A multi-source perception power grid fault accurate positioning intelligent detection system

Abstract

The application relates to the technical field of power grid fault detection, and discloses a multi-source sensing power grid fault accurate positioning intelligent detection system. The system realizes comprehensive sensing of multi-source information by collecting electrical quantities, images, environment and text data in real time through a physical space sensing module; a data synchronous mapping module aligns the multi-source data in time and space and maps the data to a digital twin space, ensuring the time and space consistency of the data; a digital twin space module constructs a multi-scale power grid twin model, realizes accurate positioning of faults in complex scenes through cross-modal attention fusion features and a hypothesis-inference-verification mechanism, and a decision execution module is internally provided with a large language model and a power fault knowledge graph, generates a diagnosis conclusion and a disposal suggestion through causal reasoning, and automatically issues a control instruction. The system forms an intelligent detection closed loop of sensing, mapping, reasoning and decision making, and significantly improves the accuracy, environmental adaptability and automation level of power grid fault detection.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and in particular to a multi-source sensing intelligent detection system for precise location of power grid faults. Background Technology

[0002] Existing technologies mainly revolve around power grid fault location and monitoring. They collect electrical quantity data of the power grid by deploying electrical quantity sensors and other equipment, and analyze data characteristics by combining state estimation, Clarke transform and other algorithms to narrow down the fault candidate area and thus achieve fault segment location. Some technologies introduce image monitoring, signal waveform detection and other means to help identify equipment abnormalities or eliminate signal interference, providing data support for fault judgment. Overall, they form a power grid fault detection and location technology system that combines single or a few types of data collection with traditional algorithm analysis.

[0003] Regarding the aforementioned and existing related technologies, the inventors believe that the following shortcomings often exist: existing technologies mostly rely on single or a few types of data as fault criteria, and the data fusion dimensions are limited, making it difficult to fully reflect complex power grid fault scenarios; positioning accuracy is affected by data quality and algorithm limitations, and positioning errors are prone to occur in complex power grid scenarios; there is a lack of adaptive learning and dynamic update mechanisms, resulting in insufficient adaptability to new faults and complex operating conditions, and the level of intelligence in fault diagnosis and handling needs to be improved.

[0004] Publication No. CN116482475A discloses a fault location method for active distribution networks based on multiple criteria. It narrows down the fault candidate region by using state estimation residual analysis and Clarke transform as dual criteria, which solves the problem of accuracy in fault location of active distribution networks. However, it does not solve the problems of limited data fusion dimensions, lack of adaptive learning mechanism, and insufficient intelligence in fault diagnosis and handling.

[0005] Publication number CN117214617B discloses a real-time fault monitoring and location system and method for smart grids. It achieves real-time fault monitoring and location through image preprocessing and feature analysis, which solves the problem of low efficiency of manual monitoring. However, it does not solve the problems of deep fusion of multi-source heterogeneous data, improvement of positioning accuracy in complex power grid scenarios, and dynamic updating of models and knowledge bases. Summary of the Invention

[0006] The technical problem to be solved by this invention is that existing power grid fault detection technologies suffer from insufficient multi-source information fusion and low level of intelligent positioning. Specifically, reliance on a single data source leads to poor positioning accuracy in complex scenarios, and reliance on human experience results in low diagnostic efficiency and an inability to learn adaptively. To address this, we propose a multi-source sensing intelligent detection system for precise positioning of power grid faults.

