Physical model and lightweight neural network fused probabilistic fault positioning method and related device
By integrating physical models with lightweight neural networks and combining them with Bayesian confidence assessment, the problems of accuracy, robustness, data dependence, and computational efficiency in power system fault location are solved, achieving high accuracy, interpretability, and real-time performance, and adapting to complex power grid environments.
Patent Information
- Application Number
- CN202511562643.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-29
- Publication Date
- 2026-02-13
AI Technical Summary
Existing power system fault location methods have shortcomings in terms of accuracy, robustness, data dependence, model interpretability, and computational efficiency, especially in that they cannot provide confidence in the location results.
By integrating physical models and lightweight neural networks, a probabilistic fault location method is constructed through transient physical feature extraction, precise localization using lightweight residual networks, and Bayesian confidence evaluation. This method combines physical feature extraction, precise localization using lightweight residual networks, and Monte Carlo random deactivation techniques.
It improves the accuracy and robustness of fault location, reduces data dependence, enhances model interpretability, enables lightweight deployment and real-time response, and provides probabilistic confidence assessment of fault location results, adapting to new energy grid connection and complex power grid environments.
Smart Images
Figure CN121522344A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system planning and operation technology, specifically a probabilistic fault location method and related device that integrates physical models and lightweight neural networks. Background Technology
[0002] As the cornerstone of modern society, the stable and reliable operation of the power system is of paramount importance. With the large-scale grid connection of new energy sources such as wind and solar power, and the increasing complexity of load centers and transmission networks, the operational challenges facing the power system are becoming increasingly severe. Transmission lines, as a critical link in power transmission, can cause not only huge economic losses but also widespread power outages if they fail, severely impacting social production and residents' lives. Therefore, quickly and accurately locating transmission line faults is a core element in ensuring the safe and stable operation of the power system, reducing outage time and losses, and improving power supply reliability.
[0003] Currently, power system fault location technologies are mainly divided into two categories: impedance-based location methods and traveling wave-based location methods. Impedance-based methods, such as single-ended or double-ended impedance methods, infer the fault location by measuring the voltage and current at the fault point to calculate the line impedance. These methods are simple in principle and easy to implement, but their accuracy is easily affected by factors such as transition resistance, inaccurate line parameters, measurement errors, and fault type. Especially in high-resistance faults and complex power grids, the location accuracy often fails to meet requirements. Traveling wave-based methods, such as single-ended and double-ended traveling wave methods, utilize the transient traveling wave signal generated at the moment of the fault and calculate the fault location by detecting the arrival time of the traveling wave front. These methods theoretically have high location accuracy and are less affected by transition resistance, but they have extremely high requirements for sampling frequency and synchronization. Furthermore, in multi-terminal lines, branch lines, and when the traveling wave signal is severely interfered with by noise, wave front identification is difficult, posing challenges to the stability and reliability of the location.
[0004] In recent years, with the rapid development of artificial intelligence technology, especially deep learning, in various fields, its powerful feature learning and nonlinear mapping capabilities have brought new ideas to power system fault diagnosis. Some studies have attempted to use neural network models to learn fault modes directly from raw waveform data or data after simple feature extraction. However, purely data-driven deep learning methods often have the following limitations: First, they require massive amounts of labeled data covering various fault types and operating conditions for training, which is costly to acquire and difficult to fully cover in practical applications; second, the model is usually regarded as a "black box," and its decision-making process lacks interpretability. In power system applications with extremely high reliability requirements, the confidence level of its results is difficult to assess, causing dispatchers to have concerns when making important decisions; third, some complex deep learning models have a large number of parameters and consume high computational resources, making them difficult to deploy in field protection devices or edge computing devices and failing to meet real-time requirements.
[0005] In summary, existing fault location methods either lack accuracy and robustness, or suffer from bottlenecks in data dependence, model interpretability, and computational efficiency. There is an urgent need for a novel fault location method that can balance physical mechanisms and data intelligence, while also outputting location result confidence and improving decision reliability. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to provide a probabilistic fault location method and related device for power systems that integrates physical models and lightweight neural networks. It aims to solve the shortcomings of existing fault location methods in terms of accuracy, robustness, data dependence, model interpretability and computational efficiency, especially the inability to provide confidence of the location results.
