A power transmission line fault location method based on digital twinning and knowledge distillation

Abstract

The application provides a power transmission line fault positioning method based on digital twinning and knowledge distillation, constructs a digital twin body in double-way data synchronization with a physical power transmission line, and generates a virtual training sample set with an accurate fault distance label based on a physical information embedding strategy; a large-capacity teacher model is constructed for pre-training; the teacher model is migrated to a lightweight student model through knowledge distillation, and is deployed on an edge node of the power transmission line after compression; when a real fault occurs in the physical line, the verified real fault data is used to perform online fine-tuning on the student model based on domain adaptive incremental learning, and the updated parameters are fed back to the digital twin body. Through the closed-loop cooperative mechanism of the digital twin body and the edge reasoning node, the model parameters are continuously optimized by using the real recording wave data after the actual fault occurs, so that the fault positioning model can be evolved online and adaptively, and the distribution difference between offline training and actual operation can be effectively overcome.

Description

Technical Field

[0001] This invention relates to the field of power system fault location, specifically a method for fault location of transmission lines based on digital twins and knowledge distillation. Background Technology

[0002] Transmission lines form the backbone of the power system, undertaking the crucial task of transmitting electricity over long distances. Timely and accurate location of line faults is essential for minimizing outage time, reducing maintenance costs, and improving power supply reliability. Traditional fault location methods primarily rely on impedance methods and traveling wave methods, calculating the relationship between electrical quantities after a fault and the fault distance. However, under complex conditions such as high-resistance grounding and arc discharge, the location accuracy often falls short of practical requirements.

[0003] With the development of artificial intelligence technology, data-driven methods have gradually become a research hotspot for fault location in power transmission lines. Long Short-Term Memory (LSTM) networks and their variants (such as bidirectional LSTM and gated recurrent units) can effectively capture long-term dependencies in temporal data, achieving good results in fault segment identification and distance prediction tasks. Hybrid Convolutional-Long Short-Term Memory (CNN-LSTM) networks further combine the spatial feature extraction capabilities of convolutional layers with the temporal modeling capabilities of recurrent layers, further improving the location accuracy.

[0004] However, existing deep learning-based fault location methods still have significant shortcomings in practical applications. Transmission line faults are low-probability events, and samples with precise distance labels in historical fault waveforms are extremely scarce, leading to insufficient model training and poor generalization ability. Existing high-precision models typically have millions of parameters, making them unsuitable for deployment on edge computing nodes along the line and failing to meet the needs of real-time fault location. Furthermore, there are distributional differences between the simulation data used for offline training and real line data, resulting in a gradual degradation of model accuracy after deployment, and there is a lack of an effective mechanism for online adaptive updates using real fault samples. Previous studies have proposed data-driven methods based on physical information embedding, estimating system parameters from single-end fault waveforms and generating training data through electromagnetic transient simulations, but this method is only used for offline data augmentation and is not combined with an online update mechanism. Summary of the Invention

[0005] To address the problems mentioned in the background, this invention proposes a transmission line fault location method based on digital twins and knowledge distillation. It utilizes digital twins to generate large-scale, multi-condition fault samples in real-time simulation, and combines knowledge distillation technology to transfer knowledge from a complex teacher model to a lightweight student model. This significantly reduces the number of model parameters and computational complexity while maintaining high location accuracy, enabling real-time inference at edge nodes along the line. Simultaneously, through a closed-loop collaborative mechanism between the digital twins and edge inference nodes, the model parameters are continuously optimized using real waveform data after actual faults occur, allowing the fault location model to adaptively evolve online, effectively overcoming the distributional differences between offline training and actual operation. The technical solution provided by this invention is as follows:

[0006] A method for fault location of transmission lines based on digital twins and knowledge distillation includes the following steps:

[0007] Step S1: Construct a digital twin that is bidirectionally synchronized with the physical transmission line, and generate a virtual training sample set with accurate fault distance labels based on the physical information embedding strategy;

[0008] Step S2: Construct a large-capacity teacher model and pre-train it using the virtual training sample set generated in step S1 to learn the mapping relationship from the fault electrical quantity time sequence waveform to the fault section and fault distance;

[0009] Step S3: Transfer the teacher model trained in step S2 to the lightweight student model through knowledge distillation, and deploy it on the edge node of the transmission line after compression;

[0010] Step S4: When a real fault occurs in the physical line, the student model is fine-tuned online using verified real fault data based on domain adaptive incremental learning, and the updated parameters are sent back to the digital twin.

