LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter

Abstract

The invention relates to the technical field of converters, in particular to an LLC perception type physical information nested neural network parameter estimation method suitable for an LLC resonant converter, and the method comprises the following steps: S1, constructing a continuous time state space model of the LLC resonant converter, and carrying out the discretization processing; s2, an LLC perception type physical information nested neural network is constructed, and the LLC perception type physical information nested neural network comprises a data reconstruction network and a physical information nested neural network which are connected and is used for online parameter identification of the LLC resonant converter; the data reconstruction network comprises a resonance state judgment layer, a K2 operation layer, a pseudo label generation layer, a constraint layer and a data reconstruction layer; the physical information nested neural network comprises a middle state mapping layer and a physical layer; defining LLC perception type physical information nested neural network input; s3, constructing a loss function; and S4, deploying and executing the LLC perception type physical information nested neural network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of converters, and more particularly, to an LLC perception type physical information nested neural network parameter estimation method suitable for an LLC resonant converter. BACKGROUND

[0002] LLC resonant converters are widely used in high-efficiency, high-power-density power conversion scenarios, such as server power supplies, chargers, and other industrial systems. Among many influencing factors, system parameters (such as resonant cavity parameters, transformer inductance, capacitance, equivalent series resistance ESR, load resistance, and power device on-resistance) play a dominant role in the working characteristics (resonant frequency, quality factor, and gain characteristics) of the LLC resonant converter. Therefore, accurate estimation of system parameters not only has important significance for realizing real-time component health monitoring and high-efficiency zero-voltage switching (ZVS), but also helps to deploy advanced control strategies and improve the operating performance of the entire system.

[0003] However, the industry lacks a practical solution for accurate identification of all parameters in the LLC resonant converter. The main obstacles are three key factors: the presence of multiple inductance and capacitance elements causes the resonant network to exhibit high nonlinearity; the use of switching frequency instead of duty cycle as a control variable causes additional nonlinear effects on system dynamic characteristics; and unobservable topology changes cause modeling uncertainty, which highly depends on high-frequency measurement methods.

[0004] Currently, some parameter estimation methods specifically designed for LLC converters have been proposed in the prior art, for example: one method is to derive the small-signal second-order discrete-time transfer function of the LLC converter, and identify the parameters through the least absolute shrinkage and selection operator (LASSO), but the complex nonlinear mapping between the coefficients and the physical parameters makes it difficult to extract the parameters; another method proposes a capacitance monitoring method based on capacitor model analysis of large-signal discharge trajectories, which can accurately estimate the output capacitance, but requires the converter to be turned off to capture the discharge curve, which is not suitable for online applications; another method derives an equivalent series resistance (ESR) estimation model for the output capacitance of the LLC converter, and realizes online ESR calculation through ripple measurement technology and peak current analysis, but the limitations of the diode peak current approximation method make it only applicable under high-power conditions. Another method introduces an active control strategy that can trigger millisecond-level discharge events, and accurately extracts the values of the resonant capacitance and output capacitance by using a pre-trained offline artificial neural network (ANN) to analyze the generated trajectories.

[0005] Although the above method can achieve high accuracy, the number of identified parameters is small, and the parameter prediction quantity is highly dependent, which significantly reduces the accuracy of the target parameters. In addition, most existing estimation techniques are only for capacitive elements, which has limited contribution to capturing the complete dynamic behavior of the converter, and lacks in implementing model-based control or digital twin development. Therefore, it is a technical problem to be solved to realize high-precision online estimation of all parameters of LLC resonant converter, reconstruct the nanosecond converter dynamic characteristics, and realize high-fidelity digital twin modeling. It also has great engineering and industrial value. SUMMARY

[0006] The present application aims to solve the problem of few identified parameters and high prediction quantity dependence of the existing LLC resonant converter parameter estimation method, and provides an LLC-aware physical information nested neural network parameter estimation method suitable for LLC resonant converter. This method can complete the online high-precision identification of all system parameters of LLC resonant converter.