[0007] A probabilistic fault localization method integrating a physical model and a lightweight neural network includes the following steps:
[0008] S1, Physical Feature Extraction
[0009] The instantaneous three-phase voltage and three-phase current waveform data after a transmission line fault are collected. A rough estimate of the fault location is calculated using at least one power system transient physical model. At least one auxiliary physical feature is extracted. The rough estimate and the auxiliary physical feature are combined to construct a standardized physical feature vector V.
[0010] S2, Lightweight Residual Network Precise Positioning
[0011] The standardized physical feature vector V output from step S1 is input into a pre-trained lightweight residual network model, and the high-precision deterministic fault location is output through inference.
[0012] S3, Bayesian confidence assessment
[0013] The Dropout layer of the lightweight residual network model in step S2 is activated using the Monte Carlo random deactivation technique. The standardized physical feature vector V output from step S1 is repeatedly input into the network model with the activated Dropout layer for N forward propagation cycles to obtain a fault location distribution set containing N samples. Statistical analysis is performed on the distribution set to calculate the sample mean μ and sample standard deviation σ. The sample mean μ is used as the final most probable fault location, and the sample standard deviation σ is used as an indicator to quantify the uncertainty or probability confidence of the location result.
[0014] Furthermore, step S1 includes the following sub-steps:
[0015] S1.1 Data Acquisition: When a fault occurs in a transmission line, the protection devices or monitoring units deployed at both ends of the line acquire instantaneous three-phase voltage and three-phase current waveform data for at least one cycle after the fault.
[0016] S1.2 Physical Model Calculation and Preliminary Feature Extraction: Calculate a rough estimate of the fault location based on the transient impedance model or transient traveling wave model. ;
[0017] S1.3, Auxiliary physical feature extraction: Extract at least one auxiliary physical feature, wherein the auxiliary physical feature includes transient energy within a specific time window after the fault occurs. Voltage phase angle at the moment of fault occurrence The transition resistance estimate obtained by using an algorithm to roughly estimate the transition resistance. And fault type identification based on comparison of phase current abrupt changes and unique thermal coding. ;
[0018] S1.4: Feature Vector Construction: The rough estimate of the fault location obtained in step S1.2 is used to construct the feature vector. Combined with the auxiliary physical features extracted in step S1.3, a standardized physical feature vector V is constructed.
[0019] Furthermore, step S1.2 specifically includes: Based on the transient impedance model: performing Fourier transform on the acquired voltage and current waveforms, calculating the phasor of the fundamental frequency component, and estimating a rough estimate of the fault location by solving the voltage equation based on the impedance per unit length of the line. Alternatively, based on the transient traveling wave model: a high-pass filter is used to extract the initial traveling wave signal generated at the instant of the fault from the raw data, and the arrival time of the traveling wave at the local end and the opposite end is identified by the traveling wave front arrival algorithm. A rough estimate of the fault location is then calculated according to the two-end traveling wave method formula. .
[0020] Furthermore, step S2 includes the following sub-steps:
[0021] S2.1 Model Input: The precise positioning module receives the standardized physical feature vector V as input;
[0022] S2.2 Network Inference: The core of the precise localization module is a pre-trained, lightweight residual network model, ResNet, which includes:
[0023] Input layer: The number of neurons in the input layer is equal to the dimension of the physical feature vector V;
[0024] Hidden layer: The hidden layer consists of 5 stacked residual blocks. Each residual block contains two fully connected layers and one shortcut connection. The activation function of the fully connected layers is LeakyReLU.
[0025] Output layer: The output layer contains a single neuron, employs a linear activation function, and directly outputs a scalar value as the high-precision deterministic fault location. .
[0026] S2.3 Model Training: The lightweight residual network model ResNet is trained using supervised learning with a large amount of fault data under different operating conditions generated on the simulation platform. The training objective is to minimize the model output. The mean square error (MSE) between the actual fault location set in the simulation and the actual fault location.
[0027] Furthermore, step S3 includes the following sub-steps:
[0028] S3.1 Activation of randomness: The confidence assessment module adopts the same or similar neural network structure as the precise positioning module, and when performing positioning calculation, it forcibly activates the Dropout layer of the lightweight residual network model. The Dropout layer randomly discards a portion of neurons with a preset probability during each forward propagation.
