Boost converter parameter identification method based on impedance network and time sequence mask
By constructing a segmented decoupled physical information neural network and a timing mask concatenation method, combined with an impedance network and the Runge-Kutta algorithm, the model mismatch and multi-condition adaptability problems in Boost converter parameter identification are solved, achieving high-precision parameter monitoring and predictive maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-05-01
AI Technical Summary
Existing Boost converter parameter identification methods face problems such as parameter degradation, poor adaptability to multiple operating conditions, physical model mismatch, and difficulty in training convergence. In particular, they are difficult to achieve high-precision parameter monitoring and predictive maintenance in non-minimum phase systems.
A hybrid modeling approach combining data-driven and physical constraints is adopted. By constructing a segmented decoupled physical information neural network, combined with a time-series mask and impedance network, a discretized physical residual equation is established. The implicit Runge-Kutta algorithm is used for joint optimization to achieve synchronous updates of global and local parameters.
It achieves high-precision identification of Boost converter parameters under complex operating conditions with an error of less than 2.5%, meeting the needs of real-time monitoring and predictive maintenance.
Smart Images
Figure CN121960159A_ABST
Abstract
Description
A Boost Converter Parameter Identification Method Based on Impedance Network and Timing Mask Technical Field
[0001] This invention belongs to the interdisciplinary field of power electronics and artificial intelligence, specifically relating to a method for identifying parameters of a Boost converter based on impedance networks and timing masks. Background Technology
[0002] Power electronic converters, as core components of energy conversion, are widely used in new energy power generation, electric vehicles, aerospace, and industrial power supplies. Boost converters, as a typical type of boost converter, face unique challenges in parameter identification due to their non-minimum phase characteristics and discontinuous current mode.
[0003] In practical applications, passive components of converters, such as inductors and capacitors, will experience performance degradation due to long-term thermal stress, electrical stress, and environmental factors. This degradation is mainly manifested in a decrease in capacitance (C) and an increase in equivalent series resistance. Increased inductance and inductance value L, among other parameters, directly affect the system's steady-state performance and dynamic response, and can even lead to system failure. Therefore, real-time and accurate monitoring of these parameters is crucial for ensuring system reliability and enabling predictive maintenance.
[0004] Currently, the main methods for Boost converter parameter identification include:
[0005] Model-based methods, such as least squares and extended Kalman filtering, rely on accurate mathematical models, are sensitive to model errors, and struggle to handle parasitic parameters and switching nonlinearities.
[0006] Frequency domain analysis method: The frequency response is measured by injecting a disturbance signal, but this requires interrupting the normal operation of the system and has high requirements for measurement accuracy.
[0007] Data-driven methods, such as neural networks and support vector machines, require a large amount of labeled data, have limited generalization ability, and lack physical interpretability.
[0008] The existing technology has the following main drawbacks:
[0009] Physical model mismatch: Traditional state-space models do not fully consider the modulation of the current path by the switching state, resulting in the failure of physical constraints in the intermittent conduction mode.
[0010] Numerical instability caused by differential coupling: when considering capacitance When the output voltage derivative depends on the inductor current derivative, a differential algebraic equation is formed, which is difficult to converge using traditional numerical methods.
[0011] Poor adaptability to multiple operating conditions: Traditional methods usually assume a constant load and cannot simultaneously identify the inherent parameters of the circuit and the load parameters that change with the operating conditions in a single training session. Summary of the Invention
[0012] The main objective of this invention is to overcome the shortcomings of existing physical information neural networks when applied to Boost converter parameter identification, such as physical model mismatch, difficulty in training non-minimum phase systems, and inability to adapt to multi-level load changes in a single training session. The invention provides a solution that can identify the entire set of circuit parameters with high accuracy under actual engineering conditions.
[0013] To achieve the above objectives, the present invention adopts the following technical solution:
[0014] The parameter identification system proposed in this invention consists of four parts: a data acquisition module, a preprocessing module, a physical information neural network module, and an optimization identification module. The system adopts a hybrid modeling approach of data-driven and physical constraints, embedding the physical laws of the Boost converter into a deep neural network in the form of soft constraints to achieve a balance between physical interpretability and data adaptability.