[0007] An LLC-aware physical information nested neural network parameter estimation method suitable for LLC resonant converter, comprising the following steps:

[0008] S1. Dynamic modeling of LLC resonant converter, first constructing a continuous-time state space model of LLC resonant converter, and then discretizing the continuous-time state space model of LLC resonant converter;

[0009] S2. Constructing an LLC-aware physical information nested neural network, the LLC-aware physical information nested neural network comprising a data reconstruction network and a physical information nested neural network connected in series, for online parameter identification of LLC resonant converter;

[0010] The data reconstruction network comprises a resonant state judgment layer, a K2 operation layer, a pseudo-label generation layer, a constraint layer and a data reconstruction layer;

[0011] The physical information nested neural network comprises an intermediate state mapping layer and a physical layer, and the discretized LLC resonant converter continuous-time state space model is taken as the training model of the physical layer;

[0012] Defining the input of LLC-aware physical information nested neural network;

[0013] S3. Constructing a loss function;

[0014] S4. Deploying and executing the LLC-aware physical information nested neural network model.

[0015] Further, in step S1, the continuous-time state space model is represented as:

[0016]

[0017] Wherein, A is transformer ratio, V F2 is the secondary side rectifier diode conduction voltage drop; K is the upper bridge arm controllable switch MOSFET (S1) switch state signal, when S1 is turned on, K=1, when S1 is turned off, K=0; K2 is the secondary side rectifier diode D1, D2 conduction state signal, when any D1, D2 is turned on, K2=1, when D1, D2 is not turned on, K2=0; Lambda is the set of parameters to be identified, including resonant inductance L r , resonant capacitance C r , resonant resistance R r , transformer excitation inductance L m , transformer leakage inductance L2, leakage inductance parasitic resistance R2, output capacitance C, output capacitance parasitic resistance R C , load resistance R load .

[0018] Further, the training model of the physical layer based on the discretized LLC resonant converter continuous-time state space model comprises:

[0019]

[0020] Wherein, u represents the system state variable, corresponding to LLC resonant converter primary current i1, secondary current i2, output voltage v o and resonant capacitor voltage v r ;{a ij ,b j ,c j} values can be consulted in the I RK butcher table; each topology initial state time is k, final state time is k+1, and intermediate state time is k+c q , q represents the IRK framework order, referring to formula (1), f[u(k+c j ); Lambda represents the intermediate state function model of system variables.

[0021] Further, in step S2, an LLC sensing type physical information nested neural network is built on an embedded platform or a cloud platform.

[0022] The specific method for defining the input of LLC sensing type physical information nested neural network is:

[0023] Under the configuration of MCU regular sampling, LLC resonant converter will collect system state at each modulation wave and carrier intersection, including primary current i1, secondary current i2, output voltage v o and resonant capacitor voltage v r and switch state signal;

[0024] The initial state of the system state variable in the first switching state is represented as {i1(1), i2(1), v o (1), v r (1)}. The final state is represented as {i1(l+1), i2(l+1), v o (l+1), v r (l+1)}. The switching state signal K and the time interval Δt of the next switching state are added to form the ith data {i1(1), i2(1), v o (1), v r (1), K(1), Δt(1), i1(l+1), i2(l+1), v o (l+1), v r (l+1)}. These observable physical quantities are used as inputs of the neural network.

[0025] Further, in the resonance state judgment layer, F represents the resonance mode, F = 1 when the switching frequency (f s ) is greater than the resonance frequency (f r ), and F = -1 when f s < f r . According to whether the secondary side current i2(l+1) corresponding to the final state of one switching state is zero, the resonance state can be determined, i2(l+1) = 0, f s < f r , F = -1, otherwise F = 1, denoted as F(l);

[0026] In the K2 operation layer, K2 is represented as:

[0027] K2 = S t (-F·K) + (1-S t )(K(F+1) / 2) (10)

[0028] Where S t is a binary indicator of the topology phase, S t = 1 corresponds to the first topology phase of each switching period, and S t = 0 corresponds to the second topology phase of each switching period.

