An LLC-aware type physical information embedding neural network parameter estimation method suitable for an LLC resonant converter

Abstract

This invention relates to the field of converter technology, and particularly to a parameter estimation method for LLC-aware physical information nested neural network suitable for LLC resonant converters, comprising the following steps: S1. Constructing a continuous-time state-space model of the LLC resonant converter and discretizing it; S2. Constructing an LLC-aware physical information nested neural network, wherein the LLC-aware physical information nested neural network includes a connected data reconstruction network and a physical information nested neural network for online parameter identification of the LLC resonant converter; the data reconstruction network includes 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 includes an intermediate state mapping layer and a physical layer; defining the input of the LLC-aware physical information nested neural network; S3. Constructing a loss function; S4. Deploying and executing the LLC-aware physical information nested neural network model.

Description

Technical Field

[0001] This invention relates to the field of converter technology, and more specifically, to a method for estimating LLC-sensory physical information nested neural network parameters suitable for LLC resonant converters. Background Technology

[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 the 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 operating characteristics (resonant frequency, quality factor, and gain characteristics) of LLC resonant converters. Therefore, accurately estimating system parameters is not only crucial for achieving real-time component health monitoring and efficient zero-voltage switching (ZVS), but also helps in deploying advanced control strategies and improving the overall system performance.

[0003] However, the industry lacks a practical solution for the accurate identification of all parameters in LLC resonant converters. The main obstacles stem from three key factors: the presence of multiple inductors and capacitors leads to a highly nonlinear resonant network; the use of switching frequency instead of duty cycle as the control variable introduces additional nonlinear effects on the system's dynamic characteristics; and unobservable topology changes result in uncertainties in modeling, thus making high-frequency measurement methods highly dependent on them.

[0004] Currently, existing technologies have proposed several parameter estimation methods specifically designed for LLC converters. For example, one method derives the small-signal second-order discrete-time transfer function of the LLC converter and identifies parameters using the Least Absolute Shrunk and Selection Operator (LASSO). However, the complex nonlinear mapping between coefficients and physical parameters makes parameter extraction difficult. Another method proposes a capacitance monitoring method based on capacitance model analysis of large-signal discharge trajectories, which can accurately estimate the output capacitance, but requires shutting down the converter to capture the discharge curve, making it unsuitable for online applications. Yet another method derives an equivalent series resistance (ESR) estimation model for the LLC converter's output capacitance and achieves online ESR calculation through ripple measurement technology and peak current analysis. However, the limitations of the diode peak current approximation method limit its applicability to high-power conditions. A third method introduces an active control strategy that can trigger millisecond-level discharge events, accurately extracting the values ​​of the resonant capacitance and output capacitance by analyzing the generated trajectory using a pre-trained offline artificial neural network (ANN).

[0005] While the aforementioned methods achieve high accuracy, they suffer from a limited number of identified parameters and high dependence on the predicted parameters, significantly reducing the accuracy of the target parameters. Furthermore, most existing estimation techniques only target capacitive elements, offering limited contribution to capturing the complete dynamic behavior of the converter and falling short in achieving model-based control or digital twin development. Therefore, achieving high-precision online estimation of all parameters of an LLC resonant converter, while simultaneously reconstructing the dynamic characteristics of the nanosecond-level converter and realizing high-fidelity digital twin modeling, is a pressing technical challenge with significant engineering and industrial value. Summary of the Invention

[0006] This invention addresses the problems of existing LLC resonant converter parameter estimation methods having a small number of identified parameters and high predictive dependence. It aims to provide an LLC-sensory physical information nested neural network parameter estimation method suitable for LLC resonant converters. This method can achieve online high-precision identification of all system parameters of the LLC resonant converter.

[0007] A parameter estimation method for LLC-sensory physical information nested neural network for LLC resonant converters includes the following steps:

[0008] S1. Dynamic modeling of LLC resonant converter: First, a continuous-time state-space model of LLC resonant converter is constructed, and then the continuous-time state-space model of LLC resonant converter is discretized.

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

[0010] The data reconstruction network includes a resonance 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 includes an intermediate state mapping layer and a physical layer, and uses the discretized LLC resonant converter continuous-time state-space model as the training model of the physical layer.