[0015] S1: Construct a multi-condition transient dataset to collect the inductor current of the Boost converter during multiple load step changes. and capacitor voltage The discrete sampling data is recorded, and the corresponding switching duty cycle signal S and time step are recorded. ;
[0016] S2: Construct a segmented, decoupled physical information neural network. This network, located in the network parameter layer of the data-driven module, contains two types of parameters: one is a globally shared parameter set, including inductance L, capacitance C, inductor current, and resistance. Equivalent series resistance of capacitor On-resistance of the switching transistor Diode forward voltage drop and input voltage Another type is a partial set of parameters, including corresponding load resistors for multiple segments. In this process, the globally shared parameters are treated as time-invariant variables, while the local segmental parameters change with the working conditions.
[0017] S3: Perform temporal index mask concatenation. During the network forward propagation, generate a temporal mask based on the segmentation index of the input data, and concatenate the local segment variables corresponding to the current time period with the global shared variables to generate the equivalent physical parameter vector at the current moment.
[0018] S4: Establish an equivalent physical model of the impedance network and construct the discretized physical residual equation for the Boost converter. This equation is derived through the algebraic impedance voltage divider factor. The coupling effect between capacitor ESR and load resistance is explicitly described, and the complementary switching state 1-y is introduced to characterize the discontinuous energy injection characteristics of the Boost converter. This equation is then embedded as a regularization term into the loss function.
[0019] S5: Joint optimization and identification. The implicit Runge-Kutta algorithm is used to discretize the differential equation. By minimizing data error and physical residual, the neural network weights, global shared parameters and local segment parameters are updated synchronously, and finally the identification results are extracted.
[0020] Optionally, the discretized physical residual equations for the Boost converter in step S4 are constructed as follows:
[0021] definition For switch state variables, Define the impedance voltage divider factor Physical equation of inductor current The formula for construction is:
[0022] ;
[0023] Physical equation of capacitor voltage The construction formula is:
[0024] ;
[0025] in, In the equation The term is used to characterize the effect of the output voltage being applied in reverse to the inductor after being divided by the impedance network during the switch-off period; In the equation The term is used to characterize the effect of inductor current being injected into the capacitor branch through the impedance network.
[0026] Optionally, the temporal index mask concatenation in step S3 specifically includes: presetting three trainable scalar variables. Based on the time period index of the input data, a vector of all ones is constructed and multiplied with the corresponding scalar. Then, a concatenation operation is used to synthesize a unified load resistance vector across the data batch dimension. This vector automatically switches over time and participates in step S4. Factor calculation enables multi-load identification in a single training session.
[0027] Optionally, the neural network described in step S2 includes a scaling layer for normalizing the input data to the [-1,1] interval; the physical parameters are activated by an exponential function and multiplied by a preset scaling factor before calculation to ensure the numerical stability of the physical parameters in the positive real number domain.
[0028] Optionally, the optimization in step S5 adopts a cascaded strategy: first, the Adam optimizer is used for 200,000 iterations of global search, and then the L-BFGS-B algorithm is used for 50,000 iterations of second-order fine search. Attached Figure Description
[0029] Figure 1 is a schematic diagram of the method of the present invention;
[0030] Figure 2 is a schematic diagram of the segmented decoupled neural network architecture design;
[0031] Figure 3 shows the equivalent circuit diagram of the Boost converter considering parasitic parameters;
[0032] Figure 4 shows the parameter estimation results. Detailed Implementation
[0033] This invention provides a method for identifying Boost converter parameters based on impedance networks and timing masks, as shown in Figures 1 to 4. The method includes:
[0034] S1: Construct a multi-condition transient dataset to collect the inductor current of the Boost converter during multiple load step changes. and capacitor voltage The discrete sampling data is recorded, and the corresponding switching duty cycle signal S and time step are recorded. ;
[0035] In this embodiment of the invention, the data acquisition strategy adopts a combination of sparse sampling and multi-condition excitation:
[0036] Non-uniform sampling: Only feature points within each switching cycle are collected, selecting peak and trough points, reducing the amount of data by 75% while preserving the dynamic characteristics of the system.
[0037] Load step excitation: Arrange 3-5 load step changes within the sampling window to fully excite the dynamic response of the system at various operating points.
[0038] Anti-interference processing: The original data is subjected to moving average filtering, outlier removal, and adaptive normalization.
[0039] The data format is defined as a six-tuple: ,in A measurement tag for load resistance, used for time segmentation.
