Method and device for non-invasive identification of rc impedance based on physical constraint transfer learning

Abstract

The application discloses an RC impedance non-intrusive identification method and device based on physical constraint migration learning, and relates to the technical field of impedance identification. The method comprises the following steps: collecting the voltage effective value, the current effective value and the voltage-current phase difference at the low-voltage end of the primary side, and using a neural network model which is pre-trained by a simulation source domain, fine-tuned by a limited number of target domain measured samples and trained by electrical response consistency, to identify the equivalent resistance value R and the capacitance value C of the secondary side online; the physical consistency constraint is that the predicted R and C are substituted into the transformer conversion equivalent circuit, the primary side electrical response is calculated inversely, and the consistency loss with the measured electrical response of the primary side is combined with the data loss, so that the model prediction result meets the measured label fitting and the circuit mechanism constraint at the same time. The application can reduce the high-voltage side measurement insulation risk and hardware cost, and can relieve the overfitting under the condition of measured small sample and the simulation-measured domain offset problem.

Description

Technical Field

[0001] This invention relates to the fields of high-voltage electrical equipment status sensing, transformer (such as intelligent large DC converter transformer) impedance identification and neural network measurement technology, specifically to a non-intrusive RC impedance identification method and device based on physical constraint transfer learning. Background Technology

[0002] Obtaining the equivalent insulation resistance and ground capacitance of a power distribution line quickly before commissioning is crucial for assessing its insulation condition, reducing the risks of blind commissioning, and improving fault handling efficiency. In AC commissioning instruments, a low-voltage power supply, inverter circuit, and step-up transformer are typically used to apply high voltage to the line under test. The secondary side of the line can be considered equivalent to a parallel RC load. The equivalent resistance R reflects insulation leakage characteristics, while the equivalent capacitance C reflects the line's ground capacitance or energy storage characteristics.

[0003] Existing methods that directly place voltage and current sensors on the secondary side of high-voltage circuits such as 6.5 kV require high-level insulation design and high-voltage side communication isolation, resulting in high hardware costs, complex on-site maintenance, and safety risks. In contrast, voltage, current, and phase information at the low-voltage end of the primary side are easy to acquire, and changes in secondary side impedance are reflected in the primary side response through transformer electromagnetic coupling and impedance conversion relationships. Therefore, it is possible to non-invasively identify secondary side impedance using primary side signals.

[0004] The analytical identification method based on the transformer T-type equivalent circuit has clear physical significance, but its accuracy is highly dependent on the accuracy of primary-side leakage impedance, secondary-side leakage impedance, excitation branch, and iron loss parameters. Actual high-voltage transformers undergoing commissioning exhibit excitation nonlinearity, temperature rise drift, iron loss variations, and parasitic wiring parameters, making it difficult for fixed-parameter analytical models to maintain high accuracy over long periods. Especially in high-ratio systems such as 220 V / 6.5 kV, the secondary-side load impedance is compressed by the square of the ratio when referred to the primary side, resulting in weaker current amplitude and phase changes on the primary side due to variations in secondary-side R and C. This makes measurement noise and phase errors easily amplified.

[0005] Pure data-driven neural network methods can learn the nonlinear mapping between primary-side features and secondary-side resistance R and capacitance C. However, obtaining labels for high-voltage experiments is difficult, and the amount of measured samples is usually small, making direct training prone to overfitting. Transfer learning can be pre-trained using simulated source domain data and then fine-tuned using a limited number of measured samples in the target domain. However, if relying solely on data errors, the model's prediction results may be affected by the simulation-measured domain offset, resulting in a lack of consistency in circuit response. In existing technologies, neither purely data-driven nor purely physical model-driven approaches have solved the problem of how to deeply integrate physical mechanisms with data-driven approaches to achieve high-precision identification under small sample conditions. This is especially true in high-ratio nonlinear applications for high-voltage test instruments, further increasing the difficulty of high-precision parameter identification.

