A fault location system and method for power distribution networks

Abstract

This application relates to the field of smart grid technology and discloses a distribution network fault location system and method. The system includes a causal generation module, an event inversion module, a dynamic pruning module, a causal co-verification module, and a metacognitive analysis module. The method iteratively calls the causal generation module through the event inversion module to solve for the optimal event parameters containing the fault location, and uses the dynamic pruning module to eliminate physically unreasonable search spaces during the iteration. Then, the causal co-verification module verifies the results and calculates the diagnostic confidence. When the confidence is below a threshold, the event model is upgraded to reconstruct complex faults. Furthermore, the system can generate a diagnosticability map offline to identify diagnostic blind spots. This invention, through generative inversion and model upgrade mechanisms, achieves precise fault location and accurate reconstruction of complex accident sequences, and possesses the ability to quantitatively evaluate its own diagnostic capabilities, significantly improving the intelligence level of fault handling.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, specifically to a distribution network fault location system and method. Background Technology

[0002] As the final link in the power system, the distribution network has a complex network structure, widespread line distribution, and variable operating environment, making it a part of the power system with a high failure rate. Therefore, the ability to quickly and accurately locate the fault point after a fault occurs is of paramount importance for shortening power outage time for users, reducing economic losses, and improving power supply reliability.

[0003] To achieve rapid fault location, current research primarily focuses on utilizing physical quantities such as traveling waves and impedance, or employing data-driven artificial intelligence methods. However, these technologies still face common technical bottlenecks in practical applications. Models based on traditional physical mechanisms rely heavily on precise knowledge of power grid topology parameters and meticulous setting of protection parameters for accurate fault location, making them less adaptable to distribution networks with frequently changing network structures.

[0004] Most existing artificial intelligence methods rely on learning from massive fault sample databases to build discriminative models that determine fault characteristics and locations. This paradigm severely limits their diagnostic performance to the completeness of the sample database. For novel or rare fault types not covered by the database, their generalization ability is insufficient, leading to misdiagnosis. Furthermore, existing technologies have limited capabilities in handling complex cascading or concurrent faults, typically providing only a single diagnostic result without the ability to quantitatively assess the physical consistency and reliability of that result. They also lack the ability to globally examine and proactively optimize the system's localization capabilities under specific sensor layouts, resulting in inadequacy when facing complex scenarios and diagnostic blind spots.

[0005] Therefore, this invention proposes a power distribution network fault location system and method to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a power distribution network fault location system and method, which solves the problems that power distribution network fault location technology faces when dealing with complex fault scenarios, such as difficulty in assessing the reliability of diagnostic results, insufficient generalization ability, and lack of proactive recognition and optimization of its own location capabilities.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a power distribution network fault location system, the system comprising:

[0008] The causal generation module uses a pre-trained Fourier neural operator surrogate model to generate a spatiotemporal transient physical field corresponding to the fault event parameters and covering the entire distribution network.

[0009] After receiving the actual fault observation waveform of the distribution network, the event inversion module takes the actual fault observation waveform as the target and iterates by calling the causal generation module to solve the optimal fault event parameters containing fault location information that can optimally generate the actual fault observation waveform.

[0010] The dynamic pruning module analyzes the spatiotemporal transient physical field during the iterative solution process of the event inversion module, and removes unreasonable fault location assumption regions from the subsequent search space of the event inversion module according to preset physical rules.

[0011] After the event inversion module solves for the optimal fault event parameters, the causal co-verification module calls the causal generation module to generate the final spatiotemporal transient physical field. Based on the consistency between the final spatiotemporal transient physical field and the actual fault observation waveform in multiple physical dimensions, the module calculates the final diagnostic confidence associated with the optimal fault location information.

[0012] The metacognitive analysis module runs offline. By injecting virtual fault event parameters and calling the causal generation module, event inversion module, and causal co-verification module to perform closed-loop simulation, it generates a diagnostic map that evaluates the system's fault location capability under the current distribution network sensor layout and identifies diagnostic blind spots.

[0013] Preferably, the causal generation module is used for:

[0014] During pre-training, the Fourier neural operator surrogate model is optimized using a composite loss function to learn the mapping relationship from fault event parameters to spatiotemporal transient physical fields.

[0015] When defining the composite loss function, the Fourier neural operator surrogate model is called to generate the corresponding spatiotemporal transient physical field based on a set of fault event parameters used for training.

[0016] Based on the generated spatiotemporal transient physical field, the data fidelity term that minimizes the difference between the spatiotemporal transient physical field and the results of high-precision simulation software is calculated, as well as the physical residual term that minimizes the degree of violation of the basic physical equations of the distribution network by the spatiotemporal transient physical field is calculated.

[0017] The composite loss function is obtained based on the sum of the data fidelity term and the physical residual term. The formula for calculating the composite loss function is as follows:

[0018] ;

[0019] In the formula, It is a composite loss function; For data fidelity items; For physical residuals; and The weighting coefficients are used to balance the data fidelity term and the physical residual term.