[0012] Define the input of the LLC-sensory physical information nested into a neural network;

[0013] S3. Construct the loss function;

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

[0015] Furthermore, in step S1, the constructed continuous-time state-space model is represented as follows: Where A is the transformer turns ratio. i 1 The primary current, i 2 For secondary current, v o For output voltage, v r V is the voltage across the resonant capacitor. F2 The forward voltage drop of the secondary rectifier diode; K For the upper bridge arm controllable switching transistor MOSFET ( S 1 Switch status signal, when S 1 When conducting, K =1, when S 1 When shut down, K =0; K 2 Secondary rectifier diode D 1 、D 2 On status signal, when any D 1 、D 2 When conducting ，K 2 =1, when D 1 , D 2 When neither is conducting, K 2 =0; λ is the set of parameters to be identified, including the resonant inductance. L r Resonant capacitor C r Resonant resistor R r Transformer magnetizing inductance L m Transformer leakage inductance L 2 Leakage inductance and parasitic resistance R 2 Output capacitor C Output capacitor parasitic resistance R C Load resistance R load .

[0016] Furthermore, the training model for the physical layer, constructed based on the discretized continuous-time state-space model of the LLC resonant converter, specifically includes: in, uThis represents the system state variable, corresponding to the primary current of the LLC resonant converter. i 1 Secondary current i 2 Output voltage v o and resonant capacitor voltage v r ； { a ij , b j , c j The value of} can be found in the IRK butcher table; the initial state time for each topology is . k The final state is k+1 The intermediate state time is k+c i , q To indicate the order of the IRK frame, refer to formula (1). f [ u ( k+c j ); λ ] represents the intermediate state function model of the system variables.

[0017] Furthermore, in step S2, an LLC-aware physical information nested neural network is built on an embedded platform or cloud platform;

[0018] The specific method for defining the input of the LLC-sensory physical information nested neural network is as follows:

[0019] Under the standard sampling configuration of the MCU, the LLC resonant converter will collect the system state, including the primary current, each time the modulated wave intersects with the carrier wave. i 1 Secondary current i 2 Output voltage v o and resonant capacitor voltage v r and switch status signal; l The initial state of the system state variables under each switching state is represented as { i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l The final state is represented as { )}. i 1 (l+ 1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 Add switch status signal K The time interval Δ between the next switching state and the next switching state t , forming the first l Data items { i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l ), K ( l ), Δ t ( l ), i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+ 1 ), v r ( l+1 These observable physical quantities serve as inputs to the neural network. Furthermore, in the resonant state determination layer... F Indicates the resonant mode, when the switching frequency ( f s >Resonant frequency ( f r )hour, F =1, when f s < f r hour, F =-1. Based on the secondary current corresponding to the final state of a switch. i 2 ( l+1 Whether the value is zero can determine the resonance state. i 2 ( l+1 )=0, f s < fr , F =-1, otherwise F =1, denoted as F ( l );

[0020] exist K 2 Computation layer K 2 Represented as: in, S t It is a binary indicator for the topology stage. 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.

[0021] Furthermore, in the pseudo-tag generation layer, F ( l The data is combined with the input data and fed into a feedforward neural network to infer the system state and the timing transition of the topology boundary within each switching interval; let ( l+m ) indicates the ( l ) and the ( l+1 At the topological transformation time between ) sampling points, the variables to be inferred include { i 1 ( l+m ), i 2 ( 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: in p Representation layer index, w ij ( p-1 ) indicates the connection of the () p-1 ) layer i The first neuron and the second p Layer j The weights of each neuron, b j ( p ) is the first p Layer j Bias of each neuron I Indicates the ( p-1The number of neurons in the layer J Indicates the first p The number of neurons in the layer σ This represents a non-linear function, namely the Sigmoid function.