[0029] Further, in the pseudo-label generation layer, F(l) is combined with the input data into a feedforward neural network for inferring the system state and the timing conversion of the topology boundary in each switching interval; let (l+m) represent the topology conversion time between the (l)th and (l+1)th sampling points, then the variables to be inferred include {i1(l+m), i2(l+m), v o (l+m), v r (l+m), t(l+m)}; the calculation of each hidden layer in the feedforward neural network is as follows:

[0030]

[0031] where p represents the layer index, w ij (p-1) represents the weight connecting the i-th neuron of the (p-1)-th layer and the j-th neuron of the p-th layer, b j (p) is the bias of the j-th neuron of the p-th layer, I represents the number of neurons of the (p-1)-th layer, J represents the number of neurons of the p-th layer, and σ represents a nonlinear function, i.e., a Sigmoid function.

[0032] In the constraint layer, the constraint layer accepts the variables output by the pseudo-label generation layer, and substitutes them into the mathematical model based on the physical constraint layer:

[0033]

[0034] where, and u respectively represent the maximum and minimum values of all sampling data of the variable u; the output is used as the pseudo-label {i1(l+m), i2(l+m), v o (l+m), v r (l+m), t(l+m)} of the downstream physical information nested neural network.

[0035] In the data reconstruction layer, the data reconstruction layer re-formats each original data into two pieces of data, which correspond to two topological stages within the exchange interval, forming the input structure required by the downstream physical information nested neural network;

[0036] Let the index of the reconstructed data be n, the (l)-th original data {i1(l), i2(l), v r (l), v o (l), i1(l+1), i2(l+1), v r (l+1), v o (l+1), Δt(l), K(l), F(l)} is divided into two pieces of reconstructed data:

[0037] The n-th data: {i1(n):=i1(l), i2(n):=i2(l), v r (n):=v r (l), v o (n):=v o (l), v o (n+1):=i1(l+m), i2(n+1):=i2(l+m), v r (n+1):=v r (l+m), v o (n+1):=v o(l+m), K(n): = K(l), At(n): = t(l+m), K2(n): = -1*K(l)*F(l)}

[0038] The nth+1 data: {i1(n+1): = i1(l+m), i2(n+1): = i2(l+m), v r (n+1): = v r (l+m), v o (n+1): = v o (l+m), i1(n+2): = i1(l+1), i2(n+2): = i2(l+1), v r (n+2): = v r (l+1), v o (n+2): = v o (l+1), K(n+1): = K(l), At(n+1): = At(l)-t(l+m), K2(n+1): = -1*K(l)*F(l)}.

[0039] Further, in the intermediate state mapping layer, the intermediate state mapping layer is implemented by a feedforward neural network, and the mathematical expression thereof is completely consistent with formula (11); the layer maps the reconstructed input vector to the potential intermediate state set in the corresponding topological interval;

[0040] In the physical layer, the physical layer accepts the intermediate state output by the intermediate state mapping layer, substitutes it into the LLC resonant converter IRK model in formulas (2)-(8), and outputs the initial and final states estimated for each switching process and The weight parameter of the physical layer is the system parameter λ to be identified.

[0041] Further, the loss function Loss is defined as the sum of the residuals of all outputs of the physical layer:

[0042]

[0043] where N represents the total number of reconstructed data.

[0044] Further, in step S4, the LLC-aware physical information nested neural network model is deployed and executed, specifically:

[0045] The constructed LLC-aware physical information nested neural network model is deployed to an MCU platform or a cloud platform, and the LLC-aware physical information nested neural network model includes periodic or event-triggered data acquisition, data processing, model training and parameter extraction processes.

[0046] Further, the specific method of data acquisition is: acquiring system state at each switching state switching point, including primary side current i1, secondary side current i2, output voltage v o , resonance capacitor voltage v r , time interval of each switching Δt and switching state K;

[0047] The specific method of data processing is: converting the collected data into training data samples suitable for LLC perception type physical information nested neural network model, and each training data sample contains i1(l), i2(l), v o (l), v r (l), K(l), Δt(l), i1(l+1), i2(l+1), v o (l+1), v r (l+1) ten elements;

[0048] The specific method of model training is: inputting the training data sample into the LLC perception type physical information nested neural network, running the LLC perception type physical information nested neural network back propagation algorithm, and gradually approaching the true parameter value of the weight parameter λ in the physical layer as the loss function converges;

[0049] The specific method of extracting parameters is: when the training residual error is reduced to a certain threshold and tends to be stable, the training process is stopped, and the estimated result of the system parameter λ is finally extracted from the physical layer.