[0011] Preferably, step S1 specifically includes:

[0012] Step S11: Establish a digital twin of the physical transmission line and achieve bidirectional data synchronization. The digital twin uses a distributed parameter line model to describe the electrical characteristics of the transmission line and couples real-time meteorological data and line aging impedance correction coefficients. The physical transmission line pushes the instantaneous values ​​of three-phase voltage, three-phase current, and circuit breaker switching status to the digital twin. The digital twin calculates the residual between the measured values ​​and the predicted values. When the residual exceeds a preset threshold, online recalibration is triggered.

[0013] Step S12: Based on single-ended fault waveform data, the physical information embedding method is used to estimate the range of system parameters;

[0014] Step S13: Inject virtual faults in batches into the digital twin to generate a virtual training sample set with precise labels.

[0015] Preferably, step S12 specifically includes:

[0016] Establish the local terminal time-domain circuit equations in the αβ0 stationary coordinate system:

[0017]

[0018] in, This is the voltage measurement value of the j-th mode at the local terminal. Let be the measured current value of the j-th mode at the local terminal; This is the equivalent power supply voltage; and These are the equivalent power supply resistance and inductance to be estimated, respectively.

[0019] Applying the above equation at time t and time t List them separately, subtract the two equations, and you get:

[0020]

[0021] in, For voltage differential, For current differential, The difference is the derivative of the current; the least squares method is used to solve for it. and This allows us to determine the range of values ​​for load conditions, fault resistance, and fault inception angle.

[0022] Preferably, step S2 specifically includes:

[0023] Step S21: Construct a high-capacity teacher model. The teacher model uses a two-layer bidirectional gated recurrent unit as a shared feature extraction backbone network to output feature vectors. Two parallel output heads are fed in: a segment classification head and a distance regression head; the segment classification head divides the entire line length into K segments and receives feature vectors. The output is a K-dimensional logits vector, which is then converted into a probability distribution for each segment using the Softmax function; the distance regression head receives the feature vector. Then output a scalar value representing the precise distance from the beginning of the line to the fault point, without using an activation function;

[0024] Step S22: Pre-train the teacher model using a virtual training sample set. The total loss function is the weighted sum of segment classification loss and distance regression loss.

[0025] Preferably, step S3 specifically includes:

[0026] Step S31: Construct a lightweight student model, using a structure of cascaded one-dimensional depthwise separable convolution and unidirectional gated recurrent units, and setting up a segment classification head and a distance regression head;

[0027] Step S32: Train the student model based on knowledge distillation. The distillation of the segment classification head involves temperature scaling of the logits vectors of the teacher and student models, using KL divergence as the distillation loss. The distillation of the distance regression head uses mean squared error loss to make the distance predictions of the student model approximate the distance predictions of the teacher model. A projection layer is also introduced into the student model to map the intermediate features of the student model to the same dimensional space as the intermediate features of the teacher model.

[0028] Step S33: Perform structured pruning and INT8 quantization perceptual training on the student model in sequence, compress the model volume, and deploy it on the edge computing nodes along the line.

[0029] Preferably, step S4 specifically includes:

[0030] Step S41: When a real fault occurs in the physical line, the waveform recorded by the fault recorder and the verified fault distance tag are sent back to the cloud to form a real sample set.