[0020] Preferably, the event inversion module is used for:

[0021] To reverse-engineer the optimal fault event parameters containing fault location information, a method is constructed and iteratively solved by minimizing the objective function;

[0022] When constructing the objective function, the causal generation module is invoked to generate corresponding waveforms based on a set of fault event parameters to be evaluated. Then, based on the generated waveforms and the actual observed fault waveforms, the similarity between the two is calculated using a dynamic time warping algorithm. The formula for calculating the objective function is as follows:

[0023] ;

[0024] In the formula, Let be the objective function. The fault event parameters, which contain fault location information, are to be solved. The number of sensors that received actual fault observation waveforms; For the first Actual fault observation waveforms collected by each sensor; The causal generation module generates data based on fault event parameters. In the Sensor locations The generated waveform; The distance is calculated using the dynamic time warping algorithm; For the first The weights of the observation data from each sensor in the objective function;

[0025] The event inversion module is also used for a solution unit based on the Bayesian optimization algorithm. The solution unit searches in the fault event parameter space with the aim of minimizing the objective function to obtain the optimal fault event parameters with the minimum value.

[0026] Preferably, the dynamic pruning module is used for:

[0027] The qualitative feature analysis unit extracts the waveform polarity and transient frequency components of the spatiotemporal transient physical field generated by the causal generation module during iteration at the sensor location;

[0028] The spatial pruning unit removes the assumed fault location region corresponding to the spatiotemporal transient physical field from the subsequent search space of the event inversion module when there is a fundamental physical contradiction between the waveform polarity or transient frequency component and the actual fault observation waveform.

[0029] Preferably, the causal co-verification module is used for:

[0030] When performing consistency judgments across multiple physical dimensions, these multiple physical dimensions include temporal verification, amplitude verification, and morphological verification.

[0031] The time verification determines whether the arrival time sequence of the disturbance in the actual fault observation waveform is consistent with the propagation time sequence predicted based on the optimal fault location information.

[0032] The amplitude verification determines whether the relative intensity of the actual fault observation waveform is consistent with the attenuation relationship with propagation distance predicted based on the optimal fault location.

[0033] The morphological verification determines whether the waveform characteristics of the actual fault observation waveform are consistent with the morphology predicted based on the optimal fault location propagating along the corresponding line and undergoing reflection and refraction.

[0034] Preferably, the causal co-verification module is further used for:

[0035] When the final diagnostic confidence level is higher than or equal to the preset confidence level threshold, it is determined that the current optimal fault event parameters are sufficient to reasonably explain the received actual fault observation waveform, the diagnostic process ends and the result is output;

[0036] When the final diagnostic confidence level is lower than the preset confidence level threshold, it is determined that the current event model describing a single fault is insufficient to explain the actual fault observation waveform, and the event inversion module is triggered to re-solve for the optimal fault event parameters in the complex event model space describing sequential faults or concurrent faults.

[0037] Preferably, the metacognitive analysis module is used for:

[0038] The virtual diagnostic loop unit calls the causal generation module to generate a virtual observation waveform for each injected virtual fault event parameter. Then, it uses the virtual observation waveform as input and calls the event inversion module and the causal co-verification module to perform a complete diagnostic process to obtain the inferred fault event parameters and the expected diagnostic confidence level.

[0039] The map generation unit takes all the inference results and expected diagnostic confidence obtained by the virtual diagnostic loop unit after traversing all virtual fault event parameters as input, integrates and generates a diagnostic map.

[0040] Preferably, the data structure of the fault event parameters is defined as an extensible event syntax, which is used for:

[0041] A single-event model describing an independent, single failure event;

[0042] A sequence event model that describes a cascading failure scenario in which an initial failure, after a certain time delay, triggers another secondary failure.

[0043] The present invention also provides a method for locating faults in a power distribution network, the method comprising the following steps:

[0044] S1. Using a pre-trained Fourier neural operator proxy model, a spatiotemporal transient physical field corresponding to the input fault event parameters and covering the entire distribution network is generated based on the input fault event parameters.

[0045] S2. After receiving the actual fault observation waveform of the distribution network, take the actual fault observation waveform as the target, and iterate by calling step S1 to solve in reverse the optimal fault event parameters that can best generate the actual fault observation waveform and contain fault location information.

[0046] S3. During the iterative solution process in step S2, the spatiotemporal transient physical field is analyzed, and unreasonable fault location assumption regions are removed from the subsequent search space in step S2 according to preset physical rules.

[0047] S4. After solving the optimal fault event parameters in step S2, call step S1 to generate the final spatiotemporal transient physical field, and calculate the final diagnostic confidence associated with the optimal fault location information based on the consistency between the final spatiotemporal transient physical field and the actual fault observation waveform in multiple physical dimensions.

[0048] S5. Run offline, inject virtual fault event parameters, and call steps S1, S2 and S4 to perform closed-loop simulation to generate a diagnostic map that evaluates the system fault location capability under the current distribution network sensor layout, and identify diagnostic blind spots.

[0049] Preferably, after step S4, the method further includes:

[0050] S4-1: Determine whether the final diagnostic confidence calculated in step S4 is lower than the preset confidence threshold;

[0051] S4-2: If the final diagnostic confidence level is lower than the preset confidence level threshold, the event model describing the fault will be upgraded from a single event model to a sequence event model, and step S2 will be re-executed.

[0052] This invention provides a fault location system and method for power distribution networks. It has the following beneficial effects:

[0053] 1. This invention transforms the fault location problem from a traditional classification and matching model to a generative inversion model based on physical processes by constructing a causal generation module and combining it with an event inversion module for fault solving. It does not rely on a sample library containing massive fault types for training, but instead learns the inherent physical laws of power grid transient processes, exhibiting stronger generalization ability and adaptability for unseen fault types or complex electromagnetic transient phenomena. Simultaneously, the similarity calculation method used in the event inversion process effectively overcomes the problem of inconsistent data acquisition time starting points from various sensors, reduces the system's dependence on high-precision clock synchronization hardware, and further ensures the accuracy of fault location.