[0006] Compared with the method of directly solving parameters by utilizing primary side electrical signals and combining equivalent circuit error minimization, this invention further introduces simulation source domain pre-training, target domain small sample transfer fine-tuning, electrical response consistency physical constraints, identifiability verification, and embedded inference-only deployment, thereby solving problems such as limited field samples, easy model overfitting, possible deviation of prediction results from circuit mechanism, and real-time operation of resource-constrained controllers. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a non-intrusive RC impedance identification method and device based on physical constraint transfer learning. Relying solely on measurable signals from the primary low-voltage side, it employs simulated source domain pre-training, noise enhancement of a limited number of measured samples in the target domain, fine-tuning of target domain transfer learning, and physical constraints on the consistency of primary side electrical response. This ensures that the equivalent resistance R and capacitance C of the secondary side parallel circuit output by the neural network identification model are both close to the measured labels and conform to the response law of the transformer equivalent circuit. This improves the safety, accuracy, and engineering deployability of online sensing for high-voltage systems with large transformer ratios.

[0008] On the one hand, a non-invasive RC impedance identification method based on physical constraint transfer learning includes:

[0009] In a high-ratio high-voltage system with electromagnetic coupling between the primary and secondary sides, primary side voltage and current signals are acquired at the low-voltage end of the primary side, and the effective values ​​of the primary side voltage, the effective values ​​of the primary side current, and the voltage-current phase difference are obtained based on the primary side voltage and current signals.

[0010] The effective value of the primary side voltage, the effective value of the primary side current, and the voltage-current phase difference are input into the trained neural network identification model, and the resistance and capacitance values ​​of the secondary side RC equivalent load are output online.

[0011] The training of the neural network identification model includes: embedding the physical constraint of electrical response consistency based on the transformer equivalent circuit as a hard constraint into the training loss function, so that at least one primary electrical response quantity obtained by back-calculating the resistance and capacitance values ​​of the secondary RC equivalent load predicted by the model through the transformer equivalent circuit is consistent with the measured corresponding electrical response quantity. The hard constraint and the data supervision loss jointly participate in the backpropagation optimization of the model parameters.

[0012] Preferably, the neural network identification model includes an input layer, at least one hidden layer, and an output layer; the input layer receives three features: the effective value of the primary side voltage, the effective value of the primary side current, and the voltage-current phase difference; the output layer outputs two parameters: resistance and capacitance; the number of hidden layers is one to five, and each hidden layer includes several neurons; during online inference, the hidden layer performs input normalization, affine operation of the fully connected layer, nonlinear activation, and output denormalization; wherein, the affine operation of the fully connected layer includes weight matrix multiplication and bias superposition.

[0013] Preferably, the training of the neural network identification model includes a simulated source domain pre-training process; the simulated source domain pre-training process generates simulated source domain samples based on the transformer equivalent circuit model or simulation model; the transformer equivalent circuit model includes primary side leakage impedance, secondary side equivalent leakage impedance, excitation branch, turns ratio, and secondary side RC equivalent load; wherein, the secondary side RC equivalent load impedance is converted to the primary side according to the square of the turns ratio, and a mapping relationship is established between the primary side electrical response and the secondary side resistance and capacitance.

[0014] Preferably, the training of the neural network identification model further includes a target domain measured sample enhancement process; the enhancement process only applies to the primary side input features of the measured training samples, while the measured test samples retain the original measured data unchanged; the enhancement process is used to simulate voltage sampling error, current sampling error, phase detection error, or fluctuations in field operating conditions.

[0015] Preferably, the training of the neural network identification model further includes a transfer fine-tuning process; the transfer fine-tuning process transfers the input layer and hidden layer parameters of the neural network model pre-trained in the simulated source domain to the target domain neural network model, freezes all hidden layer parameters of the source domain neural network model, and only fine-tunes the weight matrix and bias vector of the output layer.

[0016] Preferably, the training loss function L corresponding to the physical constraint of primary-side electrical response consistency includes the data supervision loss L. data And physical consistency loss L phy , satisfying L=L data +λL phy ; among which, L dataL is used to constrain the error between the resistance and capacitance values ​​predicted by the model and the measured labels. phy The physical constraint weight λ is used to constrain the consistency between the primary electrical response obtained by back-calculation of the predicted resistance and capacitance values ​​through the transformer equivalent circuit and the measured primary electrical response.