[0031] Step S42: Input the virtual training sample set and the real sample set into the student model respectively, extract the intermediate layer features, and denot them as follows: and Calculate MMD distance:

[0032]

[0033] in, and These are the features of virtual samples and real samples, respectively. and These represent the number of virtual samples and the number of real samples, respectively. This is a nonlinear mapping function that maps the original features to the reproducing kernel Hilbert space. Denotes the norm in Hilbert space;

[0034] Step S43: Update the student model based on the elastic weight consolidation strategy; the loss function is:

[0035]

[0036] in, For localization loss on real samples, For the diagonal elements of the Fisher information matrix, These are the original parameters obtained from distillation training. These are the parameters after incremental updates. This is the regularization intensity coefficient;

[0037] Step S44: The updated student network model parameters from step S43 are sent back to the digital twin as the initial model for the next round of virtual fault simulation and data generation, thus completing the closed-loop co-evolution of the digital twin and edge reasoning.

[0038] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By constructing a digital twin synchronized with the physical transmission line's bidirectional data and generating large-scale virtual training samples with precise labels based on a physical information embedding strategy, this invention effectively solves the problem of insufficient model training caused by the scarcity of fault samples. Furthermore, this invention transfers a large-capacity teacher model to a lightweight student model through knowledge distillation, and deploys it on edge nodes after structured pruning and quantization, achieving rapid real-time fault inference while ensuring positioning accuracy. In addition, this invention, through a domain-adaptive incremental learning mechanism, enables the model to be fine-tuned online using real fault data and completes closed-loop evolution by feeding parameters back to the digital twin, thereby continuously adapting to changes in the line's operating state and effectively mitigating the model degradation problem caused by domain differences between simulation and real data. Attached Figure Description

[0039] 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:

[0040] Figure 1 This is the main flowchart of the present invention. Detailed Implementation

[0041] 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.

[0042] To make the above-mentioned objectives, features and effects of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1: A method for fault location of transmission lines based on digital twins and knowledge distillation, such as... Figure 1 As shown, it includes the following steps:

[0044] Step S1: Construct a digital twin that is bidirectionally synchronized with the physical transmission line, and generate a virtual training sample set with accurate fault distance labels based on the physical information embedding strategy;

[0045] Step S2: Construct a large-capacity teacher model and pre-train it using the virtual sample set generated in step S1 to learn the mapping relationship from the fault electrical quantity time sequence waveform to the fault section and fault distance;

[0046] Step S3: Transfer the teacher model trained in step S2 to the lightweight student model through knowledge distillation, and deploy it on the edge node of the transmission line after compression to realize rapid fault location and reasoning.

[0047] Step S4: When a real fault occurs in the physical circuit, the student model is fine-tuned online using verified real fault data based on domain adaptive incremental learning, and the updated parameters are sent back to the digital twin to complete the closed-loop evolution.

[0048] Furthermore, step S1 specifically includes:

[0049] Step S11: Establish a digital twin of the physical transmission line and achieve bidirectional data synchronization.

[0050] A digital twin is a high-fidelity mapping of a physical power line in digital space. This step establishes the digital twin in the cloud or at the dispatch master station. The twin uses a distributed parameter line model to describe the electrical characteristics of the transmission line. Unlike simplified lumped parameter models, the distributed parameter model considers the continuous distribution characteristics of line resistance, inductance, and capacitance along the line length, and can more accurately reflect the propagation behavior of voltage and current waves during fault transients. Simultaneously, the twin also couples real-time meteorological data (including ambient temperature, wind speed, and icing thickness) and line aging impedance correction coefficients to reflect the impact of environmental factors and line aging on electrical parameters.

[0051] To achieve state synchronization between the digital twin and the physical circuit, the physical side pushes instantaneous three-phase voltage, instantaneous three-phase current, and circuit breaker switching status to the twin at a set refresh rate via a synchronous phasor measurement device. The synchronous phasor measurement device has a high sampling frequency, enabling it to capture detailed transient changes in voltage and current before and after a fault. After receiving the data from the physical side, the twin continuously solves the circuit state equations and outputs predicted electrical quantities. Simultaneously, the twin also calculates the residual between the measured values ​​from the physical side and the twin's predicted values. When this residual exceeds a preset threshold, it indicates a significant drift in the circuit parameters (e.g., a sudden change in conductor temperature leading to sag or impedance shift). This triggers online recalibration of the twin, realigning the virtual model with the physical entity.