[0054] 2. This invention extends the fault location capability from a single fault point to the reconstruction of complex accident sequences by setting up a causal co-verification module and an extensible event model. When a simple single-event model is insufficient to explain all actual fault observation waveforms, the system can make a judgment based on the calculated final diagnostic confidence level and automatically upgrade the event model to a sequence event model for re-solution. This allows the invention not only to locate the initial fault but also to further reconstruct the spatiotemporal location of subsequent cascading faults triggered by the initial fault, thereby providing a complete analysis of the overall accident situation.

[0055] 3. This invention introduces a metacognitive analysis module, enabling a comprehensive assessment and review of the system's positioning capabilities even during offline operation. This module, by injecting virtual fault event parameters and performing closed-loop simulation, proactively generates a diagnostic map covering the entire network. This map visually indicates the system's expected positioning accuracy in different areas and clearly identifies positioning blind spots, providing quantitative decision-making basis for optimizing or supplementing the sensor network, thereby improving the overall reliability of the solution before system deployment.

[0056] 4. This invention improves the efficiency and accuracy of fault location by designing a dynamic pruning module. During the iterative search process of the event inversion module, the dynamic pruning module performs rapid qualitative physical feature analysis on the intermediately generated spatiotemporal transient physical fields. When a fundamental physical contradiction is found between the waveform characteristics generated by a broad assumed fault location region and the actual observed waveform, that region is removed from the subsequent search space. This avoids the optimization algorithm performing ineffective searches in unsolvable regions, thereby accelerating the convergence of the location process towards a more physically plausible solution. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the system functional module structure of the present invention;

[0058] Figure 2This is a flowchart of the online fault diagnosis process of the present invention;

[0059] Figure 3 This is a flowchart of the offline metacognitive analysis and diagnostic map generation process of the present invention;

[0060] Figure 4 This is a schematic diagram of the diagnosticability map of the present invention.

[0061] Among them, 10. Causal generation module; 20. Event inversion module; 30. Dynamic pruning module; 40. Causal synergistic verification module; 50. Metacognitive analysis module; 60. Diagnosticability map; 61. Diagnostic blind spot. Detailed Implementation

[0062] The technical solutions in 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.

[0063] Reference Figure 1 The present invention provides a power distribution network fault location system, which can be deployed on a server or a dedicated edge computing device. The system includes a causal generation module 10, an event inversion module 20, a dynamic pruning module 30, a causal co-verification module 40, and a metacognitive analysis module 50.

[0064] The causal generation module 10 is configured to establish a deterministic mapping between fault event parameters and the network-wide physical response triggered by those fault event parameters. This module employs a pre-trained Fourier neural operator surrogate model, which, upon receiving a set of fault event parameters as input, generates a spatiotemporal transient physical field corresponding to the fault event parameters and covering the entire distribution network as output.

[0065] The event inversion module 20 is connected to the causal generation module 10. After the system receives the actual fault observation waveform collected by one or more sensors in the distribution network, the event inversion module 20 uses the actual fault observation waveform as the target and repeatedly calls the causal generation module 10 to perform iterative calculations to solve for a set of optimal fault event parameters that can optimally generate the actual fault observation waveform and include fault location information.

[0066] The dynamic pruning module 30 is connected to the event inversion module 20 and the causal generation module 10. During the iterative solution process of the event inversion module 20, the dynamic pruning module 30 simultaneously analyzes the spatiotemporal transient physical field generated by the causal generation module 10 during the iteration, and identifies and removes physically unreasonable fault location assumption regions from the subsequent search space of the event inversion module 20 according to preset physical rules.

[0067] The causal co-verification module 40 has its input connected to the output of the event inversion module 20. After the event inversion module 20 solves for the optimal fault event parameters, the causal co-verification module 40 first calls the causal generation module 10 to generate a final spatiotemporal transient physical field using the optimal fault event parameters. Subsequently, based on the consistency between the final spatiotemporal transient physical field and the actual fault observation waveform across multiple physical dimensions, this module calculates a final diagnostic confidence level associated with the localization result.

[0068] The metacognitive analysis module 50, connected to the causal generation module 10, the event inversion module 20, and the causal co-verification module 40, runs offline. This module injects a series of virtual fault event parameters into the system and calls other modules to perform a complete closed-loop simulation to generate parameters for evaluating the system's fault location capability under the current distribution network sensor layout. Figure 4 The diagnosticability map 60 is shown, and diagnostic blind spots 61 are identified within it.

[0069] Reference Figure 2 This invention also provides a method for locating faults in a power distribution network. This method is an online diagnostic process and specifically includes the following steps:

[0070] Step S1: Using the causal generation module 10, the corresponding spatiotemporal transient physical field is generated based on the input fault event parameters.

[0071] In step S2, after receiving the actual fault observation waveform, the event inversion module 20 uses the waveform as the target and iterates by calling step S1 to solve for the optimal fault event parameters containing fault location information.

[0072] In step S3, during the iterative solution process in step S2, the dynamic pruning module 30 analyzes the intermediate spatiotemporal transient physical field and removes unreasonable fault location assumption regions from the subsequent search space in step S2 according to physical rules.