[0007] To achieve the above objectives, this application adopts the following technical solution: a multi-source sensing intelligent detection system for precise location of power grid faults, comprising: a physical space sensing module deployed on the physical side of the power grid, including various types of sensors and data interfaces, for real-time acquisition of multi-dimensional heterogeneous data of power grid operation; a data synchronization mapping module for performing spatiotemporal alignment processing on the multi-dimensional heterogeneous data and mapping it to a digital twin space; a digital twin space module including a multi-scale power grid twin model, a cross-modal feature association and inference unit, and a fault inference and location unit; the multi-scale power grid twin model is used to characterize the physical characteristics and operating behavior of the power grid; the cross-modal feature association and inference unit is used to receive the mapped multi-source data and use a bidirectional cross-modal attention mechanism to perform feature interaction and fusion on the multi-source data, outputting fused features; the bidirectional cross-modal attention mechanism uses electrical quantity features as queries to perform attention calculation on image features and text features, and simultaneously uses image features as queries to perform reverse attention calculation on electrical quantity features and text features. This system enables deep interactive fusion of multi-source heterogeneous data. The fault inference and location unit generates candidate fault sections based on the fusion characteristics using state estimation residual analysis. In the multi-scale power grid twin model, hypothetical faults are set for each candidate fault section, and electromagnetic transient simulations are performed. The simulated waveforms are compared with real-time acquired electrical quantity data to calculate waveform matching degree. This waveform matching degree calculation combines dynamic time warping algorithm and cross-correlation analysis to determine the precise fault point based on the matching degree. The decision execution module includes a fault diagnosis and handling unit. This unit has a built-in large language model and performs causal reasoning based on the fusion characteristics and the precise fault point, combined with a pre-constructed power fault knowledge graph. It generates fault cause analysis and handling suggestions, and outputs a diagnostic report or issues control commands. The digital twin space module also feeds back the data and results of each fault handling process to the multi-scale power grid twin model and the power fault knowledge graph, enabling iterative evolution of model parameters and the knowledge base.

[0008] Preferably, the physical space perception module constructs a multi-dimensional holographic perception network covering electrical, visual, environmental, and semantic dimensions. The electrical quantity sensors are deployed at key nodes of the power grid and employ a high sampling frequency capable of fully recording fault traveling waves. The visual sensors, controlled by a gimbal, capture images of key equipment and lines from a top-down angle and simultaneously acquire visible light and infrared images. The environmental sensors monitor micro-meteorological data in real time. The text data interface connects to the power dispatch data network to obtain unstructured text data. The high sampling frequency refers to the sampling rate required to capture high-frequency components during fault transients.

[0009] Preferably, the data synchronization mapping module includes: a clock synchronization unit, which uses the second pulse of a satellite navigation system to synchronize the clocks of each sensor; an abnormal data preprocessing unit, which uses a statistical distribution-based outlier detection algorithm to identify outliers in electrical quantity data, and uses a sequence learning model to perform context completion and repair on the identified abnormal data; the statistical distribution-based outlier detection algorithm refers to an algorithm that uses data distribution characteristics to identify sampling points that deviate from the normal pattern; the sequence learning model refers to a machine learning model that uses temporal context information to predict missing or abnormal data; and an interpolation alignment unit, which uses an interpolation method to uniformly interpolate multi-source data with different sampling rates to a high-precision time axis, and extracts feature vectors from image data before performing time alignment to generate a spatiotemporally consistent multimodal fusion data stream; the interpolation method refers to a data reconstruction method that maps discrete sampling points to continuous time coordinates to generate a unified temporal sequence.

[0010] Preferably, the cross-modal feature association and inference unit employs a bidirectional cross-modal attention mechanism to perform feature interaction and fusion on multidimensional heterogeneous data. The bidirectional cross-modal attention mechanism uses electrical quantity features as queries to perform attention calculations on image features and text features, and simultaneously uses image features as queries to perform reverse attention calculations on electrical quantity features and text features. After fusing the results of multiple attention calculations, a fused feature vector of a preset dimension is output. The electrical quantity features are obtained by time-series feature extraction and frequency domain decomposition, the image features are obtained by target detection and the attention mechanism, and the text features are obtained by a pre-trained language model in the power field. The time-series feature extraction refers to the process of extracting time-domain statistical features from electrical quantity waveforms, and the frequency domain decomposition refers to the method of converting time-domain signals into frequency-domain components to obtain frequency band energy distribution.