[0017] Preferably, the non-invasive RC impedance identification method based on physical constraint transfer learning further includes a parameter identifiability verification step, specifically including:

[0018] A perturbation is applied in the space of secondary-side resistance and capacitance parameters. The Jacobian matrix of the primary-side feature pairs of resistance and capacitance is calculated using finite difference. Based on the rank, singular value, or condition number of the Jacobian matrix, it is determined whether the resistance and capacitance values ​​in the current working range satisfy the local identifiability condition.

[0019] Preferably, the primary side electrical response includes the effective value of the primary side current, the primary side active power, the primary side apparent power, or the primary side equivalent impedance, and all the primary side electrical response are calculated based on the same transformer equivalent circuit model.

[0020] Preferably, the high-voltage system with large transformation ratio includes a high-voltage test instrument, a step-up transformer, or a nonlinear transformation ratio system with electromagnetic coupling between the primary low-voltage end and the secondary high-voltage end, wherein the secondary RC equivalent load can characterize the line insulation state and the ground distribution parameters.

[0021] On the other hand, a non-invasive RC impedance identification device based on physical constraint transfer learning includes:

[0022] A primary-side voltage sampling module is installed on the primary side of the electrical equipment to collect primary-side voltage signals;

[0023] A primary-side current sampling module is installed on the primary side of the electrical equipment to collect primary-side current signals;

[0024] The model storage module is used to store the neural network model parameters obtained after simulation source domain pre-training and target domain transfer fine-tuning; the controller is connected to the primary side voltage sampling module, the primary side current sampling module and the model storage module respectively, and is configured to execute the method to complete the online identification of the secondary side parallel equivalent resistance and capacitance values.

[0025] The result output module is used to output the resistance and capacitance values ​​identified by the controller.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) The present invention completes the secondary side RC impedance identification at the low voltage end of the primary side, without the need to arrange sensors and communication links on the high voltage side, thus reducing the high voltage insulation risk, hardware cost and on-site maintenance difficulty;

[0028] (2) This invention utilizes simulated source domain pre-training to learn a wide range of basic mappings, and then uses a limited number of target domain measured samples for transfer fine-tuning, which can alleviate the problems of scarce measured label samples and easy overfitting of pure measured training.

[0029] (3) The present invention introduces a physical constraint on the consistency of primary current, so that the predicted resistance and capacitance values ​​not only meet the data fitting requirements, but also remain consistent with the equivalent circuit response of the transformer, thereby improving the stability and generalization ability under small sample conditions.

[0030] (4) This invention evaluates whether the secondary side resistance and capacitance values ​​can be stably recovered from the primary side features within a given working range through parameter identifiability test, which can help determine the training sample range and reduce the risk of misidentification caused by weakly sensitive areas.

[0031] (5) This invention decouples training and inference. In the online phase, normalization, matrix multiplication, activation function and denormalization are performed only on the controller. It has the engineering advantages of low computing power, low latency and easy embedding into high-pressure test instruments. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the non-invasive RC impedance identification method based on physical constraint transfer learning in this embodiment.

[0034] Figure 2 This is a schematic diagram of the non-invasive parameter identification device for the 10 kV high-voltage test instrument in this embodiment.

[0035] Figure 3 This is a schematic diagram of the overall process of training and deploying the non-invasive identification method in this embodiment;

[0036] Figure 4 This is a schematic diagram of the transformer equivalent circuit in this embodiment;

[0037] Figure 5 This is a schematic diagram of the neural network recognition model structure in this embodiment;

[0038] Figure 6This is a schematic diagram illustrating the physical constraint loss for electrical response consistency in this embodiment;

[0039] Figure 7 This is a schematic diagram of the embedded online identification process in this embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0041] like Figure 1 As shown, this application provides a non-intrusive RC impedance identification method based on physical constraint transfer learning, which includes the following steps:

[0042] S101, the primary side voltage signal and the primary side current signal are collected at the low-voltage end of the primary side of the high-voltage system with electromagnetic coupling relationship between the primary side and the secondary side, and the effective value of the primary side voltage, the effective value of the primary side current, and the voltage and current phase are obtained based on the primary side voltage signal and the primary side current signal.