[0073] Step S4: After solving for the optimal fault event parameters in step S2, the causal co-verification module 40 calls step S1 to generate the final spatiotemporal transient physical field and calculates the final diagnostic confidence associated with the location result.

[0074] Step S4-1: Determine whether the final diagnostic confidence level calculated in step S4 is lower than a preset confidence threshold. If the determination result is no, output the optimal fault event parameters containing the optimal fault location as the final diagnostic result, and the process ends. If the determination result is yes, proceed to step S4-2.

[0075] Step S4-2 upgrades the event model used to describe the fault from a single-event model to a sequential event model, and returns to re-execute step S2 to search in a higher-dimensional parameter space.

[0076] Reference Figure 3 In offline mode, the metacognitive analysis module 50 executes metacognitive analysis step S5. This step performs a closed-loop virtual diagnostic process, including steps S1, S2, and S4, for each parameter in a predefined set of virtual fault event parameters, obtaining an inferred fault event parameter and an expected diagnostic confidence level. After traversing all virtual fault event parameters, all results are integrated to generate a result as shown below. Figure 4 Diagnosticity map 60 is shown.

[0077] The systems and methods of this invention can be deployed in computing environments including, but not limited to, servers, workstations, cloud platforms, or embedded computing units. This computing environment includes hardware such as processors, memory, and communication interfaces. The memory stores a computer program that, when executed by the processor, implements the steps of the aforementioned methods. The communication interface is used to interact with distribution network automation systems, wide-area measurement systems (WAMS), or other data acquisition terminals to obtain actual fault observation waveforms.

[0078] The following is about Figures 1 to 4 The specific modules and steps in the system and method shown are described in detail.

[0079] To provide a unified and structured description of single faults and even complex cascading fault scenarios, this invention defines an extensible event grammar to standardize the data structure of fault event parameters. This event grammar serves as the technical foundation for causal generation, event inversion, and model upgrading throughout the system. Specifically, the event grammar includes:

[0080] Single-Event Model: Used to describe a single, independent failure event. Its data structure is a set of parameters containing all the key physical attributes describing the failure. A typical single-event model can be defined as:

[0081] ;

[0082] In the formula, Define the fault location, consisting of a line identifier and the relative distance (e.g., percentage) from the line origin; Define the fault type, such as single-phase grounding, two-phase short circuit, three-phase short circuit, etc. (enumerated values); Define the initial phase angle at which the fault occurs, which affects the initial state of the transient process; Define the transition resistance at the fault point, which affects the magnitude and waveform characteristics of the fault current.

[0083] The Sequential Event Model (SAM) describes a cascading failure scenario where an initial failure triggers another secondary failure after a certain time delay. This model is built upon the single-event model, combining two single-event models with a time parameter. Its data structure can be defined as follows:

[0084] ;

[0085] In the formula, It is a parameter set that follows a single-event model structure and is used to describe the initial fault; For another parameter set that follows the single-event model structure, used to describe the event... The resulting secondary malfunction; A time interval parameter, representing from happened to The time elapsed since the event occurred.

[0086] This extensible event syntax design enables the solution objective of the event inversion module 20 to be parameterized and structured, and provides a clear data structure foundation for the causal co-verification module 40 to automatically upgrade from a single event model to a sequence event model, thereby realizing the reconstruction of the overall picture of complex accidents.

[0087] Reference Figure 1 The detailed implementation of the causal generation module 10 is described below. The function of this module is to construct an efficient and accurate proxy model from fault event parameters to the transient physical field of the entire network, providing a foundation for subsequent event inversion and verification. After pre-training, this module executes step S1 in the system's online diagnostic process.

[0088] The core of the causal generation module 10 is a Fourier neural operator surrogate model. This model structure was chosen because the transient traveling wave propagation process caused by faults in the distribution network can essentially be described by partial differential equations (specifically, telegraph equations). The Fourier neural operator surrogate model, by directly learning and solving operators in the Fourier domain, can handle distribution networks with complex topologies without relying on specific mesh partitioning. Furthermore, its computational efficiency is significantly improved compared to traditional time-domain finite-difference or finite element numerical simulation methods when solving the physical processes defined by partial differential equations. This characteristic enables the model to meet the technical requirements of generating spatiotemporal transient physical fields covering the entire distribution network.

[0089] When pre-training this Fourier neural operator surrogate model, the goal is to learn an accurate mapping from fault event parameters to a spatiotemporal transient physics field. Fault event parameters are a set of structured data describing the causes of fault occurrence, which may include: fault location (defined by line identifier and percentage relative distance from the line head), fault type (e.g., single-phase grounding, phase-to-phase short circuit), initial phase angle of the fault, and fault transition resistance value. The spatiotemporal transient physics field is a data structure used to characterize the changes in physical quantities (e.g., voltage or current) at all discrete points in the distribution network model over time within a preset time window.

[0090] The training process is optimized using a composite loss function. To calculate this composite loss function, firstly, based on a set of fault event parameters used for training, a corresponding spatiotemporal transient physical field is generated by calling a Fourier neural operator surrogate model; then, based on this generated spatiotemporal transient physical field, a data fidelity term and a physical residual term are calculated. Finally, a composite loss function is constructed based on the data fidelity term and the physical residual term, and its calculation formula is as follows:

[0091] ;

[0092] In the formula, It is a composite loss function; For data fidelity items; For physical residuals; and These are the weighting coefficients used to balance the data fidelity term and the physical residual term.