[0050] Compared with the prior art, the beneficial effects of the present application are:

[0051] The LLC perception type physical information nested neural network parameter estimation method suitable for the LLC resonant converter in the present application can realize full parameter identification and get rid of the dependence on parameter prediction. The present application directly embeds the discrete time topology related dynamic characteristics of the LLC resonant converter into the neural network, innovatively introduces a semi-supervised data reconstruction mechanism, can infer the system state at the unobservable topology switching point, so as to ensure that the continuous and effective supervised learning is realized in the dynamic discontinuous area. The present application overcomes the problems of strong parameter coupling, high system nonlinearity and unobservable topology conversion of the LLC resonant converter, so that only relying on the conventional sampling configuration (collecting i1, i2, v o and v r once at each switching time) can complete the identification of multiple system parameters (such as resonant inductance L r , resonant capacitance C r , resonant resistance R r , transformer excitation inductance L m , transformer leakage inductance L2, leakage inductance parasitic resistance R2, output capacitance C, output capacitance parasitic resistance R C , load resistance R loadonline identification of the LLC resonant converter.

[0052] The physical constraints of the LLC-aware physical information nested neural network in the application contain all dynamic information of the system, so that calibration or feature extraction through external large signal interference is not required, and the system complexity and operation cost are reduced. At the same time, the training of the LLC-aware physical information nested neural network is a continuous optimization process, so that parameter estimation and state awareness are always kept in the optimal state, and therefore the existing system can be directly embedded without additional development or configuration, realizing seamless integration and rapid deployment.

[0053] The LLC resonant converter continuous-time dynamic model is discretized by using the implicit Runge-Kutta (IRK) method in the application, so as to convert the continuous-time dynamics into algebraic relationships between discrete-time point variables. The existence of the data reconstruction network ensures the prediction of unobservable points, and finally the parameters are accurately identified while only using the system state sampled at the switching frequency to reconstruct the nanosecond-level converter dynamics. BRIEF DESCRIPTION OF DRAWINGS

[0054] The accompanying drawings are included to provide a further understanding of the application and constitute a part of the specification, which together with the embodiments of the application are used to explain the application and do not constitute a limitation on the application. In the drawings:

[0055] Figure 1 is a flowchart of the LLC-aware physical information nested neural network parameter estimation method for the LLC resonant converter in the application.

[0056] Figure 2 is a schematic diagram of the LLC resonant converter in the application.

[0057] Figure 3 is a structural schematic diagram of the LLC-aware physical information nested neural network in the application.

[0058] Figure 4 is a data acquisition schematic diagram in the over-resonance state.

[0059] Figure 5 is a data acquisition schematic diagram in the under-resonance state. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0061] Embodiment:

[0062] As Figure 1 shown, the LLC resonance converter-aware physical information nested neural network parameter estimation method suitable for the LLC resonance converter in the embodiment includes the following steps:

[0063] S1. Dynamic modeling of the LLC resonance converter.

[0064] S11. Constructing a continuous-time state space model of the LLC resonance converter.

[0065] The main parameters of the LLC resonance converter are as Figure 2 shown, a continuous-time state space model is constructed and represented as:

[0066]

[0067] where A is the transformer turns ratio; V F2 is the on-state voltage drop of the secondary rectifier diode; K is the on-state signal of the upper bridge arm controllable switch MOSFET (S1), K = 1 when S1 is on, and K = 0 when S1 is off; K2 is the on-state signal of the secondary rectifier diodes D1 and D2, K2 = 1 when either D1 or D2 is on, and K2 = 0 when neither D1 nor D2 is on. λ is the set of parameters to be identified, including the resonant inductance L r , the resonant capacitance C r , the resonant resistance R r , the transformer excitation inductance L m , the transformer leakage inductance L2, the leakage inductance parasitic resistance R2, the output capacitance C, the output capacitance parasitic resistance R C , and the load resistance R load , etc.