[0029] S3.2, Multiple random inferences: The same standardized physical feature vector V is used as input and repeatedly fed into the neural network with the Dropout layer activated for N forward propagation, so as to obtain a fault location distribution set D containing N slightly different localization results;
[0030] S3.3 Statistical Analysis and Results Output:
[0031] Calculate the statistical measure: Perform statistical analysis on the fault location distribution set D and calculate its sample mean. and sample standard deviation Among them, the sample mean The calculation formula is:
[0032] (1);
[0033] Sample standard deviation The calculation formula is:
[0034] (2);
[0035] Results presentation: The mean of the samples As the final most likely location of the failure; the standard deviation of the sample This serves as an indicator to quantify the uncertainty or probability confidence of the positioning result; and the positioning result and its credibility are presented in a visual manner, outputting a probabilistic positioning result including the most likely fault location and confidence interval.
[0036] Furthermore, the visualization of the positioning results and their reliability includes plotting a normal distribution curve centered on the sample mean μ and with the sample standard deviation σ as its width.
[0037] Furthermore, a probabilistic fault location device integrating a physical model and a lightweight neural network is characterized by comprising:
[0038] The physical feature extraction module is used to collect instantaneous three-phase voltage and three-phase current waveform data after a transmission line fault, calculate a rough estimate of the fault location using at least one power system transient physical model, extract at least one auxiliary physical feature, and combine the rough estimate with the auxiliary physical feature to construct a standardized physical feature vector V.
[0039] The lightweight residual network precise localization module is used to input the standardized physical feature vector V into the pre-trained lightweight residual network model and output a high-precision deterministic fault location through inference.
[0040] The confidence assessment module is used to activate the Dropout layer of the lightweight residual network model using Monte Carlo random deactivation technology. The standardized physical feature vector V is repeatedly input into the network model that activates the Dropout layer for N forward propagation to obtain a set of fault location distributions containing N samples. Statistical analysis is performed on the distribution set to calculate the sample mean μ and sample standard deviation σ. The sample mean μ is used as the final most probable fault location, and the sample standard deviation σ is used as an indicator to quantify the uncertainty or probability confidence of the location result.
[0041] This invention achieves a multi-dimensional technological breakthrough by integrating physical models with lightweight neural networks and combining them with a Bayesian confidence assessment mechanism, addressing the core shortcomings of existing fault location technologies. Specific beneficial effects are as follows:
[0042] 1. Improve positioning accuracy and robustness
[0043] This invention utilizes a physical feature extraction module to calculate the preliminary fault location using a transient impedance model or a traveling wave model. It then integrates auxiliary physical features such as transient energy and the initial fault angle to construct a standardized feature vector containing the core physical laws of the fault, providing reliable input for subsequent precise location. The lightweight residual network model learns the nonlinear relationship between physical features and the actual fault location, correcting the coarse estimate. This effectively overcomes the problems of traditional impedance methods being affected by transition resistance and traveling wave methods having difficulty identifying wavefronts, significantly improving the location accuracy and environmental adaptability under complex operating conditions (such as high-resistance faults and noise interference).
[0044] 2. Reduce data dependency and enhance model interpretability
[0045] Compared to purely data-driven deep learning methods, this invention extracts prior features through physical models, significantly reducing reliance on massive amounts of labeled data and lowering data acquisition and labeling costs. Simultaneously, the construction process of physical feature vectors (such as transient energy reflecting fault impact intensity and fault initial angle relating to transient process characteristics) gives the model inputs clear physical meaning, avoiding "black box" decision-making, enhancing the interpretability of fault location results, and meeting the stringent reliability and safety requirements of power systems.
[0046] 3. Achieve lightweight deployment and real-time response.
[0047] To address the issues of high computational resource consumption and difficulty in deploying complex deep learning models on edge devices, this invention employs a lightweight residual network structure (five stacked residual blocks, with LeakyReLU activation function selected for fully connected layers). This significantly reduces the number of model parameters and computational complexity while maintaining positioning accuracy. This design allows for efficient deployment in field protection devices or edge computing units, meeting the real-time requirements of transmission line fault location (millisecond-level response after a fault).