[0043] S102, the effective value of the primary side voltage, the effective value of the primary side current, and the voltage-current phase difference are input into the trained neural network identification model, and the resistance and capacitance values ​​of the secondary side RC equivalent load are output online.

[0044] The training of the neural network identification model includes: embedding the physical constraint of electrical response consistency based on the transformer equivalent circuit as a hard constraint into the training loss function, so that at least one primary electrical response quantity obtained by back-calculating the resistance and capacitance values ​​of the secondary RC equivalent load predicted by the model through the transformer equivalent circuit is consistent with the measured corresponding electrical response quantity. The hard constraint and the data supervision loss jointly participate in the backpropagation optimization of the model parameters.

[0045] Furthermore, the high-voltage system with large transformation ratio includes a high-voltage test instrument, a step-up transformer, or a nonlinear transformation ratio system with electromagnetic coupling between the primary low-voltage end and the secondary high-voltage end. The secondary RC equivalent load can characterize the line insulation state and the ground distribution parameters.

[0046] The following explanation uses the primary current consistency constraint, 3-18-18-2 fully connected neural network, and 220 V / 6.5 kV step-up transformer experimental platform as examples.

[0047] like Figure 2As shown, the non-invasive RC impedance identification method based on physical constraint transfer learning in this embodiment is applied to the non-invasive RC impedance identification device based on physical constraint transfer learning.

[0048] The non-intrusive RC impedance identification device based on physical constraint transfer learning includes a low-voltage power supply or battery, a DC-DC boost circuit, a single-phase inverter, a 220 V / 6.5 kV boost transformer, a secondary-side RC equivalent load, a primary-side voltage sampling module, a primary-side current sampling module, a model storage module, a controller, and a result output module.

[0049] In this embodiment, the secondary-side RC equivalent load is used to simulate the insulation resistance and ground capacitance of the power distribution line. The controller outputs the identification results of the secondary-side resistance value R and capacitance value C based on the primary-side sampling information. Specifically, the secondary-side RC equivalent load can be a secondary-side RC parallel equivalent load.

[0050] Specifically, the primary-side voltage sampling module is located on the primary side of the electrical equipment and is used to collect the primary-side voltage signal; the primary-side current sampling module is located on the primary side of the electrical equipment and is used to collect the primary-side current signal; the model storage module is used to store the neural network model parameters obtained after simulation source domain pre-training and target domain transfer fine-tuning; the controller is connected to the primary-side voltage sampling module, the primary-side current sampling module and the model storage module respectively, and is configured to execute the non-intrusive RC impedance identification method to complete the online identification of the parallel equivalent resistance and capacitance values ​​on the secondary side; the result output module is used to output the resistance and capacitance values ​​identified by the controller.

[0051] The controller includes an RMS value calculation module, a phase difference calculation module, and a model processing module. The RMS value calculation module is used to process the primary side voltage signal and the primary side current signal to obtain the RMS value U of the primary side voltage. p and the effective value of the primary side current I p The phase difference calculation module is used to calculate the voltage-current phase difference φ based on the primary side voltage signal and the primary side current signal; the model processing module includes a normalization unit, a neural network inference unit, and an inverse normalization unit. The normalization unit calls the input normalization parameters in the model storage module to normalize the three features U. p I p The φ values ​​are normalized, and the normalized feature vectors are input into the neural network inference unit. The neural network inference unit calls the model parameters in the model storage module to perform matrix multiplication, bias superposition, and activation function operations layer by layer to obtain the normalized output results, which are then input into the denormalization unit. The denormalization unit calls the denormalization parameters in the model storage module to restore the output results to the actual resistance value R and capacitance value C.

[0052] like Figure 3 As shown, the training and deployment process of this invention includes seven main steps: collecting features of the primary low-voltage side; generating simulated source domain samples based on equivalent circuits or Simulink models; training the source domain neural network model using the simulated source domain samples; collecting a limited number of measured samples of the target domain and performing noise enhancement; transferring the source domain model to the target domain model; adding physical constraints for electrical response consistency for fine-tuning; and deploying the trained model to the controller to output the resistance value R and capacitance value C online.