[0052] Step S12: Based on single-ended fault recording data, the range of system parameters is estimated using the physical information embedding method.

[0053] The purpose of this step is to extract physical information from limited real fault waveform data to guide the generation of subsequent large-scale simulation data, thereby reducing the distributional discrepancy between simulation and real data. Specifically, this step uses measurable three-phase voltage and current data from one end (i.e., one side of the line) to obtain modal components in the stationary coordinate system through Clarke transformation, establishes circuit equations, and estimates the value ranges of the following four types of system parameters: equivalent power supply impedance, load conditions, fault inception angle (i.e., the phase angle of the voltage waveform at the moment of fault occurrence), and fault resistance.

[0054] In the αβ0 stationary coordinate system obtained by Clarke transformation, for the local end (labeled as 1), the time-domain circuit equation for the j-th modal component (j takes the form of α, β, or 0) is:

[0055]

[0056] in, This is the voltage measurement value of the j-th mode at the local terminal. These are the measured current values ​​for the j-th mode at the local terminal; both of these are known quantities that can be directly obtained from the fault recorder. The voltage is the equivalent power supply voltage, which is an unknown quantity. and These are the equivalent power supply resistance and inductance to be estimated, respectively.

[0057] Due to the equivalent power supply voltage It is unknown, and the above equation cannot be solved directly. Therefore, the above equation is applied to the first half-cycle of the fault (time). The equations for the fault and the second half-cycle (time t) are listed separately. Then, the two equations are subtracted to eliminate the unknown power supply voltage term, resulting in:

[0058]

[0059] in, For voltage differential, For current differential, It is the difference between the derivatives of the current.

[0060] Within at least half a cycle of time following the fault, multiple sets of the above difference equations can be obtained. These equations are then combined to form an overdetermined linear system, which can be solved using the least squares method to determine the equivalent source impedance. and After obtaining the equivalent impedance, and combining it with the steady-state voltage and current phasors before the fault, we can further estimate the load conditions (phase difference between the power supply voltages on both sides) and the possible range of values ​​for the fault resistance.

[0061] Step S13: Inject virtual faults in batches into the twin to generate a virtual training sample set with precise labels.

[0062] Within the parameter range estimated in step S12, this step systematically injects a large number of virtual faults into the digital twin. Since these faults occur in virtual space, the fault location, fault type, fault resistance, fault inception angle, and system operating conditions of each virtual fault can be precisely set. Therefore, the generated training samples have accurate fault distance labels, eliminating the need for manual annotation.

[0063] Specifically, the virtual fault settings cover a combination of variations in the following dimensions:

[0064] Fault location: Multiple fault injection points are set at equal intervals along the entire length of the line to divide the line into several sections;

[0065] Fault types include single-phase ground fault, two-phase short-circuit fault, two-phase ground short-circuit fault, and three-phase short-circuit fault;

[0066] Fault resistance: Take multiple discrete values ​​within the range estimated in step S12, covering low-resistance faults and high-resistance faults;

[0067] Fault inception angle: Takes multiple discrete values ​​within the range of 0 degrees to 360 degrees;

[0068] System load level and meteorological conditions: Combined within the range estimated in step S12.

[0069] For each fault combination, the twin runs an electromagnetic transient simulation, recording the three-phase voltage and three-phase current time-series waveforms at the local terminal for a period of time after the fault occurs, and using the corresponding fault distance (the precise distance from the beginning of the line to the fault point) as a label. Summarizing all the time-series waveform-distance label pairs constitutes a virtual training sample set. Since the number of virtual faults is not limited by physical conditions, the size of this sample set can reach tens of thousands or even more, thus solving the problem of scarce real fault samples.

[0070] Furthermore, step S2 specifically includes:

[0071] Step S21: Construct a large-capacity teacher model; the teacher model uses a two-layer bidirectional gated recurrent unit (BiGRU) as a shared feature extraction backbone network, followed by two parallel output heads: a segment classification head and a distance regression head.