[0048] 4. Innovatively provides probabilistic confidence assessment
[0049] This invention utilizes a Bayesian confidence assessment module and Monte Carlo random deactivation technique to perform N independent inferences on the same feature vector, generating a set of fault location distributions and calculating the mean and standard deviation. This mechanism achieves, for the first time, the quantification of uncertainty in fault location results: the sample mean represents the most probable fault location, and the sample standard deviation directly reflects the reliability of the result (the smaller the σ, the higher the confidence), presenting a probabilistic result through a 95% confidence interval (μ±1.96σ). This function provides dispatchers with a basis for decision-making, avoiding misjudgments caused by potential errors from a single deterministic result, and improving the reliability of decisions in engineering applications.
[0050] 5. Adapting to new energy grid integration and complex power grid environments
[0051] With the large-scale grid connection of new energy sources and the increasing complexity of power grid structures, traditional methods are struggling to cope with multi-source disturbances and dynamic operating conditions. This invention, through the integration of physical models and data intelligence, combines the stability of physical mechanisms with the adaptability of neural networks, effectively handling transient characteristic changes brought about by the access of new energy sources, and providing key technical support for the safe and stable operation of modern power systems.
[0052] In summary, this invention, through a three-layer architecture of "physical feature extraction - lightweight network correction - probabilistic confidence assessment," comprehensively solves the bottlenecks of existing technologies in terms of accuracy, robustness, data dependence, interpretability, and computational efficiency, and significantly improves the engineering practical value and decision support capabilities of fault location. Attached Figure Description
[0053] Figure 1 This is a flowchart illustrating the probabilistic fault location method that integrates a physical model and a lightweight neural network according to an embodiment of the present invention. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.
[0055] Please see Figure 1 This invention provides a probabilistic fault location method that integrates a physical model and a lightweight neural network, comprising the following steps:
[0056] S1, Physical Feature Extraction
[0057] The instantaneous three-phase voltage and three-phase current waveform data after a transmission line fault are collected. A rough estimate of the fault location is calculated using at least one transient physical model of the power system. At least one auxiliary physical feature is extracted. The rough estimate and the auxiliary physical feature are combined to construct a standardized physical feature vector V.
[0058] Specifically, the goal is to extract low-dimensional feature vectors representing the core physical laws of faults from raw electrical measurement data. This is achieved by combining multi-source data acquisition with calculations using various physical models to achieve preliminary quantification of the fault state. When a fault occurs on a transmission line, protection devices or monitoring units (such as digital fault recorders, PMUs, etc.) deployed at both ends of the line acquire the instantaneous three-phase voltage at a high sampling rate (e.g., 10 kHz) for at least one cycle (20 ms) after the fault. , , and three-phase current , , The waveform data. This step is performed by the physical feature extraction module, whose core task is to extract low-dimensional feature vectors that can characterize the core physical laws of the fault from these raw, high-dimensional electrical measurement data.
[0059] Preferably, the sub-step of S1 is:
[0060] S1.1: When a fault occurs in a transmission line, the instantaneous three-phase voltage for at least one power frequency cycle (20ms) after the fault is synchronously acquired at a high sampling rate (e.g., 10 kHz) through the protection devices or dedicated monitoring units (e.g., digital fault recorders, PMUs) configured at both ends of the line. , , and three-phase current , , Waveform data;
[0061] S1.2: The physical feature extraction module (A) has at least one built-in power system transient physical model, which is used to calculate the original waveform data collected in S1.1 to obtain a rough estimate of the fault location. ;
[0062] Preferably, based on the transient impedance method: by performing a Fourier transform on the acquired voltage and current waveforms, the phasor of the fundamental frequency component is obtained. , ); and combined with the known impedance per unit length of the line Solve the voltage equation:
[0063] (3)
[0064] Calculate the apparent impedance, and then estimate it. .
[0065] Preferably, based on the transient traveling wave method: a high-pass filter is used to extract the initial traveling wave signal generated at the instant of the fault from the original data, and a precise traveling wave front arrival algorithm (such as wavelet transform or slope change detection) is used to identify the arrival of the traveling wave at this end. and the other end The time is determined according to the two-terminal traveling wave method formula:
[0066] (4)
[0067] in For the total length of the line, The wave speed is given by the formula. .
[0068] S1.3: To comprehensively characterize the fault features, the physical module will also extract at least one auxiliary physical feature to enhance the descriptive power of the feature vector.
[0069] Preferably, transient energy ( It can calculate the integral of transient power within a specific short time window (e.g., 3ms) after a fault, reflecting the impact intensity of the fault.
[0070] Preferably, the initial angle of the fault ( Record the voltage phase angle at the moment the fault occurs, which affects the transient process of the fault.