[0022] In the constraint layer, the variables output by the pseudo-label generation layer are substituted into the mathematical model based on the physical constraint layer: in, and Representing variables respectively u The maximum and minimum values ​​of all sampled data; the output is used as pseudo-labels for the downstream physical information nested neural network. i 1 ( l+m) , i 2 ( l+m ), v o ( l+m ), v r ( l+m ), t ( l+m )};

[0023] In the data reconstruction layer, each original data item is reformatted into two data items, corresponding to two topological stages within the exchange interval, forming the input structure required by the downstream physical information nested neural network; the index of the reconstructed data is denoted as... n , No. ( l ) original data { i 1 ( l ) , i 2 ( l ) , v r ( l ) , v o ( l ) , i 1 ( l+1 ) , i 2 ( l+ 1 ) , v r ( l+1 ) , v o ( l+1), ∆t ( l ) , K ( l ) , F (l The data is divided into two reconstructed data sets:

[0024] No. n Data item:{ i 1 ( n ):= i 1 ( l ), i 2 ( n ):= i 2 ( l ), v r ( n ):= v r ( l ), v o ( n ):= v o ( l ), v o ( n+ 1 ):= i 1 ( l+m ), i 2 ( n+1 ):= i 2 ( l+m ), v r ( n+1 ):= v r ( l+m ), v o ( n+1 ):= v o ( l+m ), K ( n ):= K ( l ),∆ t ( n ):= t ( l+m ), K 2 ( n ):=-1* K ( l )* F ( l )}

[0025] No. n +1 data entry: { i 1 ( n+1 ):= i 1 ( l+m ), i 2 ( n+1 ):= i 2 ( l+m ), v r ( n+1 ):= v r ( l+m ), v o ( n +1 ):= v o ( l+m ), i 1 ( n+2 ):= i 1 ( l+1 ), i 2 ( n+2 ):= i 2 ( l+1 ), v r ( n+2 ):= v r ( l+1 ), v o ( n+2 ):= v o ( l +1 ), K ( n+1 ):= K ( l ), ∆ t ( n+1 ):=∆ t ( l )- t ( l+m ), K 2 ( n+1 ):= -1* K ( l )* F (l )}.

[0026] Furthermore, in the intermediate state mapping layer, a feedforward neural network is used, and its mathematical expression is completely consistent with formula (11); this layer maps the reconstructed input vector to the set of potential intermediate states in the corresponding topological interval.

[0027] At the physical layer, the physical layer receives the intermediate states output by the intermediate state mapping layer, substitutes them into the LLC resonant converter IRK model in formulas (2)-(8), and outputs the estimated initial and final states for each switching process. and The weight parameters of the physical layer are the system parameters to be identified. λ Furthermore, the loss function Loss Defined as the sum of the residuals of all outputs from the physical layer:

[0028] in N This indicates the total number of reconstructed data.

[0029] Furthermore, in step S4, the deployment and execution of the LLC-aware physical information nested neural network model specifically involves:

[0030] The constructed LLC-aware physical information nested neural network model is deployed to an MCU platform or cloud platform. Executing the LLC-aware physical information nested neural network model includes periodically or event-triggered execution of data acquisition, data processing, model training, and parameter extraction processes.

[0031] Furthermore, the specific method for data acquisition is as follows: The system status, including the primary current, is acquired at each switching point. i 1 Secondary current i 2 Output voltage v o Resonant capacitor voltage v r The time interval Δ between each switch t and switch status K；

[0032] The specific data processing method is as follows: the collected data is transformed into training data samples adapted to the LLC perceptual physical information nested neural network model. Each training data sample contains... i 1 ( l ), i 2 ( l ), v o ( l ),v r ( l ), K ( l ),Δ t ( l ), i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 Ten elements;

[0033] The specific method for model training is as follows: Input the training data samples into the LLC perceptual physical information nested neural network, run the LLC perceptual physical information nested neural network backpropagation algorithm, and as the loss function converges, the weight parameters in the physical layer... λ Gradually approaching the true parameter value;

[0034] The specific method for extracting parameters is as follows: when the training residual decreases to a certain threshold and tends to stabilize, the training process is stopped, and the system parameters are... λ The estimation results are ultimately extracted from the physical layer.