[0068] S12. Discretization processing of the continuous-time state space model of the LLC resonance converter,

[0069] The LLC resonance converter continuous-time state space model in step S11 is discretized using the implicit Runge-Kutta (IRK) method to obtain the system state evolution relationship within each switching state, thereby constructing a physical layer model that can be used for neural network training and accelerating the parameter identification efficiency.

[0070]

[0071] where u represents the system state variable (corresponding to the LLC resonance converter primary current i1, secondary current i2, output voltage v o , and resonant capacitor voltage v r ), {a ij ,b j ,c jThe values of I RK can be found in the I RK table. Each topology has an initial state at time k, an intermediate state at time k+c q , and a final state at time k+1. q represents the I RK frame order, referring to formula (1), f[u(k+c j ); λ] represents the intermediate state function model of the system variable.

[0072] S2. Constructing LLC-aware physical information nested neural network.

[0073] On an embedded platform such as DSP, ARM or cloud platform, an LLC-aware physical information nested neural network LLC-Aware PINN algorithm framework is built, as shown in Figure 3 , which includes two connected networks: a data reconstruction network DRN and a physical information nested neural network PINN, for online parameter identification of the LLC resonant converter. The data reconstruction network includes a resonant state judgment layer, a K2 operation layer pseudo label generation layer, a constraint layer and a data reconstruction layer. The physical information nested neural network includes an intermediate state mapping layer and a physical layer, which takes the discretized LLC resonant converter continuous-time state space model as the training model of the physical layer.

[0074] S21. Defining LLC-aware physical information nested neural network input.

[0075] Figure 4 Figure 5 are the over-resonance state and the under-resonance state, respectively. Under the configuration of the MCU regular sampling, the LLC resonant converter will collect the system state at each modulation wave and carrier intersection, including the primary current i1, the secondary current i2, the output voltage v o , the resonant capacitor voltage v r , and the switching state signal. For example, the initial state of the system state variable under the lth switching state is represented as {i1(l), i2(l), v o (l), v r (l)}, and the final state is represented as {i1(l+1), i2(l+1), v o (l+1), v r (l+1)}. Adding the switching state signal K and the time interval Δt of the next switching state, the lth data {i1(l), i2(l), v o (l), v r (l), K(l), Δt(l), i1(l+1), i2(l+1), v o (l+1), v r (l+1)} is formed. These observable physical quantities will be used as the input of the neural network.

[0076] S22. Judging the resonant state in the resonant state judgment layer.

[0077] F represents the resonance mode, when the switching frequency (f s ) is greater than the resonance frequency (f r ), F = 1, when f s < f r , F = -1. According to whether the secondary side current i2(l+1) corresponding to the final state of one switching state is zero, the resonance state can be determined, when i2(l+1) = 0, f s < f r , F = -1, otherwise F = 1, denoted as F(l).

[0078] S23. Operating K2 at the K2 operation layer.

[0079] K2 cannot be directly observed, but can be expressed as:

[0080] K2 = S t (-F·K) + (1-S t )(K(F+1) / 2) (10)

[0081] where S t is the binary indicator of the topology phase, S t = 1 corresponds to the first topology phase of each switching period, and S t = 0 corresponds to the second topology phase of each switching period.

[0082] S24. Generating pseudo labels at the pseudo label generation layer.

[0083] F(l) is combined with the input data into a feedforward neural network for inferring the system state and the timing transition of the topology boundary in each switching interval. Let (l+m) represent the topology transition time between the (l)th and (l+1)th sampling points, then the variables to be inferred include {i1(l+m), i2(l+m), v o (l+m), v r (l+m), t(l+m)}. The calculation of each hidden layer in the feedforward neural network is as follows:

[0084]

[0085] where p represents the layer index, w ij (p-1) represents the weight connecting the i th neuron of the (p-1)th layer and the j th neuron of the p th layer, b j (p) is the bias of the j th neuron of the p th layer, I represents the number of neurons of the (p-1)th layer, J represents the number of neurons of the p th layer, and σ represents a nonlinear function, i.e., a Sigmoid function.