[0071] Preferably, the transition resistance is roughly estimated ( A specific algorithm can be used to make a preliminary assessment of the transition resistance at the fault point.
[0072] Preferably, the fault type is uniquely encoded (…). Based on the characteristics of sudden changes in phase current, the fault type (such as single-phase grounding, phase-to-phase short circuit or three-phase short circuit) can be identified and represented by a unique thermal code.
[0073] S1.4: Construction of the comprehensive physical feature vector: The preliminary estimate of the fault location obtained in S1.2 is used to construct the comprehensive physical feature vector. The physical features extracted in S1.3 are integrated and standardized to form the final physical feature vector. :
[0074] (5)
[0075] This vector is passed as the output of this step to the next step.
[0076] S2, Lightweight Residual Network Precise Positioning
[0077] The standardized physical feature vector V output from step S1 is input into a pre-trained lightweight residual network model, and the high-precision deterministic fault location is output through inference.
[0078] Specifically, with the goal of nonlinearly correcting the preliminary physical features, a pre-trained lightweight neural network model is used to achieve high-precision deterministic fault location output. This step is performed by the precise localization module, which aims to perform nonlinear correction on the coarse physical features to obtain a high-precision localization result.
[0079] Preferably, the sub-step of S2 is as follows:
[0080] S2.1: The precise positioning module receives the standardized physical feature vector output from S1.4. This serves as the input to its model. This input is not the original waveform, but rather the key features that have been "dimension-reduced" and "condensed" through physical mechanisms.
[0081] S2.2: Model Inference Based on Lightweight Neural Networks: At the core of module (B) is a lightweight neural network model pre-trained on a large amount of offline data. This model is able to learn the complex nonlinear relationship between physical characteristics and the true fault location, and perform inference to output accurate results.
[0082] Preferably, the neural network model is a lightweight residual network (ResNet). Its structural characteristics are as follows:
[0083] Input layer: The number of neurons is precisely matched to the dimension of the input physical feature vector V.
[0084] Hidden layers: Composed of 5 stacked residual blocks. Each residual block contains two fully connected (dense) layers, supplemented by the non-linear activation function LeakyReLU. The key is the introduction of a shortcut connection, which directly adds the block's input to its output, facilitating the network's learning of residual corrections.
[0085] Output layer: Composed of a single neuron, using a linear activation function, it directly outputs a scalar value, i.e., a high-precision deterministic fault location. .
[0086] S2.3: Offline Training and Optimization of the Model: The lightweight ResNet model is trained offline using supervised learning with massive amounts of multi-condition fault data (including different fault locations, fault types, transition resistances, fault initial angles, etc.) generated through the PSCAD professional simulation platform. The training objective is to minimize the model's predictions. The actual fault location set in the simulation The mean squared error (MSE) between the two is used to optimize the model parameters.
[0087] S3, Bayesian confidence assessment
[0088] The Dropout layer of the lightweight residual network model in step S2 is activated using the Monte Carlo random deactivation technique. The standardized physical feature vector V output from step S1 is repeatedly input into the network model with the activated Dropout layer for N forward propagation cycles to obtain a fault location distribution set containing N samples. Statistical analysis is performed on the distribution set to calculate the sample mean μ and sample standard deviation σ. The sample mean μ is used as the final most probable fault location, and the sample standard deviation σ is used as an indicator to quantify the uncertainty or probability confidence of the location result.
[0089] Specifically, aiming to quantify the uncertainty of the location results, this invention combines Monte Carlo random deactivation technology and statistical analysis to achieve a probabilistic assessment and presentation of the fault location. This step, executed by the confidence assessment module, is one of the key innovations of this invention, used to quantify the uncertainty of the location results.
[0090] Preferably, the sub-step of S3 is as follows:
[0091] 3.1: The confidence assessment module uses the same neural network model as the precise location module. During fault location inference, the Dropout layer in the neural network is forcibly activated. This Dropout layer, during each forward propagation, will activate with a preset probability. Randomly set the output of a portion of neurons to zero (i.e., "discard" them).