[0093] Data fidelity item The calculation process is as follows: For the same set of training fault event parameters, a spatiotemporal transient physical field is generated by a Fourier neural operator surrogate model, and a reference spatiotemporal transient physical field is generated by high-precision electromagnetic transient simulation software (such as PSCAD or EMTP). The data fidelity term is obtained by calculating the norm difference (e.g., mean square error) between these two physical fields. The purpose of this term is to ensure the numerical accuracy of the model's generated results.

[0094] Physical residuals The calculation process is as follows: The spatiotemporal transient physical field generated by the Fourier neural operator surrogate model is substituted into the discrete form of the fundamental physical equations describing the transient processes of the distribution network, such as Kirchhoff's current law at nodes and the telegraph equations on branches. Since the solution generated by the model is an approximate solution, the difference between the two sides of the equation after substitution is not zero; this difference is the physical residual. The physical residual term is obtained by calculating the norm of this physical residual over the entire spatiotemporal domain. The purpose of this term is to ensure that the physical field output by the model still follows the basic physical laws even in parameter regions not covered by the training data, thereby improving the model's generalization ability.

[0095] Reference Figure 1 and Figure 2 The detailed implementation of the event inversion module 20 is described below. This module executes step S2 in the system's online diagnostic process, and its core task is to solve an inverse problem:

[0096] After receiving the actual fault observation waveform collected by the distribution network sensor, the cause of the result is deduced and solved in reverse, namely, a set of optimal fault event parameters containing fault location information.

[0097] To achieve this reverse engineering, the event inversion module 20 first constructs an objective function and minimizes its value through iterative calculation. When constructing the objective function, the module calls the causal generation module 10 to generate a set of corresponding transient waveforms based on a set of fault event parameters to be evaluated. Subsequently, the module calculates the similarity between the generated waveforms and the actual fault observation waveforms. The similarity calculation here employs a dynamic time warping algorithm. This algorithm is chosen because there are slight deviations in the start times of data acquisition between different sensors, and the waveform shape of transient traveling waves undergoes stretching or compression to some extent during propagation. The dynamic time warping algorithm, by calculating the optimal matching path between two time series, can measure their morphological similarity, rather than a strict point-to-point Euclidean distance, thereby reducing the system's requirement for high-precision, strict clock synchronization of all sensors.

[0098] The specific formula for calculating the objective function is as follows:

[0099] ;

[0100] In the formula, Let be the objective function, whose value represents the parameters of the fault event to be evaluated. The degree of dissimilarity between the generated waveform and the actual observed waveform, The fault event parameters, which contain fault location information, are to be solved. The number of sensors that received actual fault observation waveforms; For the first Actual fault observation waveforms collected by each sensor; The causal generation module generates data based on fault event parameters. In the Sensor locations The generated waveform; The distance is calculated using the dynamic time warping algorithm; For the first The weights of the observation data from each sensor in the objective function can be preset based on the sensor's signal-to-noise ratio, data quality, or its importance in the power grid topology.

[0101] Event inversion module 20 also includes a solution unit for searching the fault event parameter space to find the objective function. Optimal fault event parameters that yield the minimum value The solution unit employs the Bayesian optimization algorithm. This algorithm was chosen because the objective function... Each evaluation requires a call to the causal generation module 10, resulting in high computational costs and classifying it as a typical expensive black-box function optimization problem. The Bayesian optimization algorithm, by constructing a probabilistic surrogate model of the objective function and using the sampling function to guide the next sampling location, can find the global optimum in fewer iterations, thus improving the overall computational efficiency of the event inversion process while maintaining solution accuracy.

[0102] Reference Figure 1 and Figure 2 The detailed implementation of the dynamic pruning module 30 is described below. When solving for optimal fault event parameters, the event inversion module 20 has a high-dimensional and wide-ranging search space. Unconstrained global optimization would lead to high computational resource consumption and the solution process would easily converge to physically invalid local optima. The dynamic pruning module 30 is designed to address this problem. This module runs in parallel with the iterative solution process of the event inversion module 20, executing step S3. It utilizes prior physical knowledge to remove unreasonable fault location assumptions from the subsequent search space, thereby improving solution efficiency and the accuracy of the results.

[0103] The dynamic pruning module 30 includes a qualitative feature analysis unit and a spatial pruning unit. The qualitative feature analysis unit is used to extract the waveform qualitative features of the spatiotemporal transient physical field generated by the causal generation module 10 during iteration at various sensor locations, such as the initial polarity and transient frequency components of transient voltage or current.

[0104] Taking waveform polarity as an example, for a single-phase ground fault in a distribution network, based on traveling wave theory, the initial polarity of the transient traveling wave current detected by all measurement points upstream of the fault point is the same, while the initial polarity of the transient traveling wave current detected by all measurement points downstream of the fault point is opposite to that upstream. In a certain iteration of the event inversion module 20, if the fault event parameters to be evaluated include a hypothetical fault location, the qualitative feature analysis unit first extracts the initial polarity of the waveform generated by the hypothesis at each sensor location and compares it with the initial polarity of the actual fault observation waveform. If it is found that the generated waveform polarity distribution contradicts this physical law, the spatial pruning unit determines that the hypothetical fault location is invalid and removes the location and a section of the line adjacent to it from the subsequent search space of the event inversion module 20.