[0086] S25. The constraint layer accepts the variables output by the pseudo label generation layer and substitutes them into the mathematical model based on the physical constraint layer:

[0087]

[0088] wherein, and u respectively represent the maximum and minimum values of all sampling data of the variable u; the output is used as the pseudo label {i1(l+m), i2(l+m), v o (l+m), v r (l+m), t(l+m)} of the downstream physical information nested neural network.

[0089] S26. The data reconstruction layer re-formats each original data into two data corresponding to two topological stages within the exchange interval, forming the input structure required by the downstream physical information nested neural network;

[0090] Let the reconstructed data index be n, the (l)th original data {i1(l), i2(l), v r (l), v o (l+1), i2(l+1), v r (l+1), v o (l+1), Δt(l), K(l), F(l)} is divided into two reconstructed data:

[0091] The n th data: {i1(n): = i1(l), i2(n): = i2(l), v r (n): = v r (l), v o (n): = v o (l), v o (n+1): = i1(l+m), i2(n+1): = i2(l+m), v r (n+1): = v r (l+m), v o (n+1): = v o (l+m), K(n): = K(l), Δt(n): = t(l+m), K2(n): = -1*K(l)*F(l)}

[0092] The n+1 th data: {i1(n+1): = i1(l+m), i2(n+1): = i2(l+m), v r (n+1): = v r (l+m), v o (n+1): = v o (l+1), i2(n+2): = i2(l+1), v r (n+2): = v r(l+1),v o (n+2): = v o (l+1),K(n+1): = K(l), At(n+1): = At(l) - t(l+m), K2(n+1): = -1 * K(l) * F(l)}.

[0093] S27. The intermediate state mapping layer is implemented using a feedforward neural network, and its mathematical expression is completely consistent with formula (11); this layer maps the reconstructed input vector to the set of potential intermediate states within the corresponding topological interval.

[0094] S28. The physical layer accepts the intermediate state output by the intermediate state mapping layer, substitutes it into the LLC resonant converter IRK model in formulas (2)-(8), and outputs the initial and final states estimated for each switching process and The weight parameter of the physical layer is the system parameter λ to be identified.

[0095] S3. Construct the loss function.

[0096] The loss function Loss is defined as the sum of the residuals of all outputs of the physical layer:

[0097]

[0098] where N represents the total number of reconstructed data.

[0099] S4. Deploy and execute the LLC-aware physical information nested neural network model.

[0100] Deploy the constructed LLC-aware physical information nested neural network model to an MCU platform or a cloud platform, and execute the LLC-aware physical information nested neural network model, including periodically or event-triggered data collection, data processing, model training, and parameter extraction processes.

[0101] The specific method of data collection is to collect system states at each switching state switching point, including primary side current i1, secondary side current i2, output voltage v r , resonant capacitor voltage v o , time interval Δt of each switching, and switching state K.

[0102] The specific method of data processing is to convert the collected data into training data samples suitable for the LLC-aware physical information nested neural network model, and each training data sample contains i1(l), i2(l), v r (l), v o (l), K(l), At(l), i1(l+1), i2(l+1), v r (l+1), vr (l+1) ten elements.

[0103] The specific method of model training is: inputting the training data sample into the LLC perception type physical information nested neural network, running the LLC perception type physical information nested neural network back propagation algorithm, and gradually approaching the true parameter value of the weight parameter λ in the physical layer as the loss function converges.

[0104] The specific method of extracting the parameter is: when the training residual error is reduced to a certain threshold and tends to be stable, the training process is stopped, and the estimation result of the system parameter λ is finally extracted from the physical layer.

[0105] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the foregoing embodiments of the present application are described in detail, for those skilled in the art, it still can be modified, or part of the technical features of the equivalent replacement of the technical solutions recorded in the foregoing embodiments. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An LLC-aware type physical information embedding neural network parameter estimation method suitable for an LLC resonant converter, characterized in that, The method comprises the following steps: S1. Dynamic modeling of LLC resonant converter, first constructing a continuous-time state space model of LLC resonant converter, and then discretizing the continuous-time state space model of LLC resonant converter; S2. Constructing LLC perception type physical information nested neural network, the LLC perception type physical information nested neural network comprising a data reconstruction network and a physical information nested neural network connected in series, and being used for online parameter identification of LLC resonant converter; The data reconstruction network comprises a resonant state judgment layer, a K2 operation layer, a pseudo-label generation layer, a constraint layer and a data reconstruction layer; The physical information nested neural network comprises an intermediate state mapping layer and a physical layer, and the discretized LLC resonant converter continuous-time state space model is taken as a training model of the physical layer; Defining an LLC perception type physical information nested neural network input; S3. Constructing a loss function; S4. Deploying and executing the LLC perception type physical information nested neural network model.