[0092] S3.2: Using the same physical feature vector V output from S1.4 as input, repeatedly feed it into the neural network with the Dropout layer activated for N=100 independent forward propagations. Since the combination of neurons randomly dropped in each propagation is different, N slightly different fault location results will be generated, thus forming a fault location distribution set containing N samples. :
[0093] (6)
[0094] S3.3: The confidence assessment module performs statistical analysis on the fault location distribution set D obtained in S3.2 and calculates its sample mean. and sample standard deviation The formula for calculating the sample mean is:
[0095] (7)
[0096] The formula for calculating the sample standard deviation is:
[0097] (8)
[0098] Finally, the sample mean The most likely location of the fault was determined. The sample standard deviation... This can be directly used as an indicator to quantify the uncertainty or probability confidence of the positioning result. A smaller value indicates a higher model confidence level. The final result will be presented to power system operators in a probabilistic form, for example: the most likely location is... km, and provides a 95% confidence interval, such as km. For a more intuitive demonstration, this result can be plotted on the monitoring interface as a graph. Centered on, with The width is a normal distribution curve.
[0099] This invention also provides a probabilistic fault location device that integrates a physical model and a lightweight neural network, comprising:
[0100] The physical feature extraction module is used to collect instantaneous three-phase voltage and three-phase current waveform data after a transmission line fault, calculate a rough estimate of the fault location using at least one power system transient physical model, extract at least one auxiliary physical feature, and combine the rough estimate with the auxiliary physical feature to construct a standardized physical feature vector V.
[0101] The lightweight residual network precise localization module is used to input the standardized physical feature vector V into the pre-trained lightweight residual network model and output a high-precision deterministic fault location through inference.
[0102] The confidence assessment module is used to activate the Dropout layer of the lightweight residual network model using Monte Carlo random deactivation technology. The standardized physical feature vector V is repeatedly input into the network model that activates the Dropout layer for N forward propagation to obtain a set of fault location distributions containing N samples. Statistical analysis is performed on the distribution set to calculate the sample mean μ and sample standard deviation σ. The sample mean μ is used as the final most probable fault location, and the sample standard deviation σ is used as an indicator to quantify the uncertainty or probability confidence of the location result.
[0103] Another embodiment of the present invention provides a probabilistic fault location system that integrates a physical model and a lightweight neural network, comprising: a computer-readable storage medium and a processor;
[0104] The computer-readable storage medium is used to store executable instructions;
[0105] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the probabilistic fault location method that integrates a physical model and a lightweight neural network.
[0106] Another embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the probabilistic fault location method that integrates a physical model and a lightweight neural network.
[0107] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0108] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0111] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A probabilistic fault location method integrating a physical model and a lightweight neural network, characterized in that, Includes the following steps: S1, Physical Feature Extraction The instantaneous three-phase voltage and three-phase current waveform data after a transmission line fault are collected. A rough estimate of the fault location is calculated using at least one power system transient physical model. At least one auxiliary physical feature is extracted. The rough estimate and the auxiliary physical feature are combined to construct a standardized physical feature vector V. S2, Lightweight Residual Network Precise Positioning The standardized physical feature vector V output from step S1 is input into a pre-trained lightweight residual network model, and the high-precision deterministic fault location is output through inference. S3, Bayesian confidence assessment The Dropout layer of the lightweight residual network model in step S2 is activated using the Monte Carlo random deactivation technique. The standardized physical feature vector V output from step S1 is repeatedly input into the network model with the activated Dropout layer for N forward propagation cycles to obtain a fault location distribution set containing N samples. Statistical analysis is performed on the distribution set to calculate the sample mean μ and sample standard deviation σ. The sample mean μ is used as the final most probable fault location, and the sample standard deviation σ is used as an indicator to quantify the uncertainty or probability confidence of the location result.
2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.1 Data Acquisition: When a fault occurs in a transmission line, the protection devices or monitoring units deployed at both ends of the line acquire instantaneous three-phase voltage and three-phase current waveform data for at least one cycle after the fault. S1.2 Physical Model Calculation and Preliminary Feature Extraction: Calculate a rough estimate of the fault location based on the transient impedance model or transient traveling wave model. ; S1.3, Auxiliary physical feature extraction: Extract at least one auxiliary physical feature, wherein the auxiliary physical feature includes transient energy within a specific time window after the fault occurs. Voltage phase angle at the moment of fault occurrence The transition resistance estimate obtained by using an algorithm to roughly estimate the transition resistance. And fault type identification based on comparison of phase current abrupt changes and unique thermal coding. ; S1.4 Feature vector construction: The rough estimate of the fault location obtained in step S1.2 above is used to construct the feature vector. Combined with the auxiliary physical features extracted in step S1.3, a standardized physical feature vector V is constructed.