[0011] Preferably, the multi-scale power grid twin model uses a graph attention network to dynamically model the power grid topology, with power grid nodes as vertices and lines as edges. Node features include fused features and physical features. The graph attention network employs a multi-head attention mechanism, where each head can be configured with different scale parameters or electrical distance definitions. The features output by multiple heads are concatenated to obtain the final fused feature containing spatial topology information. The attention coefficients between nodes are calculated using the following formula: ,in, For nodes The multi-source fusion feature vector is composed of the fusion features output by the cross-modal feature association and inference unit. This vector integrates information from multiple sources such as electrical quantities, images, environment, and text. Nodes provided for digital twin models With nodes The electrical distance between them, which is calculated based on line impedance parameters, topology and real-time operating status, reflects the physical possibility of fault propagation in the power grid; This is a scale parameter used to adjust the intensity of the influence of physical distance on attention weights, and can be set according to the scale of the power grid or experience. For nodes The set of neighboring nodes; the graph attention network adopts a multi-head attention mechanism, where each head can be configured with different scale parameters. Alternatively, different electrical distance definitions can be used to concatenate the features output by multiple heads to obtain high-dimensional node features, which are then associated with the cross-modal features and fused with the fusion features output by the inference unit to obtain the final fusion features containing spatial topology information.

[0012] Preferably, the fault simulation and location unit includes: a candidate region generation subunit, which uses state estimation residual analysis to calculate the residuals between the measured values ​​of each node and the estimated values ​​of the twin model, and identifies nodes and their adjacent lines with residuals exceeding a preset threshold as candidate fault sections. This preset threshold is determined based on historical data statistics and is used to quickly narrow down the location range; a rapid simulation subunit, which sets hypothetical fault points in the candidate sections within the multi-scale power grid twin model with a preset step size, and uses model reduction techniques to reduce the full-order electromagnetic transient model to a low-order equivalent model for rapid electromagnetic transient simulation. The model reduction techniques include vector matching or balanced truncation methods, with the simulation time controlled in milliseconds, generating simulated voltage and current waveforms for each node; and a waveform matching subunit, which calculates the matching degree between the simulated waveform and the measured waveform. Before calculating the matching degree, adaptive time windows are used to truncate the simulated and measured waveforms, and the time window is determined based on a fault initiation point detection algorithm. The fault initiation point detection algorithm includes wavelet transform modulus maxima or Hilbert-Huang transform, and the matching degree is calculated using a combination of dynamic time warping algorithm and cross-correlation analysis. The calculation formula is: ,in, For the simulation waveform sequence, This is a measured waveform sequence. For dynamic time-normalized distance, The maximum value of DTW distance across all simulation sequences. It is a cross-correlation function. , For signal power, The weighting coefficient ranges from 0.5 to 0.8; the fault point with the highest matching degree is selected as the final accurate fault point.

[0013] Furthermore, for multi-terminal power supply scenarios, the waveform matching unit adopts a two-layer adaptive weighted voting mechanism to comprehensively determine the location of the fault point. The first layer of weights is based on adaptive allocation of electrical distance, using a function inversely proportional to the electrical distance. The second layer of weights is based on dynamic evaluation of data quality, using a confidence evaluation unit with data quality parameters as input and output confidence coefficients. The two layers of weights are then fused to perform a weighted average of the matching degree of each terminal. When the confidence degree of a certain terminal is lower than a preset threshold, that terminal is automatically removed from the voting set. For the remaining terminals, uncertainty estimation is introduced to calculate the posterior probability. The uncertainty estimation refers to the method of calculating the distribution of the probability of each candidate position through a probability model, and the position with the highest posterior probability is selected as the final fault point.

[0014] Preferably, the large language model built into the fault diagnosis and handling unit is a generative pre-trained model based on power domain corpus with fine-tuned instructions. The power domain corpus includes historical fault reports, equipment manuals, and dispatching procedures, enabling the model to master power professional knowledge and fault causal logic. The power fault knowledge graph is built based on a graph database and includes equipment nodes, fault type nodes, cause nodes, handling measure nodes, and the relationships between them. It stores historical fault cases, forming a structured experience knowledge base that supports fast retrieval and analogical reasoning.