[0053] The parameters used in this embodiment are shown in Table 1 below.

[0054] Table 1. Parameters of the Example;

[0055]

[0056] Furthermore, such as Figure 4 As shown, the equivalent circuit of the transformer includes the primary side leakage impedance Z. p Secondary side leakage impedance Z s ′ Excitation branch Z m and the secondary load equivalent impedance Z L ′ Among them, Z p =R p +jX p R p X represents the equivalent resistance of the primary winding. p Indicates primary side leakage resistance; Z s ′ =R s ′ +jX s ′ R s ′ X represents the equivalent secondary resistance referred to the primary side. s ′ Z represents the secondary side leakage reactance converted to the primary side; m =R c ∥jX m R c X represents the iron loss resistance. m N represents the magnetizing reactance; p and N s These represent the number of turns in the primary and secondary windings, respectively, k=N p / N s Indicates the ratio; U p U represents the effective value of the primary side voltage. s This represents the effective value of the secondary voltage.

[0057] The secondary side RC equivalent load satisfies Z L =1 / (1 / R+jωC), where R represents the equivalent resistance of the secondary side in parallel, C represents the equivalent capacitance of the secondary side in parallel, and ω represents the operating angular frequency; the secondary side load referred to the primary side satisfies Z L ′ =k²Z L .based on Figure 4 The primary input impedance of the circuit shown can be expressed as Z. in =Z p +[Z m ∥(Z s ′ +Z L ′ Substituting the R and C predicted by the neural network into the above expression, the primary current response I corresponding to the predicted parameters can be obtained. calc =|U p / Z in |。 By the same Z in The active power P on the primary side can also be obtained from the current phasor. calc =Re(U p ·I calc * Apparent power S calc =|U p |·|I calc |and the primary-side equivalent impedance Z calc =U p / I calc This allows for the construction of physical constraints on the consistency of electrical responses in the form of current, power, or impedance.

[0058] Furthermore, the neural network identification model includes an input layer, at least one hidden layer, and an output layer; the input layer receives three features: the effective value of the primary side voltage, the effective value of the primary side current, and the voltage-current phase difference; the output layer outputs two parameters: the resistance value and the capacitance value; the number of hidden layers is one to five, and the number of neurons in each hidden layer is 4 to 128; during online inference, the hidden layer performs input normalization, fully connected forward propagation, nonlinear activation, and output denormalization operations.

[0059] In this embodiment, as Figure 5 As shown, this embodiment employs a lightweight fully connected neural network structure of 3-18-18-2. The input layer receives U p I p The system uses three primary side features: φ, R, and C. Two hidden layers each contain 18 neurons and employ the ReLU activation function. The output layer outputs two parameters, R and C. Before training, the input and output are normalized, and then denormalized after online inference to obtain the actual physical quantities.

[0060] It should be noted that the above 3-18-18-2 structure is only a preferred embodiment. In other embodiments, the number of hidden layers can be one to five, the number of neurons in each layer can be four to 128, and the activation function can be one or more of ReLU, Leaky ReLU, tanh, sigmoid, or GELU; the neural network can also be replaced with other differentiable regression models, as long as they can establish a mapping relationship between first-order features and parameters to be identified.

[0061] In this embodiment, the simulation source domain data during the simulation source domain pre-training process can be generated in batches using a transformer simulation model. In one embodiment, the simulation time is set to 0.06 s, and the effective value U of the primary side voltage is recorded after the high-voltage system with a large transformer ratio enters steady state. p , current effective value I p The simulation source domain data, including the phase difference φ, and randomly configured load combinations of secondary-side resistance R and capacitance C, generated approximately 1700 sets of simulation samples. Some samples were used for training, and others for testing. The simulation source domain pre-training employed the Adam optimizer and mean squared error loss function. This simulation source domain data covers a wide load parameter space, facilitating the learning of fundamental mapping relationships.

[0062] During the training of the simulated source domain model, a lightweight neural network recognition model is constructed, with U as the input. p I p The input and output data are R and C, respectively. The input and output data are normalized, and a supervised loss model is used to train the source domain model, enabling the model to initially learn the nonlinear mapping from the primary weak response characteristics to the secondary impedance parameters.