[0035] Compared with existing technologies, the beneficial effects of this invention are as follows: The LLC-sensory physical information nested neural network parameter estimation method for LLC resonant converters in this invention can achieve full parameter identification, eliminating the dependence on parameter prediction quantities. This invention directly embeds the discrete-time topology-dependent dynamic characteristics of the LLC resonant converter into a neural network, innovatively introducing a semi-supervised data reconstruction mechanism. This enables the inference of system state at unobservable topology switching points, thereby ensuring continuous and effective supervised learning in dynamic discontinuous regions. This invention overcomes the problems of strong parameter coupling, high system nonlinearity, and unobservable topology transitions in LLC resonant converters, enabling the estimation to rely solely on conventional sampling configurations (collecting data once at each switching moment). i 1 , i 2 , v o and v r This allows for the control of multiple system parameters (such as resonant inductance). L r Resonant capacitor C r Resonant resistor R r Transformer magnetizing inductance Lm Transformer leakage inductance L 2. Leakage inductance parasitic resistance R 2 Output capacitor C Output capacitor parasitic resistance R C Load resistance R load Online identification of ).

[0036] In this invention, the physical constraints of the LLC-aware physical information nested neural network contain all the dynamic information of the system, eliminating the need for calibration or feature extraction through external large-signal interference, thus reducing system complexity and operating costs. Furthermore, the training of the LLC-aware physical information nested neural network is a continuous optimization process, ensuring that parameter estimation and state awareness are always kept in optimal condition. Therefore, it can be directly embedded into existing systems without additional development or configuration, achieving seamless integration and rapid deployment.

[0037] This invention employs the implicit Runge-Kutta (IRK) method to discretize the continuous-time dynamic model of the LLC resonant converter, thereby transforming the continuous-time dynamics into an algebraic relationship between discrete-time variables. The existence of the data reconstruction network ensures the prediction of unobservable points. Ultimately, while accurately identifying the parameters, the system state sampled at the switching frequency is used only to reconstruct the nanosecond-level converter dynamics. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 This is a flowchart of the LLC sensing-type physical information nested neural network parameter estimation method applicable to LLC resonant converters in this invention.

[0040] Figure 2 This is a schematic diagram of the LLC resonant converter in this invention.

[0041] Figure 3 This is a schematic diagram of the LLC-sensory physical information nested neural network in this invention.

[0042] Figure 4 This is a schematic diagram of data acquisition under over-resonance conditions.

[0043] Figure 5 This is a schematic diagram of data acquisition under underresonant conditions. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0045] like Figure 1 As shown in this embodiment, the LLC sensing physical information nested neural network parameter estimation method for LLC resonant converters includes the following steps:

[0046] Dynamic modeling of S1.LLC resonant converter.

[0047] S11. Construct a continuous-time state-space model of the LLC resonant converter.

[0048] The main parameters of the LLC resonant converter are as follows: Figure 2 As shown, a continuous-time state-space model is constructed and represented as: in, A For transformer turns ratio; V F2 The forward voltage drop of the secondary rectifier diode; K For the upper bridge arm controllable switching transistor MOSFET ( S 1 Switch status signal, when S 1 When conducting, K =1, when S 1 When shut down, K =0; K 2 Secondary rectifier diode D 1 、D 2 On status signal, when any D 1 、D 2 When conducting, K 2 =1, when D 1 、D 2 When neither is conducting, K 2 =0. λ The set of parameters to be identified includes the resonant inductor. L r Resonant capacitor Cr Resonant resistor R r Transformer magnetizing inductance L m Transformer leakage inductance L 2. Leakage inductance parasitic resistance R 2 Output capacitor C Output capacitor parasitic resistance R C Load resistance R load wait.

[0049] S12. Discretize the continuous-time state-space model of the LLC resonant converter.

[0050] The implicit Runge-Kutta (IRK) method is used to discretize the continuous-time state-space model of the LLC resonant converter in step S11, so as to obtain the system state evolution relationship in each switching state, thereby constructing a physical layer model that can be used for neural network training and accelerating the efficiency of parameter identification. in, u Represents the system state variables (corresponding to the primary current of the LLC resonant converter). i 1 Secondary current i 2 Output voltage v o and resonant capacitor voltage v r ), { a ij , b j , c j The value of} can be found in the IRK butcher table. The initial state time for each topology is... k The final state is k+1 The intermediate state time is k+c i , q To indicate the order of the IRK frame, refer to formula (1). f [ u ( k+c j ); λ ] represents the intermediate state function model of the system variables.