[0040] S2: Construct a segmented, decoupled physical information neural network. This network, located in the network parameter layer of the data-driven module, contains two types of parameters: one is a globally shared parameter set, including inductance L, capacitance C, inductor current, and resistance. Equivalent series resistance of capacitor On-resistance of the switching transistor Diode forward voltage drop and input voltage Another type is a partial set of parameters, including corresponding load resistors for multiple segments. In this process, the globally shared parameters are treated as time-invariant variables, while the local segmental parameters change with the working conditions.
[0041] In this embodiment of the invention, the neural network described in step S2 includes a scaling layer for normalizing the input data to the range of [-1, 1]; the physical parameters are activated by an exponential function and multiplied by a preset scaling factor before calculation to ensure the numerical stability of the physical parameters in the positive real number domain.
[0042] S3: Perform temporal index mask concatenation. During the network forward propagation, generate a temporal mask based on the segmentation index of the input data, and concatenate the local segment variables corresponding to the current time period with the global shared variables to generate the equivalent physical parameter vector at the current moment.
[0043] In this embodiment of the invention, the temporal index mask concatenation in step S3 specifically includes: presetting three trainable scalar variables. Based on the time period index of the input data, a vector of all ones is constructed and multiplied with the corresponding scalar. Then, a concatenation operation is used to synthesize a unified load resistance vector across the data batch dimension. This vector automatically switches over time and participates in step S4. Factor calculation enables multi-load identification in a single training session.
[0044] S4: Establish an equivalent physical model of the impedance network and construct the discretized physical residual equation for the Boost converter. This equation is derived through the algebraic impedance voltage divider factor. The coupling effect between capacitor ESR and load resistance is explicitly described, and the complementary switching state 1-y is introduced to characterize the discontinuous energy injection characteristics of the Boost converter. This equation is then embedded as a regularization term into the loss function.
[0045] In this embodiment of the invention, the discretized physical residual equations for the Boost converter in step S4 are constructed as follows:
[0046] definition For switch state variables, Define the impedance voltage divider factor Physical equation of inductor current The formula for construction is:
[0047] ;
[0048] Physical equation of capacitor voltage The construction formula is:
[0049] ;
[0050] in, In the equation The term is used to characterize the effect of the output voltage being applied in reverse to the inductor after being divided by the impedance network during the switch-off period; In the equation The term is used to characterize the effect of inductor current being injected into the capacitor branch through the impedance network.
[0051] S5: Joint optimization and identification. The implicit Runge-Kutta algorithm is used to discretize the differential equation. By minimizing data error and physical residual, the neural network weights, global shared parameters and local segment parameters are updated synchronously, and finally the identification results are extracted.
[0052] In this embodiment of the invention, the optimization in step S5 adopts a cascaded strategy: first, the Adam optimizer is used to perform a global search for 200,000 iterations, and then the L-BFGS-B algorithm is used to perform a second-order fine search for 50,000 iterations to solve the gradient vanishing problem in the high-dimensional parameter space.
[0053] The specific workflow of this invention is as follows:
[0054] S1. Construct a multi-condition dataset for the converter. A Boost converter operating in continuous conduction mode is used as the research object. Circuit parameters are set as follows: input voltage... The target output voltage is 72V, and the switching frequency is 20kHz. The global intrinsic parameters to be identified include: inductance. ,capacitance and parasitic parameters .
[0055] Data acquisition is achieved through a non-uniform sampling strategy, mainly including the following processes:
[0056] Sparse sampling: Only inductor current is sampled. and capacitor voltage The peak and trough values within each switching cycle have a data volume that is only equivalent to that of high-frequency sampling. .
[0057] Multi-condition construction: Load resistance during sampling time Three step changes occur to stimulate the transient response of the system.
[0058] Data preprocessing: Preprocessing and standardizing the collected data.
[0059] S2, Segmented Decoupled Neural Network Architecture Design: This invention constructs a deep physical information neural network, which includes the following core modules:
[0060] Input layer: Input vector , representing the current, voltage, duty cycle, complementary duty cycle, and time step at the current moment, respectively.
[0061] Hidden layers: Five fully connected layers are set, each with 50 neurons. The activation function is tanh to ensure second-order differentiability when calculating the physical equations.