[0105] Taking transient frequency components as an example, due to the distributed capacitance effect, the transient traveling waves propagating on cable lines in distribution networks typically contain richer transient high-frequency components than those on overhead lines. If, after analysis, the actual fault observation waveform shows that its main energy is concentrated in the lower frequency band, and the fault location currently being evaluated by the event inversion module 20 is assumed to be located at or adjacent to a section of cable line, then there is a physical contradiction between this assumption and the observation. Based on this, the spatial pruning unit determines to remove the fault location region containing that section of cable line from the subsequent search space. In this way, the dynamic pruning module 30 introduces physical constraints into the mathematical optimization process, avoiding invalid iterative calculations by the solution algorithm in unsolvable regions.

[0106] Reference Figure 1 and Figure 2 The detailed implementation of the causal co-verification module 40 is described below. After the event inversion module 20 solves for the optimal fault event parameters, traditional diagnostic methods typically output this as the final result, but cannot evaluate the reliability of the location result. The causal co-verification module 40 is designed to solve this technical problem. It is used in step S4 to verify the physical self-consistency of the solution output by the event inversion module 20 and calculate a quantified final diagnostic confidence level to reflect the reliability of the location result.

[0107] When performing consistency checks, this module verifies the final spatiotemporal transient physical field generated from the optimal fault event parameters against the actual fault observation waveform from multiple physical dimensions. Specific verification dimensions include:

[0108] Time Verification: Based on the optimal fault location calculated by the event inversion module 20, combined with pre-stored distribution network topology data (including the length and connection relationships of each line) and the propagation speed of transient traveling waves in the corresponding medium, the theoretical propagation time from the fault location to each sensor installation location is calculated. The theoretical propagation times of all sensors are sorted to obtain a theoretical disturbance arrival time series. This theoretical time series is then compared with the arrival time series of disturbance signals actually detected by each sensor, extracted from the actual fault observation waveform. The higher the degree of consistency between the two, the higher the reliability of the solution.

[0109] Amplitude Verification: The amplitude of a transient traveling wave attenuates as the propagation distance increases along the line. Based on the calculated optimal fault location and the distribution of sensors on the line, the relative intensity relationship that the transient waveform amplitudes observed at each sensor should conform to can be predicted. This theoretical relative amplitude relationship is then compared with the actual relative amplitude relationships of the waveforms measured from actual fault observations. The higher the degree of consistency between the two, the higher the reliability of the solution.

[0110] Morphological Verification: When a transient traveling wave encounters impedance discontinuities such as line branches, ends, or connections between different types of cables during its propagation path, refraction and reflection occur. These phenomena are superimposed on the original traveling wave, causing the waveform observed at the sensor to exhibit specific morphological characteristics, such as oscillation or attenuation. The causal co-verification module 40 determines whether the morphological characteristics of the actual fault observation waveform match the waveform predicted based on the optimal fault location, after propagation along the corresponding line and reflection at impedance discontinuities. The higher the degree of consistency between the two, the higher the reliability of the solution.

[0111] The causal co-verification module 40 calculates a final diagnostic confidence value based on the consistency across multiple physical dimensions.

[0112] Reference Figure 2 Following step S4, the module further executes step S4-1, comparing the calculated final diagnostic confidence level with a preset confidence threshold. If the diagnostic confidence level is higher than or equal to the threshold, it is considered that the current optimal fault event parameters are sufficient to reasonably explain the received actual fault observation waveform, the diagnostic process ends, and the result is output.

[0113] If the diagnostic confidence level is lower than the threshold, step S4-2 is executed. At this point, the module determines that the currently used single-event model describing a single fault is insufficient to explain complex observed waveforms. Subsequently, the module upgrades the event model used to describe the fault from a single-event model to a complex event model, such as a sequence event model defined in the system to describe a fault triggering another fault after a specific time interval. The parameter structure of this sequence event model can be defined as containing two sets of independent fault event parameters and a time interval parameter. After the model upgrade, the module triggers the event inversion module 20 to return to and re-execute step S2, performing a new round of search and solution in the higher-dimensional parameter space defined by the sequence event model, thereby achieving the analysis and localization of complex accident scenarios such as cascading faults or concurrent faults.

[0114] Reference Figure 1 , Figure 3 and Figure 4 The detailed implementation of the metacognitive analysis module 50 is described below. This module runs in the offline state of the system and executes step S5. Its function is to quantitatively evaluate the overall fault location capability of the current system under a specific power distribution network topology and sensor layout, and identify areas with weak or uncovered location capabilities (i.e., location blind spots).

[0115] The metacognitive analysis module 50 includes a virtual diagnostic loop unit and a map generation unit. The module's operation flow is as follows:

[0116] First, in offline mode, the established digital model of the distribution network is discretized into a series of virtual fault points along all its lines at a preset spatial step size (e.g., every 50 meters).

[0117] Subsequently, the virtual diagnostic loop unit injects one or more preset virtual fault event parameters representing typical fault types for each virtual fault point. For each injected virtual fault event parameter, the unit executes a complete closed-loop virtual diagnostic process once:

[0118] First, the causal generation module 10 is called, and a set of virtual observation waveforms at the existing sensor layout locations in the power distribution network are generated using the virtual fault event parameters as input.

[0119] Next, the virtual observation waveform is used as input, and the event inversion module 20 and the causal co-verification module 40 are invoked to execute a complete online diagnostic process including steps S2 to S4.