[0015] Furthermore, the fault diagnosis and handling unit recalls similar cases from the knowledge graph as context input to the large language model through graph retrieval enhancement generation. The graph retrieval enhancement generation method refers to the method of retrieving subgraphs related to the current features from the knowledge graph and converting them into natural language prompts. The large language model combines fused features, fault location and retrieved cases to perform causal reasoning, generate fault cause analysis, confidence assessment and handling suggestions, and automatically generate control commands based on the handling suggestions and issue them to the execution agency.

[0016] Preferably, the digital twin space module is further used to collaboratively update the multi-scale power grid twin model and the power fault knowledge graph by using waveform data, location results, diagnostic conclusions, and handling effects generated during each fault handling process as new samples: wherein, for updating the multi-scale power grid twin model, a regularized incremental learning algorithm is adopted, which updates the network parameters by applying penalty terms to important parameters of the model; the important parameters refer to the weight parameters that contribute significantly to the model output, determined by the Fisher information matrix or other importance assessment methods; the regularized incremental learning algorithm is an optimization method that adds parameter change penalties to the loss function to balance the new and old knowledge; at the same time, an adversarial training strategy is adopted to generate adversarial examples during the incremental learning process; the adversarial training strategy is a method of adding perturbation samples during training to enhance the robustness of the model.

[0017] For updating the power fault knowledge graph, new fault cases are transformed into a knowledge graph structure, and graph neural networks are used to extract embedding representations. The knowledge graph is dynamically updated through link prediction and entity alignment techniques. Furthermore, a meta-learning mechanism is introduced to construct a shared feature extractor across fault types. The meta-learning mechanism refers to a training method that enables the model to quickly adapt to new tasks through cross-task training. The shared feature extractor refers to a feature extraction network that is universal across multiple fault types.

[0018] The technical effects and advantages of this invention are as follows: The physical space perception module collects electrical quantities, images, environmental data, and text data in real time through multi-source sensors, solving the problems of single perception dimensions and incomplete information in traditional detection methods. The data synchronization mapping module aligns multi-source data spatiotemporally and maps it to a digital twin space, resolving analysis errors caused by data spatiotemporal inconsistency. The digital twin space module constructs a multi-scale power grid twin model, solving the problems of low positioning accuracy and poor environmental adaptability in complex power grid scenarios through cross-modal attention fusion features and a hypothesis-deduction-verification fault location mechanism. The decision execution module incorporates a large language model and a power fault knowledge graph, generating diagnostic conclusions and handling suggestions through causal reasoning, solving the problems of high reliance on manual intervention and low decision-making efficiency. The system as a whole achieves deep fusion of multi-source information and intelligent closed-loop handling, significantly improving the accuracy and automation level of power grid fault detection. Attached Figure Description

[0019] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0020] Figure 1 This is a system block diagram of the present invention; Figure 2 This is the overall logic flowchart of fault deduction and waveform matching in this invention; Figure 3 This is a schematic diagram of the process for generating candidate regions in step 1 of the fault deduction and waveform matching process of the present invention; Figure 4 This is a schematic diagram of the adaptive time window extraction process in step 2 of the fault deduction and waveform matching process of the present invention. Figure 5 This is a flowchart illustrating step 3, rapid electromagnetic transient simulation, in the fault deduction and waveform matching process of this invention. Figure 6 This is a flowchart illustrating step 4, waveform matching degree calculation, in the fault deduction and waveform matching process of this invention. Detailed Implementation

[0021] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0022] Reference Figure 1-6 As shown, the present invention provides a technical solution:

[0023] I. System Overall Architecture: This invention provides a multi-source sensing intelligent detection system for precise location of power grid faults, such as... Figure 1 As shown, the system includes a physical space perception module, a data synchronization and mapping module, a digital twin space module, and a decision execution module. The system is deployed in a 220kV substation and its radiating transmission and distribution lines, covering typical active power distribution network scenarios, including distributed photovoltaic access points and wind power grid connection points. The modules communicate with each other via high-speed industrial Ethernet or fiber optic communication, using standard protocols such as IEC 61850 and IEEE C37.118 for data exchange.