2. The LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 1, wherein in step S1, the continuous-time state space model is represented as:

3. The LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 2, wherein the training model of the physical layer based on the discretized LLC resonant converter continuous-time state space model comprises: Wherein, A is transformer ratio, V F2 is the secondary side rectifier diode conduction voltage drop; K is the upper bridge arm controllable switch MOSFET (S1) switch state signal, when S1 is turned on, K=1, when S1 is turned off, K=0; K2 is the secondary side rectifier diode D1, D2 conduction state signal, when any D1, D2 is turned on, K2=1, when D1, D2 are not turned on, K2=0; λ is the to-be-identified parameter set, including resonant inductance L r , resonant capacitance C r , resonant resistance R r , transformer excitation inductance L m , transformer leakage inductance L2, leakage inductance parasitic resistance R2, output capacitance C, output capacitance parasitic resistance R C , load resistance R load .

4. The LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 3, wherein in step S2, the LLC perception type physical information nested neural network is built on an embedded platform or a cloud platform; The specific method for defining the LLC perception type physical information nested neural network input is: wherein u represents the system state variable, corresponding to the LLC resonant converter primary side current i1, secondary side current i2, output voltage v o and resonant capacitor voltage v r ;{a ij ,b j ,c j} values can be consulted in the I RK butcher table; each topology initial state time is k, final state time is k+1, intermediate state time is k+c q , q represents the IRK framework order, referring to formula (1), f[u(k+c j ); λ] represents the system variable intermediate state function model.

5. The LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 4, wherein in the K2 operation layer, K2 is represented as:

6. The LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 5, wherein in the constraint layer, the constraint layer accepts variables output by the pseudo-label generation layer, and substitutes the variables into a mathematical model based on the physical constraint layer: In the data reconstruction layer, the data reconstruction layer re-formats each original data into two pieces of data, the two pieces of data corresponding to two topology stages in an exchange interval, and forms an input structure required by a downstream physical information nested neural network; Under the MCU regular sampling configuration, LLC resonant converter will collect system state at each modulation wave and carrier intersection, including the primary current i1, secondary current i2, output voltage v o and resonant capacitor voltage v r and switch state signal; The initial state of the system state variable in the first switching state is represented as {il(l), i2(l), vl(l)} o (l), vl(l) r (l)} and the final state is represented as {il(l+1), i2(l+1), vl(l+1)} o (l+1), vl(l+1) r (l+1)} The switching state signal K and the time interval Δt of the next switching state are added to form the ith data {il(l), i2(l), vl(l)} o (l), vl(l) r (l), K(l), Δt(l), il(l+1), i2(l+1), vl(l+1) o (l+1), vl(l+1) r (l+1)} These observable physical quantities are used as the input of the neural network.