[0072] Specifically, the input to the teacher model consists of the three-phase voltage and three-phase current time-series waveforms at the local terminal over a period before and after the fault. Assuming the input sequence contains T time steps, and each time step is characterized by a 6-dimensional vector (three-phase voltage amplitude and three-phase current amplitude), the input can be represented as follows: Each of them BiGRU consists of two parallel gated loop units: a forward gated loop unit and a backward gated loop unit. The forward gated loop unit processes the input sequence in chronological order (from t=1 to t=T), while the backward gated loop unit processes the input sequence in reverse chronological order (from t=T to t=1). For the t-th time step, the internal computation process of the gated loop unit is as follows:

[0073] First, calculate the reset gate. and Update Gate :

[0074]

[0075]

[0076] in, , The weight matrix is ​​the input to the gate. , The circular weight matrix from the hidden state to the gate. , For bias vectors, For the Sigmoid function, This is the hidden state from the previous moment.

[0077] Then, calculate the candidate hidden state. :

[0078]

[0079] in, This represents element-wise multiplication (Hadamard product). It is the hyperbolic tangent function.

[0080] Finally, calculate the hidden state at the current time step:

[0081]

[0082] Update Gate Controlling the proportion of the hidden state retained from the previous moment to the current moment, resetting the door. This controls how much information from the previous hidden state is used to compute the candidate hidden state. Through this gating mechanism, the gated recurrent unit can effectively capture long-term dependencies in time-series data.

[0083] The final hidden state of the forward-gated loop unit is denoted as The final hidden state of the backward-gated loop unit is denoted as (Because backward processing starts from the last time step). Concatenating the two results yields the output feature vector of the BiGRU. Where D is the hidden state dimension of the one-way gated loop unit. This indicates vector concatenation.

[0084] BiGRU's output feature vector Simultaneously, two parallel output heads are fed in:

[0085] Segment Classification Header: Divides the entire line length into K segments (for example, for a 100km line, each segment is 5km long, for a total of 20 segments). The segment classification header contains a fully connected layer that receives... The output is a K-dimensional logits vector. Then, the probability distribution of each segment is converted using the Softmax function:

[0086]

[0087] The loss function for the segment classification head is cross-entropy loss. Let the true fault segment label of a sample be... The classification loss is then:

[0088]

[0089] Distance regression head: Based on determining the location of the faulty segment, it further outputs the precise fault distance. The distance regression head contains a fully connected layer that receives... The output is a scalar value representing the precise distance from the beginning of the line to the fault point. The distance regression head does not use an activation function (i.e., it's a linear output), and the loss function is the mean squared error loss. Let the true fault distance label for a sample be 'd', and the distance prediction value of the teacher model be... The regression loss is:

[0090]

[0091] Step S22: Pre-train the teacher model using a virtual sample set.

[0092] The virtual training sample set generated in step S13 is input into the teacher model constructed in step S21. Each sample contains the input time-series waveform. Corresponding fault section labels and fault distance label The total loss function of the teacher model is a weighted sum of the segment classification loss and the distance regression loss:

[0093]

[0094] Where N is the total number of virtual training samples, For the teacher model, the i-th sample is in the real segment The predicted probability, Let be the fault distance prediction value of the teacher model for the i-th sample. The weight coefficients for the segment classification task and the distance regression task are used to adjust the relative importance of the two tasks in joint training.

[0095] Furthermore, step S3 specifically includes:

[0096] Step S31: Construct a lightweight student model.

[0097] The student model is a fault location model deployed at edge computing nodes along the transmission line in this invention. Since the computing resources, storage capacity, and power consumption of edge nodes are strictly limited, the student model must have an extremely small number of parameters and extremely fast inference speed. Therefore, the student model constructed in this step adopts a structure of cascaded one-dimensional depthwise separable convolutions and unidirectional gated recurrent units, and also sets up two parallel outputs: a segment classification head and a distance regression head.

[0098] One-dimensional depthwise separable convolution decomposes standard convolution into two steps: the first step is depthwise convolution, which performs a one-dimensional convolution on each input channel to extract local features for each electrical quantity channel; the second step is pointwise convolution, which uses 1×1 convolutions to linearly combine the features of each channel, achieving cross-channel information fusion. Compared to standard convolution, depthwise separable convolution significantly reduces the number of parameters and computational cost while maintaining the ability to extract similar features.