[0051] S2. Construct an LLC-perceptual physical information nested neural network.

[0052] On embedded platforms, such as DSPs, ARMs, or cloud platforms, build the LLC-Aware PINN algorithm framework, which is an LLC-aware physical information nested neural network. Figure 3 As shown, the system comprises two interconnected networks: a Data Reconstruction Network (DRN) and a Physical Information Nested Neural Network (PINN), used for online parameter identification of LLC resonant converters. The DRN includes a resonant state determination layer, a K2 operation layer, a pseudo-label generation layer, a constraint layer, and a data reconstruction layer. The PINN includes an intermediate state mapping layer and a physical layer, using the discretized continuous-time state-space model of the LLC resonant converter as the training model for the physical layer.

[0053] S21. Define the input of the LLC-sensory physical information nested neural network.

[0054] Figure 4 Figure 5 These represent over-resonance and under-resonance states, respectively. Under standard MCU sampling configuration, the LLC resonant converter samples the system state, including the primary current, each time the modulated wave intersects with the carrier wave. i 1 Secondary current i 2 Output voltage v o and resonant capacitor voltage v r And switch status signals. For example, the first... l The initial state of the system state variables under each switching state is represented as { i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l The final state is represented as { )}. i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 Add switch status signal K The time interval Δ between the next switching state and the next switching state t , forming the first l Data items { i 1 (l ), i 2 ( l ), v o ( l ), v r ( l ), K ( l ), Δ t ( l ), i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 These observable physical quantities will serve as the input to the neural network.

[0055] S22. Determine the resonance state in the resonance state determination layer.

[0056] F Indicates the resonant mode, when the switching frequency ( f s >Resonant frequency ( f r )hour, F =1, when f s < f r hour, F =-1. Based on the secondary current corresponding to the final state of a switch. i 2 ( l+1 Whether the value is zero can determine the resonance state. i 2 ( l+1 )=0, f s < f r , F =-1, otherwise F =1, denoted as F ( l ).

[0057] S23. Perform operations on K2 in the K2 operation layer. K 2 It cannot be directly observed, but can be represented as: in, St It is a binary indicator for the topology stage. 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.

[0058] S24. Generate pseudo-tags in the pseudo-tag generation layer. F(l) The data is combined with the input data and fed into a feedforward neural network to infer the system state and the timing transitions of the topology boundaries within each switching interval. Let ( l+m ) indicates the ( l ) and the ( l+ 1 At the topological transformation time between ) sampling points, the variables to be inferred include The calculation for each hidden layer in a feedforward neural network is as follows: .in p Representation layer index, w ij ( p-1 ) indicates the connection of the () p-1 ) layer i The first neuron and the second p Layer j The weights of each neuron, b j ( p ) is the first p Layer j Bias of each neuron I Indicates the ( p-1 The number of neurons in the layer J Indicates the first p The number of neurons in the layer σ This represents a non-linear function, specifically the Sigmoid function. 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: in, and Representing variables respectively u The maximum and minimum values ​​of all sampled data; the output is used as pseudo-labels for downstream physical information nested neural networks. S26. The data reconstruction layer reformats each original data into two data lines, which correspond to two topological stages within the exchange interval, forming the input structure required by the downstream physical information nested neural network;