[0120] The output of this process, namely an inferred fault event parameter and a corresponding expected diagnostic confidence level, is recorded and associated with the injected virtual fault point.

[0121] After the virtual diagnostic loop unit has traversed all preset virtual fault points and virtual fault event parameters, the map generation unit integrates all recorded expected diagnostic confidence data. Using the distribution network topology as a background, this unit visualizes the expected diagnostic confidence value for each location on the map using a visualization method (e.g., a color-coded heatmap), thereby generating a map like... Figure 4 Diagnosticity map 60 is shown.

[0122] In this diagnostic capability map 60, different colors or grayscale values ​​represent the expected diagnostic confidence level when a system failure occurs at that location. Areas with expected diagnostic confidence significantly lower than a preset threshold are clearly marked as diagnostic blind zones 61. This diagnostic capability map provides planners with intuitive and quantitative assessments of the sensor network's diagnostic capabilities, offering direct technical decision-making support for subsequent sensor optimization, adding sensors, or adjusting installation locations. Its purpose is to eliminate diagnostic blind zones and improve the overall system's positioning coverage and reliability.

[0123] To more clearly illustrate the collaborative workflow of the system and method provided by this invention, a hypothetical fault scenario will be used as an example below.

[0124] Reference Figure 1 and Figure 2 Suppose a single-phase ground fault occurs on a line in a power distribution network. Multiple sensors installed in the power grid capture the transient traveling wave caused by this fault and transmit the collected actual fault observation waveform data to the system. After receiving this data, the event inversion module 20 initiates the iterative solution process in step S2, searching for the optimal fault event parameters containing the fault location within a preset single-event model parameter space. In the early stages of iteration, when the hypothetical fault location evaluated by the event inversion module 20 deviates significantly from the actual location, the dynamic pruning module 30 executes step S3 in parallel. By comparing the initial polarity of the intermediate waveform generated by the causal generation module 10 with the initial polarity of the actual observed waveform, it quickly removes large areas of the line that are physically invalid from the subsequent search space. Subsequently, the event inversion module 20 continues to search within the reduced space using a Bayesian optimization algorithm and finally converges to obtain a set of optimal fault event parameters containing the optimal fault location.

[0125] Subsequently, the process proceeds to step S4, where the causal co-verification module 40 receives the set of parameters. The module first calls the causal generation module 10 to generate the final spatiotemporal transient physical field, and then verifies it against the actual observed waveform from three dimensions: time, amplitude, and morphology, calculating the final diagnostic confidence level. If this confidence level is higher than a preset threshold, the system outputs the optimal fault event parameters, which contain the optimal fault location, as the final diagnostic result.

[0126] In another complex scenario, a secondary fault is induced at another location within a very short time after the initial fault. In this case, any set of optimal parameters calculated by the event inversion module 20 based on the single-event model, containing the single fault location, cannot simultaneously satisfy the physical self-consistency requirements of all sensor-observed waveforms under the verification of the causal co-verification module 40. This results in the calculated final diagnostic confidence level being lower than a preset threshold. Accordingly, the causal co-verification module 40 executes step S4-2, upgrading the event model from a single-event model to a sequence event model, and triggers the event inversion module 20 to return to step S2. The event inversion module 20 then re-searches and solves in the higher-dimensional sequence event parameter space, ultimately outputting a diagnostic result that reconstructs the cascading fault process, containing the two fault locations, fault types, and time intervals.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power distribution network fault location system characterized by, The system comprises: a causal generation module, which uses a pre-trained Fourier neural operator proxy model to input fault event parameters to generate a corresponding spatio-temporal transient physical field covering the entire power distribution network; an event inversion module, which, after receiving an actual fault observation waveform of the power distribution network, takes the actual fault observation waveform as a target, iteratively calls the causal generation module, and reversely solves optimal fault event parameters containing fault location information that can optimally generate the actual fault observation waveform; a dynamic pruning module, which analyzes the spatio-temporal transient physical field during the iterative solving process of the event inversion module, and removes unreasonable fault location hypothesis areas from the subsequent search space of the event inversion module according to pre-set physical rules; wherein the dynamic pruning module is configured to: a qualitative feature analysis unit configured to extract waveform polarity and transient frequency components of the spatio-temporal transient physical field generated by the causal generation module in iteration at sensor locations; a spatial pruning unit configured to remove the fault location hypothesis area corresponding to the spatio-temporal transient physical field from the subsequent search space of the event inversion module when there is a fundamental physical contradiction between the waveform polarity or transient frequency components and the actual fault observation waveform; a causal collaborative verification module, which, after the event inversion module solves the optimal fault event parameters, calls the causal generation module to generate a final spatio-temporal transient physical field, and calculates a final diagnostic confidence associated with the optimal fault location information according to the consistency between the final spatio-temporal transient physical field and the actual fault observation waveform in multiple physical dimensions; a meta-cognition analysis module, which runs in an offline state, injects virtual fault event parameters, and calls the causal generation module, the event inversion module, and the causal collaborative verification module for closed-loop simulation to generate a diagnosability map evaluating the fault location capability of the system under the current sensor layout of the power distribution network, and identify diagnostic blind areas.