[0099] In the student model, two one-dimensional depthwise separable convolutional layers are used at the front end to quickly extract fault-related local features (such as voltage drops and current surges at the moment of a fault) from the input three-phase voltage and three-phase current time-series waveforms. The output of the convolutional layers serves as the input to the back-end unidirectional gated recurrent unit. Unlike the bidirectional structure in the teacher model, the student model uses a unidirectional gated recurrent unit (processed only in ascending time sequence) to further reduce the number of parameters and computational latency. The internal computation of the unidirectional gated recurrent unit is consistent with that described in step S21, and its final hidden state is input to the subsequent output head. Similar to the teacher model, the final hidden state of the unidirectional gated recurrent unit in the student model is simultaneously fed into two parallel output heads:

[0100] Segment classification head: Contains a fully connected layer, outputting a K-dimensional logits vector, which is then processed by a softmax function to obtain the probability distribution of each segment. .

[0101] Distance Regression Head: Contains a fully connected layer and outputs accurate predicted fault distances. .

[0102] Step S32: Train the student model based on knowledge distillation.

[0103] This step utilizes the teacher model trained in step S22 to guide the training of the student model. Because the teacher model has stronger feature learning capabilities, its output probability distributions for each segment contain information about the relative relationships between segments. By having the student model mimic the soft output of the teacher model, segment recognition accuracy close to that of the teacher model can be achieved while maintaining a smaller model size. Knowledge distillation training is performed separately for the two output heads of the student model, while simultaneously aligning the intermediate layer features.

[0104] Distillation of the segment classification head: The input time-series waveforms from the virtual training sample set are simultaneously fed into both the teacher and student models. The teacher model performs forward computation using its trained parameters, and the segment classification head outputs a K-dimensional logits vector. Then, temperature scaling is applied to the logits of the teacher model to obtain the softened probability distribution. The temperature scaling process is as follows:

[0105]

[0106] Among them, temperature . The larger the value, the smoother the output probability distribution. The probability differences between segments are reduced, but the relative ranking relationship is preserved, thus conveying more information about the similarity relationship between segments to the student model.

[0107] Similarly, the segment classification head of the student model outputs a logits vector. After scaling at the same temperature, the result is obtained :

[0108]

[0109] The distillation loss of the segment classification head is measured using the Kullback-Leibler divergence, which measures the difference between the teacher soft distribution and the student distribution.

[0110]

[0111] Distillation of the distance regression head: Distillation of the distance regression head uses mean squared error loss to make the distance predictions of the student model approximate the distance predictions of the teacher model.

[0112]

[0113] in, and These are the distance predictions for the same sample from the teacher model and the student model, respectively.

[0114] Alignment of intermediate layer features: The intermediate layer features of the teacher model are derived from the output of BiGRU. The intermediate layer features of the student model come from the final hidden state of the one-way gated recurrent unit. Because the two models have different network topologies, their intermediate features differ in dimensionality and semantics, making direct distance calculation impossible. Therefore, a projection layer is introduced in the student model to map the student model's intermediate features to the same dimensional space as the teacher model's intermediate features. The projection layer is a fully connected layer.

[0115]

[0116] in, The projection weight matrix is... This is the projection bias vector. Through this projection layer, the intermediate features of the student model are transformed to the same dimension as the intermediate features of the teacher model. .

[0117] Feature alignment loss is the L2 distance between projected student features and teacher features:

[0118]

[0119] Combining the above three losses, the total loss function for distillation training is:

[0120]

[0121] in, is a coefficient.

[0122] Step S33: Compress the student model and deploy it on the edge nodes.

[0123] After distillation training, the student model is further compressed to adapt to the hardware limitations of the edge nodes. This step involves structured pruning and INT8 quantization-aware training in sequence.

[0124] Structured pruning refers to removing unimportant neurons or channels from a network according to certain rules. Specifically, the L1 norm of each hidden unit in a one-way gated recurrent unit is calculated as an importance measure, and hidden units and their corresponding connections with importance below a set threshold are removed. After pruning, the network topology is more regular, which is beneficial for efficient hardware implementation.