[0059] Let the reconstructed data index be denoted as n , No. ( l ) original data {i 1 ( l ) , i 2 ( l ) , v r ( l ) , v o ( l ) , i 1 ( l+1 ) , i 2 ( l+1 ) , v r ( l+1 ) , v o ( l+1), ∆t ( l ) , K ( l ) , F ( l The data is divided into two reconstructed data sets: the first... n Data item:{ i 1 ( n ):= i 1 ( l ), i 2 ( n ):= i 2 ( l ), v r ( n ):= v r ( l ), v o ( n ):= v o ( l ), v o ( n+1 ):= i 1 ( l+m ), i 2 ( n+1 ):= i 2 ( l+m ), v r ( n+1 ):= vr ( l+m ), v o ( n+1 ):= v o ( l+m ), K ( n ):= K ( l ), ∆ t ( n ):= t ( l+m ), K 2 ( n ):=-1* K ( l )* F ( l )}. No. n +1 data entry: { i 1 ( n+1 ):= i 1 ( l+m ), i 2 ( n+1 ):= i 2 ( l+m ), v r ( n+1 ):= v r ( l+m ), v o ( n+1 ):= v o ( l+m ), i 1 ( n+2 ):= i 1 ( l+1 ), i 2 ( n+2 ):= i 2 ( l+1 ), v r ( n+2 ):= v r ( l+1 ), v o (n+2 ):= v o ( l+1 ), K ( n+1 ):= K ( l ), ∆ t ( n+1 ):=∆ t ( l )- t ( l+m ), K 2 ( n+1 ):= -1* K ( l )* F ( l )}.

[0060] 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 in the corresponding topological interval.

[0061] S28. The physical layer receives the intermediate state output from the intermediate state mapping layer, substitutes it into the LLC resonant converter IRK model in formulas (2)-(8), and outputs the estimated initial and final states for each switching process. The weight parameters of the physical layer are the system parameters to be identified. λ .

[0062] S3. Construct the loss function.

[0063] loss function Loss Defined as the sum of the residuals of all outputs from the physical layer: in N This indicates the total number of reconstructed data.

[0064] S4. Deploy and execute the LLC-aware physical information nested neural network model.

[0065] The constructed LLC-aware physical information nested neural network model is deployed to an MCU platform or cloud platform. Executing the LLC-aware physical information nested neural network model includes periodically or event-triggered execution of data acquisition, data processing, model training, and parameter extraction processes.

[0066] The specific method for data acquisition is as follows: system status, including primary current, is acquired at each switching point. i 1 Secondary current i 2 Output voltagev o Resonant capacitor voltage v r The time interval Δ between each switch t and switch status K .

[0067] The specific data processing method is as follows: the collected data is transformed into training data samples adapted to the LLC perceptual physical information nested neural network model. Each training data sample contains... i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l ), K ( l ),Δ t ( l ), i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 Ten elements.

[0068] The specific method for model training is as follows: Input the training data samples into the LLC perceptual physical information nested neural network, run the LLC perceptual physical information nested neural network backpropagation algorithm, and as the loss function converges, the weight parameters in the physical layer... λ Gradually approaching the true parameter values.

[0069] The specific method for extracting parameters is as follows: when the training residual decreases to a certain threshold and tends to stabilize, the training process is stopped, and the system parameters are... λ The estimation results are ultimately extracted from the physical layer.

[0070] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An LLC-aware type physical information embedding neural network parameter estimation method suitable for an LLC resonant converter, characterized in that, Includes the following steps: S1. Dynamic modeling of LLC resonant converter: First, a continuous-time state-space model of LLC resonant converter is constructed, and then the implicit Runge-Kutta (IRK) method is used to discretize the continuous-time state-space model of LLC resonant converter. S2. Construct an LLC-aware physical information nested neural network, wherein the LLC-aware physical information nested neural network includes a connected data reconstruction network and a physical information nested neural network, for online parameter identification of the LLC resonant converter; The data reconstruction network includes a resonance state judgment layer, a K2 operation layer, a pseudo-label generation layer, a constraint layer, and a data reconstruction layer; in the resonance state judgment layer... F Indicates the resonant mode, when the switching frequency ( f s >Resonant frequency ( f r )hour, F =1, when f s < f r hour, F =-1; based on the secondary current corresponding to the final state of a switch. i 2 ( l+1 Whether the value is zero can determine the resonance state. i 2 ( l +1 )=0, f s < f r , F =-1, otherwise F =1, denoted as F ( l ); exist K 2 Computation layer K 2 Represented as: in, S t It is a binary indicator for the topology stage. 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. K For the upper bridge arm controllable switching transistor MOSFET ( S 1 Switch status signal, when S 1 When conducting, K =1, when S 1 When shut down, K =0; In the pseudo-tag generation layer, F ( l The data is combined with the input data and fed into the feedforward neural network to infer the system state and the timing transition of the topology boundary within each switching interval; In the constraint layer, the variables output by the pseudo-label generation layer are substituted into the mathematical model based on the physical constraint layer: In the data reconstruction layer, each original data is reformatted into two data points, which correspond to two topological stages within the exchange interval, forming the input structure required by the downstream physical information nested neural network. In the intermediate state mapping layer, a feedforward neural network is used; At the physical layer, the physical layer receives the intermediate state output from the intermediate state mapping layer, substitutes it into the LLC resonant converter IRK model, and outputs the estimated initial and final states for each switching process. The physical information nested neural network includes an intermediate state mapping layer and a physical layer, and uses the discretized LLC resonant converter continuous-time state-space model as the training model of the physical layer. Define the input of the LLC-sensory physical information nested into a neural network; S3. Construct the loss function; S4. Deploy and execute the LLC-aware physical information nested neural network model.

2. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 1, characterized in that, In step S1, the constructed continuous-time state-space model is represented as follows: in, i 1 The primary current, i 2 For secondary current, v o For output voltage, v r V is the voltage across the resonant capacitor, A is the transformer turns ratio, and V is the voltage across the capacitor. F2 The forward voltage drop of the secondary rectifier diode; K For the upper bridge arm controllable switching transistor MOSFET ( S 1 Switch status signal, when S 1 When conducting, K =1, when S 1 When shut down, K =0; K 2 Secondary rectifier diode D 1 、D 2 On status signal, when any D 1 、D 2 When conducting ，K 2 =1, when D 1 , D 2 When neither is conducting, K 2 =0; λ is the set of parameters to be identified, including the resonant inductance. L r Resonant capacitor C r Resonant resistance R r Transformer magnetizing inductance L m Transformer leakage inductance L 2 Leakage inductance and parasitic resistance R 2 Output capacitor C Output capacitor parasitic resistance R C Load resistance R load .

3. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 2, characterized in that, The training model for the physical layer, constructed based on the discretized continuous-time state-space model of the LLC resonant converter, specifically includes: in, u This represents the system state variable, corresponding to the primary current of the LLC resonant converter. i 1 Secondary current i 2 Output voltage v o and resonant capacitor voltage v r ； { a ij , b j , c j The value of} can be found in the IRK butcher table; the initial state time for each topology is . k The final state is k+1 The intermediate state time is k+c i , q To indicate the order of the IRK frame, refer to formula (1). f [ u ( k+c j ); λ ] represents the intermediate state function model of the system variables.

4. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 3, characterized in that, In step S2, an LLC-aware physical information nested neural network is built on an embedded platform or cloud platform; The specific method for defining the input of the LLC-sensory physical information nested neural network is as follows: Under the standard sampling configuration of the MCU, the LLC resonant converter will collect the system state, including the primary current, each time the modulated wave intersects with the carrier. i 1 Secondary current i 2 Output voltage v o and resonant capacitor voltage v r and switch status signals; No. l The initial state of the system state variables under each switching state is represented as { i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l The final state is represented as { )}. i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 Add switch status signal K The time interval Δ between the next switching state and the next switching state t , forming the first l Data items { i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l ), K ( l ), Δ t ( l ), i 1 ( l+1 ), i 2 ( l +1 ), v o ( l+1 ), v r ( l+1 These observable physical quantities serve as the inputs to the neural network.

5. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 4, characterized in that, In the pseudo-tag generation layer, let ( l+m ) indicates the ( l ) and the ( l+1 At the topological transformation time between ) sampling points, the variables to be inferred include { i 1 ( l+m ), i 2 ( 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: in p Representation layer index, w ij(p-1) Indicates the connection of the ( p-1 ) layer i The first neuron and the second p Layer j The weights of each neuron, b j(p) It is the first p Layer j Bias of each neuron I Indicates the ( p-1 The number of neurons in the layer. J Indicates the first p The number of neurons in the layer σ This represents a nonlinear function, specifically the Sigmoid function; in the constraint layer, the mathematical model based on the physical constraint layer is: in, and Representing variables respectively u The maximum and minimum values ​​of all sampled data; the output is used as pseudo-labels for the downstream physical information nested neural network. i 1 ( l+m) , i 2 ( l+m ), v o ( l+m ), v r ( l+m ), t ( l+m In the data reconstruction layer, the reconstructed data index is denoted as... n , No. ( l ) original data { i 1 ( l ) , i 2 ( l ) , v r ( l ) , v o ( l ) , i 1 ( l+1 ) , i 2 ( l+1 ) , v r ( l+1 ) , v o ( l+1), ∆t ( l ) , K ( l ) , F ( l The data is divided into two reconstructed data sets: ● The first n Data item:{ i 1 ( n ):= i 1 ( l ), i 2 ( n ):= i 2 ( l ), v r ( n ):= v r ( l ), v o ( n ):= v o ( l ), v o ( n+1 ):= i 1 ( l+m ), i 2 ( n+1 ):= i 2 ( l+m ), v r ( n+1 ):= v r ( l+m ), v o ( n+1 ):= v o ( l+m ), K ( n ):= K ( l ), ∆ t ( n ):= t ( l+m ), K 2 ( n ):=-1* K ( l )* F ( l )}, ●No. n +1 data entry: { i 1 ( n+1 ):= i 1 ( l+m ), i 2 ( n+1 ):= i 2 ( l+m ), v r ( n+1 ):= v r ( l+ m ), v o ( n+1 ):= v o ( l+m ), i 1 ( n+2 ):= i 1 ( l+1 ), i 2 ( n+2 ):= i 2 ( l+1 ), v r ( n+2 ):= v r ( l+1 ), v o ( n+ 2 ):= v o ( l+1 ), K ( n+1 ):= K ( l ), ∆ t ( n+1 ):=∆ t ( l )- t ( l+m ), K 2 ( n+1 ):=-1* K ( l )* F ( l )}。 6. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 5, characterized in that, In the intermediate state mapping layer, the mathematical expression implemented by the feedforward neural network is completely consistent with formula (11); this layer maps the reconstructed input vector to the set of potential intermediate states in the corresponding topological interval; At the physical layer, the physical layer receives the intermediate state output from the intermediate state mapping layer. Substituting this into the LLC resonant converter IRK model in formulas (2)-(8), the estimated initial and final states for each switching process are: and The weight parameters of the physical layer are the system parameters to be identified. λ .

7. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 6, characterized in that, loss function Loss Defined as the sum of the residuals of all outputs from the physical layer: in N This indicates the total number of reconstructed data.

8. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 7, characterized in that, In step S4, the deployment and execution of the LLC-aware physical information nested neural network model specifically involves: The constructed LLC-aware physical information nested neural network model is deployed to an MCU platform or cloud platform. Executing the LLC-aware physical information nested neural network model includes periodically or event-triggered execution of data acquisition, data processing, model training, and parameter extraction processes.

9. The LLC sensing-type physical information nested neural network parameter estimation method for LLC resonant converters according to claim 8, characterized in that, The specific method for data acquisition is as follows: system status, including primary current, is acquired at each switching point. i 1 Secondary current i 2 Output voltage v o Resonant capacitor voltage vr The time interval Δ between each switch t and switch status K； The specific data processing method is as follows: the collected data is transformed into training data samples adapted to the LLC perceptual physical information nested neural network model. Each training data sample contains... i 1 ( l ), i 2 ( l ), v o ( l ), v r ( l ), K ( l ), Δ t ( l ), i 1 ( l+1 ), i 2 ( l+1 ), v o ( l+1 ), v r ( l+1 Ten elements; The specific method for model training is as follows: Input the training data samples into the LLC perceptual physical information nested neural network, run the LLC perceptual physical information nested neural network backpropagation algorithm, and as the loss function converges, the weight parameters in the physical layer... λ Gradually approaching the true parameter value; The specific method for extracting parameters is as follows: when the training residual decreases to a certain threshold and tends to stabilize, the training process is stopped, and the system parameters are... λ The estimation results are ultimately extracted from the physical layer.

Images