[0063] Furthermore, the actual test sample data in the target domain was collected by the experimental platform. In this embodiment, a total of 69 sets of test samples were collected, of which 52 sets were used as training samples and 17 sets were used as independent test samples. To avoid leakage of test information, noise injection enhancement was performed only on the training samples. The noise could be applied to U separately. p I p φ and φ are used to simulate voltage sampling error, current sampling error, phase detection deviation, and fluctuations in field operating conditions.

[0064] During transfer fine-tuning, the input and hidden layer parameters of the simulated source domain model are transferred to the target model. Specifically, a limited number of measured samples from the target domain are collected, and noise injection enhancement is performed only on the first-order input features of the training samples to simulate sampling errors and field disturbances. The input and hidden layer parameters of the neural network model pre-trained in the simulated source domain are transferred to the target domain neural network model, and all hidden layer parameters of the source domain neural network model are frozen. Only the weight matrix and bias vector of the output layer are fine-tuned, thereby correcting the real platform domain offset while preserving the global mapping capability of the simulated source domain. In other scenarios, some hidden layer parameters can be made available depending on the sample size. This process can reduce the number of trainable parameters, reduce overfitting on small samples, and utilize the basic electromagnetic mapping relationships already learned by the source domain model.

[0065] Furthermore, such as Figure 6 As shown, the construction process of the physical consistency loss constraint is as follows: First, the neural network calculates the effective value U of the primary side voltage. p The effective value of the primary side current I p The phase difference φ between voltage and current is used to predict the resistance value R and the capacitance value C; then, the predicted R and C are substituted into the equivalent circuit of the transformer to calculate the primary current I. calc ; Again I calc Compared with the measured primary side current I p In comparison, the current consistency loss L is formed. phy Finally, L phy Data loss L corresponding to R and C label errors data Weighted summation yields the total loss L=L data +λL phy The physical weight λ can be determined jointly based on the validation set error and the current residual to avoid either overly weak constraints failing to provide physical guidance or overly strong constraints amplifying the equivalent parameter error. This physical consistency loss constraint does not directly replace the network output with an analytical model; instead, it uses a relatively reliable primary-side current response relationship as a soft constraint to improve prediction stability under small sample conditions.

[0066] In one embodiment, the data supervision loss can be expressed as: L data =(1 / N)Σ[α R (R pred,i -R i )²+α C (C pred,i -C i )²], where R pred,i and C pred,i Let R and C be the predicted values ​​of the i-th sample, respectively. i and C i These are the corresponding labels, α R and α CThis is the dimensional balance coefficient.

[0067] The physical consistency loss can be expressed as: L phy =(1 / N)Σ|I calc -I pi |², where I calc I represents the primary current response obtained by substituting the predicted R and C into the equivalent circuit and then back-calculating it. pi This is the effective value of the measured current on the primary side.

[0068] Therefore, the total loss during fine-tuning of the target domain is: L = L data +λL phy This loss function constrains the model to be consistent with the measured R and C labels and the primary current response.

[0069] In addition to primary-side current consistency constraints, the physical constraints for electrical response consistency can also take the form of primary-side active power consistency, primary-side apparent power consistency, primary-side input equivalent impedance consistency, or a combination thereof. For the secondary-side RC parallel equivalent resistance R and capacitance C predicted by the neural network, firstly, Z... L = 1 / (1 / R + jωC) to obtain the secondary load impedance, then from Z L ' = k²Z L Converted to the primary side. Combined with the primary side leakage impedance Z. p Secondary side leakage impedance Z s ' and the excitation branch impedance Z m The equivalent input impedance Z on the primary side can be obtained. in calc = Z p + [Z m ∥ (Z s ' + Z L ' )], and then by I p calc = U p / Z in calc The primary current response is calculated in reverse, and P is further obtained. p calc = Re(U p ·(I p calc ) * ), S p calc = |U p ·(I p calc)*| or input equivalent impedance Z in calc During training, the calculated electrical response quantities are compared with the corresponding measured quantities to construct the physical loss for electrical response consistency.