[0125] Quantization-aware training refers to simulating low-precision computation (in this step, 8-bit integers, i.e., INT8) during the training process, so that the model can still maintain high accuracy after quantization. Specifically, based on the student model trained by distillation, the weights and activation values ​​are quantized from 32-bit floating-point to 8-bit integers, and a few rounds of fine-tuning training are performed on the quantized model to adapt the model to the accuracy loss caused by quantization. The compressed student model is then deployed on edge computing nodes (such as merging units, smart terminals, or edge gateways) along the transmission line.

[0126] Furthermore, step S4 specifically includes:

[0127] Step S41: Collect and label real fault samples. When a real fault occurs in the physical line, the three-phase voltage and current waveforms recorded by the fault recorder, as well as the real fault distance labels verified by traveling wave ranging or manual line inspection, are transmitted back to the cloud to form a real sample set.

[0128] Step S42: Domain Adaptation Based on Maximum Mean Difference. A domain adaptation method based on maximum mean difference (MMD) is used to align the feature distributions of virtual and real samples. The virtual training sample set and the real sample set are input into the student model respectively, and the feature representations of both in the intermediate layer are extracted, denoted as... and The MMD distance is defined as:

[0129]

[0130] in, and These represent the number of virtual samples and the number of real samples, respectively. This is a nonlinear mapping function that maps the original features to the reproducing kernel Hilbert space. denoted as the norm in Hilbert space.

[0131] Step S43: Incremental learning based on Elastic Weight Consolidation. Based on the Elastic Weight Consolidation (EWC) strategy, while retaining the student network's existing knowledge, the network parameters are incrementally updated using real samples; the EWC regularization loss is:

[0132]

[0133] in, For localization loss on real samples, For the diagonal elements of the Fisher information matrix, These are the original parameters obtained from distillation training. These are the parameters after incremental updates. This is the regularization intensity coefficient;

[0134] Step S44: Model Parameter Backhaul and Closed-Loop Evolution. The updated student network model parameters from Step S43 are backhauled to the digital twin as the initial model for the next round of virtual fault simulation and data generation, completing the closed-loop co-evolution of the digital twin and edge inference.

[0135] Example 2: The computer-readable storage medium of this example stores a computer program that, when executed by a processor, implements the steps in the transmission line fault location method based on digital twin and knowledge distillation in Example 1.

[0136] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0137] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0138] Example 3: The computer device of this example includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the transmission line fault location method based on digital twin and knowledge distillation in Example 1.

[0139] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0140] Those skilled in the art will clearly understand that each implementation can be achieved using software plus the necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0141] The above description is merely a preferred embodiment of the present invention and is 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. A method for fault location of transmission lines based on digital twins and knowledge distillation, characterized in that, Includes the following steps: Step S1: Construct a digital twin that is bidirectionally synchronized with the physical transmission line, and generate a virtual training sample set with accurate fault distance labels based on the physical information embedding strategy; Step S2: Construct a large-capacity teacher model and pre-train it using the virtual training sample set generated in step S1 to learn the mapping relationship from the fault electrical quantity time sequence waveform to the fault section and fault distance; Step S3: Transfer the teacher model trained in step S2 to the lightweight student model through knowledge distillation, and deploy it on the edge node of the transmission line after compression; Step S4: When a real fault occurs in the physical line, the student model is fine-tuned online using verified real fault data based on domain adaptive incremental learning, and the updated parameters are sent back to the digital twin.

2. The method for fault location of transmission lines based on digital twins and knowledge distillation according to claim 1, characterized in that, Step S1 is as follows: Step S11: Establish a digital twin of the physical transmission line and achieve bidirectional data synchronization. The digital twin uses a distributed parameter line model to describe the electrical characteristics of the transmission line and couples real-time meteorological data and line aging impedance correction coefficients. The physical transmission line pushes the instantaneous values ​​of three-phase voltage, instantaneous values ​​of three-phase current, and circuit breaker switching status to the digital twin. The digital twin calculates the residual between the measured value and the predicted value, and triggers online recalibration when the residual exceeds a preset threshold; Step S12: Based on single-ended fault waveform data, the physical information embedding method is used to estimate the range of system parameters; Step S13: Inject virtual faults in batches into the digital twin to generate a virtual training sample set with precise labels.