3. The method according to claim 2, characterized in that, Step S1.2 specifically includes: Based on the transient impedance model: performing Fourier transform on the acquired voltage and current waveforms, calculating the phasor of the fundamental frequency component, and estimating a rough estimate of the fault location by solving the voltage equation based on the impedance per unit length of the line. Alternatively, based on the transient traveling wave model: a high-pass filter is used to extract the initial traveling wave signal generated at the instant of the fault from the raw data, and the arrival time of the traveling wave at the local end and the opposite end is identified by the traveling wave front arrival algorithm. A rough estimate of the fault location is then calculated according to the two-end traveling wave method formula. .
4. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1 Model Input: The precise positioning module receives the standardized physical feature vector V as input; S2.2 Network Inference: The core of the precise localization module is a pre-trained, lightweight residual network model, ResNet, which includes: Input layer: The number of neurons in the input layer is equal to the dimension of the physical feature vector V; Hidden layer: The hidden layer consists of 5 stacked residual blocks. Each residual block contains two fully connected layers and one shortcut connection. The activation function of the fully connected layers is LeakyReLU. Output layer: The output layer contains a single neuron, employs a linear activation function, and directly outputs a scalar value as the high-precision deterministic fault location. ; S2.3 Model Training: The lightweight residual network model ResNet is trained using supervised learning with a large amount of fault data under different operating conditions generated on the simulation platform. The training objective is to minimize the model output. The mean square error (MSE) between the actual fault location set in the simulation and the actual fault location.
5. The method according to claim 4, characterized in that, Step S3 includes the following sub-steps: S3.1 Activation of randomness: The confidence assessment module adopts the same or similar neural network structure as the precise positioning module, and when performing positioning calculation, it forcibly activates the Dropout layer of the lightweight residual network model. The Dropout layer randomly discards a portion of neurons with a preset probability during each forward propagation. S3.2, Multiple random inferences: The same standardized physical feature vector V is used as input and repeatedly fed into the neural network with the Dropout layer activated for N forward propagation, so as to obtain a fault location distribution set D containing N slightly different localization results; S3.3 Statistical Analysis and Results Output: Calculate the statistical measure: Perform statistical analysis on the fault location distribution set D and calculate its sample mean. and sample standard deviation Among them, the sample mean The calculation formula is: (1)； Sample standard deviation The calculation formula is: (2)； Results presentation: The mean of the samples As the final most likely location of the failure; the standard deviation of the sample This serves as an indicator to quantify the uncertainty or probability confidence of the positioning result; and the positioning result and its credibility are presented in a visual manner, outputting a probabilistic positioning result including the most likely fault location and confidence interval.
6. The method according to claim 5, characterized in that, The visualization of the location results and their reliability includes plotting a normal distribution curve centered on the sample mean μ and with the sample standard deviation σ as its width.
7. A probabilistic fault location device integrating a physical model and a lightweight neural network, characterized in that, include: The physical feature extraction module is used to collect instantaneous three-phase voltage and three-phase current waveform data after a transmission line fault, calculate a rough estimate of the fault location using at least one power system transient physical model, extract at least one auxiliary physical feature, and combine the rough estimate with the auxiliary physical feature to construct a standardized physical feature vector V. The lightweight residual network precise localization module is used to input the standardized physical feature vector V into the pre-trained lightweight residual network model and output a high-precision deterministic fault location through inference. The confidence assessment module is used to activate the Dropout layer of the lightweight residual network model using Monte Carlo random deactivation technology. The standardized physical feature vector V is repeatedly input into the network model that activates the Dropout layer for N forward propagation to obtain a set of fault location distributions containing N samples. Statistical analysis is performed on the distribution set to calculate the sample mean μ and sample standard deviation σ. The sample mean μ is used as the final most probable fault location, and the sample standard deviation σ is used as an indicator to quantify the uncertainty or probability confidence of the location result.
8. A probabilistic fault location system integrating a physical model and a lightweight neural network, comprising: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the probabilistic fault location method of any one of claims 1-6, which integrates a physical model and a lightweight neural network.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the probabilistic fault location method of any one of claims 1-6, which integrates a physical model and a lightweight neural network.