is less than the preset expected loss value of the (j-1)th step sub-time period after the last training is completed, the first neural network that has completed multiple iterations of training in the (j-1)th step sub-time period will continue to undergo multiple iterations of training in the jth step sub-time period, wherein the multiple iterations of training in the step sub-time period are performed sequentially. If the loss value of the first neural network is greater than or equal to the preset expected loss value of the (j-1)th step sub-time period after the last training is completed, the training is terminated.

5. The method according to claim 2 or 3, characterized in that, The first neural network includes an encoder and a decoder, wherein the encoder is a Koopman feature function-based encoder, and the decoder corresponds to the encoder; The step of training the first neural network, which has completed the (i-1)th parameter update, based on the value of the nonlinear state variable, includes: The values ​​of the nonlinear state variables are mapped from the nonlinear space to the linear space by the encoder to obtain the values ​​of the linear state variables; Based on the general solution of the nodal voltage equations, and according to the Koopman eigenvalues ​​and the values ​​of the linear state variables, the predicted values ​​of the linear state variables are obtained by dynamically evolving in the linear space. The predicted values ​​of the linear state variables are mapped from the linear space back to the nonlinear space by the decoder.

6. The method according to claim 5, characterized in that, The first neural network also includes an auxiliary neural network; The modeling method also includes: Based on the values ​​of the linear state variables, the Koopman feature values ​​are determined using the auxiliary neural network.

7. The method according to claim 2 or 3, characterized in that, The step of determining the i-th loss value of the first neural network after completing the i-th training, based on a preset loss function and according to the value of the nonlinear state variable, includes: Based on the preset autoregressive error function and according to the value of the nonlinear state variable, the i-th autoregressive error of the first neural network after the i-th training is determined; Based on a preset prediction error function and according to the value of the nonlinear state variable, the i-th prediction error of the first neural network after completing the i-th training is determined; Based on a preset linear error function and according to the value of the nonlinear state variable, the i-th linear error of the first neural network after the i-th training is determined; Based on the i-th autoregressive error and its corresponding preset first weighting coefficient, the i-th prediction error and its corresponding preset second weighting coefficient, and the i-th linear error and its corresponding preset third weighting coefficient, the i-th loss value of the first neural network after completing the i-th training is determined.

8. A device for fusing and modeling the mechanism data of power electronic equipment, characterized in that, The power electronic equipment includes at least a nonlinear system, and the nonlinear system includes multiple components; The modeling device includes: A nonlinearization module is used to, for each of the aforementioned components, obtain multiple hardware components and multiple preset nodes in the equivalent circuit based on the equivalent circuit of the component, wherein the multiple preset nodes include multiple internal nodes and multiple external terminal nodes of the component; and for each of the aforementioned hardware components, based on the electrical parameters of the hardware, list the functional relationship between the current flowing through the hardware and the node voltage of the preset nodes at both ends of the hardware; and based on Kirchhoff's current law, associate the functional relationships corresponding to the multiple hardware components to obtain the node voltage equation set of the component; wherein the hardware is a discrete electronic device, at least two of the multiple hardware components are of different types, the node voltage equation set is a nonlinear differential-algebraic equation set, and solve the node voltage equation set corresponding to the multiple components at the current time according to preset conditions to obtain multiple internal node voltages and multiple external terminal node voltages of each component at the current time, and use the multiple internal node voltages and multiple external terminal node voltages of each component at multiple consecutive times as the values ​​of the nonlinear state variables of the nonlinear system at multiple consecutive times; The neural network module is used to construct a neural network based on the Koopman operator, which serves as the first neural network. The network training module is used to perform multiple iterative trainings on the first neural network within a preset time period using the values ​​of the nonlinear state variables of the nonlinear system at multiple consecutive time points. The values ​​of the nonlinear state variables are numerical solutions in the time domain, and multiple different numerical solutions at the current time constitute the numerical solution vector at the current time. These numerical solution vectors at multiple different time points constitute a dataset, which is then input into the first neural network for training. The multiple iterative trainings within the preset time period are performed sequentially, and the loss value of the first neural network is determined after each training iteration. The model building module is used to obtain the mechanism data fusion model of the nonlinear system based on the first neural network that has completed multiple iterations of training within the preset time period, provided that the loss value of the first neural network is less than the preset target loss value after the last training within the preset time period.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images