2. The power distribution network fault location system of claim 1, wherein, The causal generation module is configured to: optimize the Fourier neural operator proxy model through a composite loss function during pre-training to learn the mapping relationship from fault event parameters to spatio-temporal transient physical fields; when defining the composite loss function, call the Fourier neural operator proxy model to generate corresponding spatio-temporal transient physical fields based on a set of training fault event parameters; based on the generated spatio-temporal transient physical fields, calculate a data fidelity term that minimizes the difference between the spatio-temporal transient physical fields and the results of high-precision simulation software, and a physical residual term that minimizes the degree of violation of the spatio-temporal transient physical fields to the basic physical equations of the power distribution network; obtain the composite loss function based on the sum of the data fidelity term and the physical residual term, and the calculation formula of the composite loss function is: ； wherein is a composite loss function; is a data fidelity term; is a physical residual term; and is a weight coefficient balancing the data fidelity term and the physical residual term.

3. The power distribution network fault location system of claim 1, wherein, The event inversion module is configured to: for reversely solving the optimal fault event parameters containing fault location information, build and iteratively solve in a way of minimizing an objective function; In constructing the objective function, the causal generation module is called to generate a corresponding waveform according to a set of fault event parameters to be evaluated, and a similarity between the generated waveform and an actual fault observation waveform is calculated based on a dynamic time warping algorithm, and a calculation formula of the objective function is: ； wherein, is the objective function, is the fault event parameter to be solved, which contains the fault location information; is the number of sensors receiving the actual fault observation waveform; is the actual fault observation waveform collected by the sensor; is the waveform generated by the causal generation module according to the fault event parameter at the sensor location ; is the distance calculated by the dynamic time warping algorithm; is the weight of the observation data of the sensor in the objective function; The event inversion module is also used for a solving unit based on a Bayesian optimization algorithm, and the solving unit searches in a fault event parameter space for optimal fault event parameters with a minimum value for the purpose of minimizing the objective function.

4. The power distribution network fault location system of claim 1, wherein, The causal collaborative corroboration module is used for: When consistency in multiple physical dimensions is judged, the multiple physical dimensions include time corroboration, amplitude corroboration, and shape corroboration, wherein, The time corroboration judges whether a disturbance arrival time sequence of the actual fault observation waveform is consistent with a propagation time sequence predicted according to optimal fault location information; The amplitude corroboration judges whether a relative intensity of the actual fault observation waveform is consistent with a relationship of attenuation with propagation distance predicted according to the optimal fault location; The shape corroboration judges whether a waveform feature of the actual fault observation waveform is consistent with a shape predicted according to the optimal fault location after propagation on a corresponding line and passing through refraction and reflection.

5. The power distribution network fault location system of claim 1, wherein, The causal collaborative corroboration module is also used for: When a final diagnostic confidence is higher than or equal to a preset confidence threshold, it is judged that the current optimal fault event parameters are sufficient to reasonably explain the received actual fault observation waveform, and a diagnosis process is ended and a result is output; When the final diagnostic confidence is lower than the preset confidence threshold, it is judged that the event model describing a single fault is insufficient to explain the actual fault observation waveform, and the event inversion module is triggered to solve the optimal fault event parameters in a complex event model space describing a sequence fault or a concurrent fault again.

6. The power distribution network fault location system of claim 1, wherein, The metacognition analysis module is used for: A virtual diagnosis circulation unit is used to, for each injected virtual fault event parameter, call the causal generation module to generate a virtual observation waveform, and then call the event inversion module and the causal collaborative corroboration module to perform a complete diagnosis process once with the virtual observation waveform as input, to obtain inferred fault event parameters and expected diagnostic confidence; A map generation unit is used to, with all the inference results and expected diagnostic confidence obtained after the virtual diagnosis circulation unit traverses all the virtual fault event parameters as input, integrate and generate a diagnosability map.

7. A power distribution network fault location system according to claim 6, characterised in that, The data structure of the fault event parameters is defined as an extensible event syntax, and the event syntax includes: A single event model for describing an independent single fault event; And a sequence event model for describing a cascading fault scenario in which an initial fault triggers another secondary fault after a certain time delay.

8. A method for fault location in a power distribution network, applied to the system of any of claims 1-7, characterized by, The method includes the following steps: S1, using a pre-trained Fourier neural operator proxy model, generating a time-space transient physical field covering the entire power distribution network according to input fault event parameters; S2, after receiving the actual fault observation waveform of the power distribution network, iteratively calling step S1 to inversely solve the optimal fault event parameters containing the fault location information which can optimally generate the actual fault observation waveform, with the actual fault observation waveform as the target; S3, in the iterative solving process of step S2, analyzing the time-space transient physical field, and removing unreasonable fault location hypothesis regions from the subsequent search space of step S2 according to the preset physical rules; S4, after solving the optimal fault event parameters in step S2, calling step S1 to generate the final time-space transient physical field, and calculating the final diagnostic confidence associated with the optimal fault location information according to the consistency between the final time-space transient physical field and the actual fault observation waveform in multiple physical dimensions; S5, in the offline state, by injecting virtual fault event parameters, calling steps S1, S2 and S4 to perform closed-loop simulation, to generate a diagnosability map to evaluate the system fault location capability under the current power distribution network sensor layout, and identify the diagnostic blind area.

9. The method of claim 8, wherein, After step S4, it further includes: S4-1: judging whether the final diagnostic confidence calculated in step S4 is lower than the preset confidence threshold; S4-2: if the final diagnostic confidence is lower than the preset confidence threshold, upgrading the event model describing the fault from a single event model to a sequence event model, and returning to re-execute step S2.

Images