[0070] In a preferred embodiment, parameter identifiability testing can be further performed. This testing is preferably performed offline during model training or before deployment to determine the training sample range, evaluate weakly sensitive regions, and assist in selecting input features, rather than requiring online identification after each output of R and C. Specifically, the quadratic side parameter vector is denoted as θ = [R, C]. T Let the first-order eigenvector be denoted as x=[U p ,I p ,φ] T For a given parameter point (R,C), small perturbations ΔR and ΔC are set respectively, and the Jacobian matrix is ​​estimated using central finite difference: J(:,1)≈[F(R+ΔR,C)-F(R−ΔR,C)] / (2ΔR), J(:,2)≈[F(R,C+ΔC)-F(R,C−ΔC)] / (2ΔC), where F represents the primary side characteristic response function obtained from the transformer equivalent circuit or simulation model. When J satisfies the full column rank condition, and the minimum singular value and condition number are within a preset range, R and C are considered to have local identifiability within this working range.

[0071] Furthermore, such as Figure 7 As shown, during online deployment, the controller synchronously samples the primary-side voltage and current. First, preprocessing is performed on the sampled signals, such as amplitude limiting, filtering, or moving average. Then, the fundamental amplitude and phase of the voltage and current are obtained through a fundamental wave extraction algorithm. This algorithm can employ discrete Fourier transform, Goertzel algorithm, synchronous sampling phasor calculation, or zero-crossing correction methods. Finally, the effective value U of the primary-side voltage is calculated based on the fundamental wave amplitude. p and the effective value of the primary side current I p φ is obtained from the difference between the voltage phase and the current phase.

[0072] Get U p I pAfter inputting φ, the controller's normalization unit calls the input normalization parameters in the model storage module to normalize the three features, and inputs the normalized feature vector into the neural network inference unit. The neural network inference unit calls the model parameters in the model storage module to perform matrix multiplication, bias superposition, and activation function operations layer by layer to obtain the normalized output result. Finally, the denormalization unit calls the denormalization parameters in the model storage module to restore the output result to the actual resistance value R and capacitance value C, and displays, uploads, or stores the result output module. The online inference process does not perform online training or iterative optimization, and is suitable for resource-constrained embedded platforms such as STM32.

[0073] This process eliminates the need to place sensors on the secondary high-voltage side, thereby reducing insulation risks and hardware complexity.

[0074] In one embodiment, the method described above was experimentally verified on a 220 V / 6.5 kV high-voltage test transformer. The experimental results show that the model with added physical constraint transfer learning can achieve high identification accuracy under a limited number of measured samples in the target domain, with mean absolute percentage errors of R and C reaching approximately 1.05% and 2.28%, respectively. This result demonstrates that the present invention can achieve high-precision online estimation of the secondary side RC impedance while avoiding direct measurement on the high-voltage side.

[0075] The physical constraints of this invention are not simple post-processing verifications, but rather the embedding of the physical model into the loss function during the training phase. This "training as constraint" approach ensures that the model is guided by physical laws when updating parameters, thereby fundamentally avoiding predictions that might arise from purely data-driven models and violate basic physical laws.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-invasive RC impedance identification method based on physical constraint transfer learning, characterized in that, include: In a high-ratio high-voltage system with electromagnetic coupling between the primary and secondary sides, primary side voltage and current signals are acquired at the low-voltage end of the primary side, and the effective values ​​of the primary side voltage, the effective values ​​of the primary side current, and the voltage-current phase difference are obtained based on the primary side voltage and current signals. The effective value of the primary side voltage, the effective value of the primary side current, and the voltage-current phase difference are input into the trained neural network identification model, and the resistance and capacitance values ​​of the secondary side RC equivalent load are output online. The training of the neural network identification model includes: embedding the physical constraint of electrical response consistency based on the transformer equivalent circuit as a hard constraint into the training loss function, so that at least one primary electrical response quantity obtained by back-calculating the resistance and capacitance values ​​of the secondary RC equivalent load predicted by the model through the transformer equivalent circuit is consistent with the measured corresponding electrical response quantity. The hard constraint and the data supervision loss jointly participate in the backpropagation optimization of the model parameters.

2. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 1, characterized in that, The neural network identification model includes an input layer, at least one hidden layer, and an output layer. The input layer receives three features: the effective value of the primary voltage, the effective value of the primary current, and the phase difference between the voltage and current. The output layer outputs two parameters: the resistance value and the capacitance value. The number of hidden layers is one to five, and each hidden layer includes several neurons. During online inference, the hidden layer performs input normalization, affine operation of the fully connected layer, nonlinear activation, and output denormalization. The affine operation of the fully connected layer includes weight matrix multiplication and bias superposition.

3. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 1, characterized in that, The training of the neural network identification model includes a simulated source domain pre-training process; the simulated source domain pre-training process generates simulated source domain samples based on the transformer equivalent circuit model or simulation model; the transformer equivalent circuit model includes primary side leakage impedance, secondary side equivalent leakage impedance, excitation branch, turns ratio, and secondary side RC equivalent load; wherein, the secondary side RC equivalent load impedance is converted to the primary side according to the square of the turns ratio, and a mapping relationship is established between the primary side electrical response and the secondary side resistance and capacitance.

4. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 3, characterized in that, The training of the neural network recognition model also includes a target domain test sample augmentation process; the augmentation process only applies to the first-order side input features of the test training samples, while the test samples retain the original test data unchanged; The enhancement process is used to simulate voltage sampling error, current sampling error, phase detection error, or fluctuations in field operating conditions.

5. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 4, characterized in that, The training of the neural network identification model also includes a transfer fine-tuning process; the transfer fine-tuning process transfers the input layer and hidden layer parameters of the neural network model pre-trained in the simulated source domain to the target domain neural network model, freezes all hidden layer parameters of the source domain neural network model, and only fine-tunes the weight matrix and bias vector of the output layer.

6. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 1, characterized in that, The training loss function L corresponding to the physical constraint of primary electrical response consistency includes the data supervision loss L. data And physical consistency loss L phy Satisfying L=L data +λL phy ; among which, L data L is used to constrain the error between the resistance and capacitance values ​​predicted by the model and the measured labels. phy The physical constraint weight λ is used to constrain the consistency between the primary electrical response obtained by back-calculation of the predicted resistance and capacitance values ​​through the transformer equivalent circuit and the measured primary electrical response.

7. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 1, characterized in that, It also includes a parameter identifiability verification step, specifically including: A perturbation is applied in the space of secondary-side resistance and capacitance parameters. The Jacobian matrix of the primary-side feature pairs of resistance and capacitance is calculated using finite difference. Based on the rank, singular value, or condition number of the Jacobian matrix, it is determined whether the resistance and capacitance values ​​in the current working range satisfy the local identifiability condition.

8. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 1, characterized in that, The primary side electrical response quantities include the effective value of the primary side current, the primary side active power, the primary side apparent power, or the primary side equivalent impedance. All of the primary side electrical response quantities are calculated based on the same transformer equivalent circuit model.

9. The non-invasive RC impedance identification method based on physical constraint transfer learning according to claim 1, characterized in that, The high-voltage system with large transformation ratio includes a high-voltage test instrument, a step-up transformer, or a nonlinear transformation ratio system with electromagnetic coupling between the primary low-voltage end and the secondary high-voltage end. The secondary RC equivalent load can characterize the line insulation state and the ground distribution parameters.

10. A non-invasive RC impedance identification device based on physical constraint transfer learning, characterized in that, include: A primary-side voltage sampling module is installed on the primary side of the electrical equipment to collect primary-side voltage signals; A primary-side current sampling module is installed on the primary side of the electrical equipment to collect primary-side current signals; The model storage module is used to store the neural network model parameters obtained after simulation source domain pre-training and target domain transfer fine-tuning; the controller is connected to the primary side voltage sampling module, the primary side current sampling module and the model storage module respectively, and is configured to execute the method described in any one of claims 1 to 9 to complete the online identification of the secondary side parallel equivalent resistance value and capacitance value. The result output module is used to output the resistance and capacitance values ​​identified by the controller.

Images