[0062] Parameter decoupling layer: Initialize two sets of trainable variables in the network computation graph.
[0063] Global shared group: To ensure their physical meaning, these variables are stored in logarithmic form within the network and mapped to positive real numbers through exponential mapping during computation.
[0064] Local segment groups: These correspond to the load resistance values under three different operating conditions.
[0065] The network operates as follows: using inductor current... Output voltage Switch state S, time step As input, the data is first passed to the data-driven module. After processing by the data-driven module using a parameter decoupling architecture, the information is output to the latent state layer. The latent state layer integrates globally shared parameters with local segment parameters adapted to the current operating condition. Subsequently, the information from the latent state layer is transmitted in two paths: one path connects to the physical model module using backward equations, outputting the inductor current at the current moment. With capacitor voltage The reconstructed result is then fed into a physical model module using forward equations, which outputs the inductor current at the next time step. With capacitor voltage Prediction results.
[0066] S3: To identify multiple load segments simultaneously in a single training iteration, this invention incorporates mask concatenation logic into the network's forward propagation path. Specific steps include:
[0067] Index splitting: Obtain the time period index to which the input data belongs. When retrieving data, 120 sampling points were collected. The load changed three times, so the index splitting is divided into three segments: splitIdx1=120, splitIdx2=120, and splitIdx3=120.
[0068] Tensor Construction: Using tensor concatenation operations, a time-varying load vector with the same batch dimension as the input data is constructed. :
[0069] Dynamic routing: In subsequent physical equation calculations, the network automatically indexes to the corresponding location based on the current data point's position. The value is used in the calculation.
[0070] S4: Construction of a physical model based on algebraic impedance equivalence
[0071] Impedance voltage division factor introduction: Considering capacitor ESR, the current injection at the moment of switch turn-off is a complex impedance shunting process. This invention defines an algebraic impedance voltage division factor. : This factor directly describes the current distribution ratio between the load branch and the capacitor branch, avoiding the direct nesting of current differential terms in the voltage differential equation. Based on Holkieff's law, the circuit equations are established, and the following physical residual equations are constructed:
[0072] Inductor current state equation :
[0073]
[0074] In the formula This is a complementary switching state. The present invention introduces... The term precisely describes the effect of the output voltage being applied in reverse to the inductor through the impedance network, which is a typical characteristic of Boost as a non-minimum phase system.
[0075] Capacitor Voltage State Equation :
[0076] Here This demonstrates the discontinuous energy injection characteristic of the Boost converter. By multiplying by... The current component of the effective injected capacitor core is directly calculated, thus achieving decoupling of voltage and current derivatives.
[0077] Based on the above state equations, in order to improve numerical accuracy, a higher-order implicit Runge-Kutta discretization method is adopted to establish the forward / backward evolution equations of the latent states, thereby solving for the intermediate latent states. The specific formulas are as follows:
[0078]
[0079] Where q=20 is the IRK order. This is the Butcher matrix. Based on the above steps, the deep neural network and physical model are deeply integrated to establish a physical data fusion architecture.
[0080] S5: Define the composite loss function: In designing the loss function, it is necessary not only to embed data loss, but more importantly, to embed physical consistency loss, which includes forward evolution loss. and backward evolution loss Therefore, the loss function is designed as follows:
[0081] ;
[0082] The data loss is ,
[0083] Physical consistency loss is .
[0084] The physical loss is calculated using implicit Runge-Kutta time step constraints, forcing the network output to conform to... equation.
[0085] S6: Model Training and Online Prediction Applications
[0086] The optimization is performed in two stages. The first stage uses a global search with the Adam optimizer, a learning rate of 1e-3, and 200,000 iterations. This stage utilizes gradient descent to quickly find an approximate solution for the global parameters.
[0087] During the fine convergence phase, the L-BFGS-B optimizer is used with a maximum of 50,000 iterations to perform high-precision calibration of minute parasitic parameters using second-order curvature information.
[0088] After obtaining the trained parameter identification model, online prediction is performed. The current state and control variable are input to obtain the potential state at the current time. The discrete physics forward equation is used to predict the potential state at the next time step. Finally, the prediction result is output to complete the online prediction.