7. The LLC perception type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 6, wherein in the intermediate state mapping layer, the intermediate state mapping layer is implemented by using a feedforward neural network, and a mathematical expression thereof is completely consistent with formula (11); the layer maps the reconstructed input vector to a potential intermediate state set in a corresponding topology interval; In the resonance state judgment layer, F represents the resonance mode, when the switching frequency (f s >Resonant frequency (f r When f = 1, F = 1. s <f r At that time, F = -1. The resonant state can be determined by whether the secondary current i2(l+1) corresponding to the final state of a switch is zero. When i2(l+1) = 0, F... s <f r If F = -1, otherwise F = 1, denoted as F(l); The loss function Loss is defined as the sum of residuals of all outputs of the physical layer: K2 = S t (-F · K) + (1 - S t )(K(F + 1) / 2) (10) where S t is a binary indicator of the topology phase, S t = 1 corresponds to the first topology phase of each switching cycle, S t = 0 corresponds to the second topology phase of each switching cycle. ​ In the pseudo-label generation layer, F(l) is combined with the input data into a feedforward neural network for inferring the system state and the timing transition of the topology boundary within each switching interval; let (l+m) represent the topology transition time between the (l)th and (l+1)th sampling points, then the variables to be inferred include {i1(l+m), i2(l+m), v o (l+m), v r (l+m), t(l+m)}; the calculation of each hidden layer in the feedforward neural network is as follows: where p represents a layer index, w ij (p-1) represents a weight connecting the (p-1)th neuron of the (p-1)th layer and the jth neuron of the pth layer, b j (p) is a bias of the jth neuron of the pth layer, I represents the number of neurons of the (p-1)th layer, J represents the number of neurons of the pth layer, and σ represents a nonlinear function, i.e., a Sigmoid function. ​ wherein, and u respectively represent the maximum and minimum values of all sampled data of the variable u; the output is used as pseudo labels {i1(l+m), i2(l+m), v o (l+m), v r (l+m), t(l+m)} for the downstream physical information nested neural network. ​ Let the reconstructed data index be n, and let the (l)th original data be {i1(l),i2(l),v}. r (l),v o (l),i1(l+1),i2(l+1),v r (l+1),v o (l+1),Δt(l),K(l),F(l)} is divided into two reconstructed data: • nth data: {il(n): = il(l), i2(n): = i2(l), v r (n): = v r (l), v o (n): = v o (l), v o (n+1): = il(l+m), i2(n+1): = i2(l+m), v r (n+1): = v r (l+m), v o (n+1): = v o (l+m), K(n): = K(l), At(n): = t(l+m), K2(n): = -1*K(l)*F(l) • nth+1 data: {il(n+1): = il(l+m), i2(n+1): = i2(l+m), v r (n+1): = v r (l+m), v o (n+1): = v o (l+m), il(n+2): = il(l+1), i2(n+2): = i2(l+1), v r (n+2): = v r (l+1), v o (n+2): = v o (l+1), K(n+1): = K(l), At(n+1): = At(l) - t(l+m), K2(n+1): = -1 * K(l) * F(l)}. ​ ​ At the physical layer, the physical layer accepts the intermediate state outputted by the intermediate state mapping layer, substitutes into the LLC resonant converter IRK model in formulas (2)-(8), and outputs the initial and final states estimated for each switching process and The weight parameter of the physical layer is the system parameter λ to be identified.

8. The LLC-aware type physical information nested neural network parameter estimation method suitable for LLC resonant converter according to claim 7, characterized in that, ​ Wherein N represents the total number of reconstructed data.

9. The LLC-aware type physical information-embedded neural network parameter estimation method for LLC resonant converter according to claim 8, wherein, In step S4, the LLC-aware physical information nested neural network model is deployed and executed. The constructed LLC-aware physical information nested neural network model is deployed to an MCU platform or a cloud platform, and the LLC-aware physical information nested neural network model is executed, including periodically or event-triggered data collection, data processing, model training and parameter extraction processes.

10. The LLC-aware physical information nested neural network parameter estimation method suitable for an LLC resonant converter according to claim 9, characterized in that, The specific method of data collection is: collecting system state at each switching state switching point, including primary side current i1, secondary side current i2, output voltage v o , resonance capacitor voltage v r , time interval of each switching Δt and switching state K; The specific method of data processing is: converting the collected data into training data samples suitable for the LLC perception type physical information nested neural network model, each training data sample contains i1(l), i2(l), v o (l), v r (l), K(l), Δt(l), i1(l+1), i2(l+1), v o (l+1), v r (l+1) ten elements; The specific method of model training is: inputting the training data sample into the LLC-aware physical information nested neural network, running the LLC-aware physical information nested neural network back propagation algorithm, and gradually approaching the true parameter value of the weight parameter λ in the physical layer as the loss function converges; The specific method of parameter extraction is: when the training residual error is reduced to a certain threshold and tends to be stable, the training process is stopped, and the estimated result of the system parameter λ is finally extracted from the physical layer.

Images