3. The method for fault location of transmission lines based on digital twins and knowledge distillation according to claim 2, characterized in that, Step S12 is as follows: Establish the local terminal time-domain circuit equations in the αβ0 stationary coordinate system: ； in, This is the voltage measurement value of the j-th mode at the local terminal. Let be the measured current value of the j-th mode at the local terminal; This is the equivalent power supply voltage; and These are the equivalent power supply resistance and inductance to be estimated, respectively. Applying the above equation at time t and time t List them separately, subtract the two equations, and you get: ； in, For voltage differential, For current differential, The difference is the derivative of the current; the least squares method is used to solve for it. and This allows us to determine the range of values ​​for load conditions, fault resistance, and fault inception angle.

4. The method for fault location of transmission lines based on digital twins and knowledge distillation according to claim 2, characterized in that, Step S2 is as follows: Step S21: Construct a high-capacity teacher model. The teacher model uses a two-layer bidirectional gated recurrent unit as a shared feature extraction backbone network to output feature vectors. Two parallel output heads are fed in: a segment classification head and a distance regression head; the segment classification head divides the entire line length into K segments and receives feature vectors. The output is a K-dimensional logits vector, which is then converted into a probability distribution for each segment using the Softmax function; the distance regression head receives the feature vector. Then output a scalar value representing the precise distance from the beginning of the line to the fault point, without using an activation function; Step S22: Pre-train the teacher model using a virtual training sample set. The total loss function is the weighted sum of segment classification loss and distance regression loss.

5. The method for fault location of transmission lines based on digital twins and knowledge distillation according to claim 4, characterized in that, Step S3 is as follows: Step S31: Construct a lightweight student model, using a structure of cascaded one-dimensional depthwise separable convolution and unidirectional gated recurrent units, and setting up a segment classification head and a distance regression head; Step S32: Train the student model based on knowledge distillation. The distillation of the segment classification head involves temperature scaling of the logits vectors of the teacher and student models, using KL divergence as the distillation loss. The distillation of the distance regression head uses mean squared error loss to make the distance predictions of the student model approximate the distance predictions of the teacher model. A projection layer is also introduced into the student model to map the intermediate features of the student model to the same dimensional space as the intermediate features of the teacher model. Step S33: Perform structured pruning and INT8 quantization perceptual training on the student model in sequence, compress the model volume, and deploy it on the edge computing nodes along the line.

6. The method for fault location of transmission lines based on digital twins and knowledge distillation according to claim 5, characterized in that, Step S4 is as follows: Step S41: When a real fault occurs in the physical line, the waveform recorded by the fault recorder and the verified fault distance tag are sent back to the cloud to form a real sample set. Step S42: Input the virtual training sample set and the real sample set into the student model respectively, extract the intermediate layer features, and denot them as follows: and Calculate MMD distance: ； in, and These are the features of virtual samples and real samples, respectively. and These represent the number of virtual samples and the number of real samples, respectively. This is a nonlinear mapping function that maps the original features to the reproducing kernel Hilbert space. Denotes the norm in Hilbert space; Step S43: Update the student model based on the elastic weight consolidation strategy; the loss function is: ； in, For localization loss on real samples, For the diagonal elements of the Fisher information matrix, These are the original parameters obtained from distillation training. These are the parameters after incremental updates. This is the regularization intensity coefficient; Step S44: The updated student network model parameters from step S43 are sent back to the digital twin as the initial model for the next round of virtual fault simulation and data generation, thus completing the closed-loop co-evolution of the digital twin and edge reasoning.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the transmission line fault location method based on digital twin and knowledge distillation as described in any one of claims 1-6.

8. A computer device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the transmission line fault location method based on digital twin and knowledge distillation as described in any one of claims 1-6.

Images