[0089] Based on the above steps, in the simulation experiment, the load resistance was continuously changed, resulting in the following step changes: 15Ω-20Ω-10Ω. After training, the identification error of inductor L was 0.36%, and the identification error of capacitor C was 0. The parasitic parameter identification error was controlled within 2.5%. Furthermore, the network successfully output three independent load resistance values, with relative errors all less than 0.2%. The specific identification results are shown in Tables 1 and 2 below:
[0090] Table 1. Identification results of globally invariant parameters
[0091] parameter True value 500 0.38 700 0.28 0.1 Identification value: 501.8 0.387 700 0.286 0.102 Error: 0.36% 0.18% 0% 2.0% 2.27% surface
[0092] Table 2. Results of Time-Varying Parameter Identification
[0093] parameter Actual value 15 20 10 Identification value: 14.994 20.022 10.006 Error: 0.04% 0.11% 0.06% surface
[0094] In this embodiment of the invention, as shown in Figure 4, the errors in the predicted and identified values are very small for both global parameters and time-varying parameters, meeting the identification accuracy requirements. This verifies that the PINN architecture based on the equivalent model of timing mask and impedance network proposed in this invention can effectively solve the parameter identification problem of non-minimum phase Boost converters under complex operating conditions.
Claims
1. A method for identifying parameters of a Boost converter based on impedance networks and timing masks, characterized in that, Includes the following steps: S1: Construct a multi-condition transient dataset to collect the inductor current of the Boost converter during multiple load step changes. and capacitor voltage The discrete sampling data is recorded, and the corresponding switching duty cycle signal S and time step are recorded. S2: Construct a segmented, decoupled physical information neural network. This network, located in the network parameter layer of the data-driven module, contains two types of parameters: one is a globally shared parameter set, including inductance L, capacitance C, inductor current, and resistance. Equivalent series resistance of capacitor On-resistance of the switching transistor Diode forward voltage drop and input voltage Another type is a partial set of parameters, including corresponding load resistors for multiple segments. In the entire training process, the global shared parameters are treated as time-invariant variables, while the local segment parameters change with the working conditions; S3: Perform temporal index mask concatenation. During the forward propagation of the network, a temporal mask is generated based on the segmentation index of the input data. The local segment variables corresponding to the current time period are concatenated with the global shared variables to generate the equivalent physical parameter vector at the current time. S4: Establish an equivalent physical model of the impedance network and construct a discretized physical residual equation for the Boost converter. This equation is expressed through an algebraic impedance voltage divider factor. The coupling effect between capacitor ESR and load resistance is explicitly described, and the complementary switching state 1-y is introduced to characterize the discontinuous energy injection characteristics of the Boost converter. This equation is embedded as a regularization term in the loss function. S5: Joint optimization and identification. The differential equation is discretized using the implicit Runge-Kutta algorithm. By minimizing data error and physical residual, the neural network weights, global shared parameters and local piecewise parameters are updated synchronously, and finally the identification results are extracted.
2. The method according to claim 1, characterized in that, The discretized physical residual equations for the Boost converter in step S4 are constructed as follows: Definition For switch state variables, Define the impedance voltage divider factor Physical equations of inductor current The formula for construction is: Physical equations of capacitor voltage The construction formula is: ;in, In the equation The term is used to characterize the effect of the output voltage being applied in reverse to the inductor after being divided by the impedance network during the switch-off period; In the equation The term is used to characterize the effect of inductor current being injected into the capacitor branch through the impedance network.
3. The method according to claim 1, characterized in that, Step S3, the temporal index mask concatenation, specifically includes: pre-setting three trainable scalar variables. Based on the time period index of the input data, a vector of all ones is constructed and multiplied with the corresponding scalar. Then, a concatenation operation is used to synthesize a unified load resistance vector across the data batch dimension. This vector automatically switches over time and participates in step S4. Factor calculation enables multi-load identification in a single training session.
4. The method according to claim 1, characterized in that, The neural network described in step S2 includes a scaling layer to normalize the input data to the range of [-1, 1]. Before calculation, the physical parameters are activated by an exponential function and multiplied by a preset scaling factor to ensure the numerical stability of the physical parameters in the positive real number domain.
5. The method according to claim 1, characterized in that, The optimization in step S5 adopts a cascade strategy: first, the Adam optimizer is used to perform a global search for 200,000 iterations, and then the L-BFGS-B algorithm is used to perform a second-order fine search for 50,000 iterations.