[0024] II. Physical Space Sensing Module: The physical space sensing module is deployed at key nodes on the physical side of the power grid, specifically including electrical quantity sensors, visual sensors, environmental sensors, and text data interfaces.

[0025] The electrical quantity sensors include a synchronous phasor measurement unit (PMU) and a fault recorder (FTU). The PMU is installed on the substation busbar, distributed generation grid connection point, and important branch lines, with a sampling frequency set to 12.8 kHz, capable of capturing high-frequency components of transient fault processes, such as traveling wave fronts and transient high-frequency oscillations. The FTU is deployed at the line sectionalizing switch, also with a sampling frequency of 12.8 kHz. All electrical quantity sensors support the IEEE C37.118 protocol, outputting the amplitude and phase of three-phase voltage and current, as well as zero-sequence and negative-sequence components.

[0026] The visual sensors include a visible light camera and an infrared thermal imager. The visible light camera, with 4K resolution and a 30fps frame rate, is mounted on top of transmission towers and poles, and uses a pan-tilt-zoom (PTZ) control to capture images from a top-down angle, covering the transmission line corridor and insulator strings. The infrared thermal imager has a resolution of 640x512 and a temperature measurement range of -20℃ to 150℃, used to detect overheating in equipment such as overheated joints and partial discharge heating of insulators. The visual sensors have a built-in image preprocessing unit that performs automatic white balance, dehazing, and enhancement processing on the raw images.

[0027] Environmental sensors are integrated into the micro-meteorological monitoring station, including wind speed and direction sensors with an accuracy of ±0.2 m / s, temperature and humidity sensors with an accuracy of ±0.3℃ and ±2%RH, rainfall sensors with a resolution of 0.1 mm, and a light sensor. Environmental data is used to analyze the impact of meteorological factors on faults, such as strong winds causing foreign object bridging and thunderstorms causing insulation flashover.

[0028] The text data interface connects to the power dispatching data network via the EMS (Electric Power Management System), operation ticket system, and equipment ledger system to acquire unstructured text data in real time, including fault alarm information, operation ticket instructions, and equipment maintenance records. The interface uses a message queue to ensure reliable real-time data transmission.

[0029] III. Data Synchronization Mapping Module: The data synchronization mapping module includes a clock synchronization unit, an abnormal data preprocessing unit, and an interpolation alignment unit.

[0030] The clock synchronization unit uses GPS second pulses to synchronize the clocks of each sensor. All sensors have built-in GPS and BeiDou dual-mode timing modules, achieving microsecond-level synchronization via IRIG-B code or PTP protocol. The data synchronization mapping module receives data packets from each sensor, extracts timestamps, and ensures that multi-source data are on the same time base.

[0031] The outlier data preprocessing unit employs an outlier detection algorithm based on interquartile range (IQR) to identify outliers in electrical quantity data, such as spikes caused by transient sensor faults. For the detected outliers, a bidirectional long short-term memory (BiLSTM) network is used for completion and repair. The BiLSTM model takes as input time-series sequences of eight normal sampling points before and after the outlier (N = 8) and outputs the repaired estimated value. The model has been pre-trained on historical data, with a repair error of less than 2%.

[0032] The interpolation alignment unit uses cubic spline interpolation to uniformly interpolate electrical quantity data, image feature data, and environmental data with different sampling rates to a time axis with 1ms intervals. Image data is first processed by a convolutional neural network to extract feature vectors of key regions, such as the location of foreign objects and areas of abnormal temperature, and then time-aligned with electrical quantities. Finally, a spatiotemporally consistent multimodal fusion data stream is generated.

[0033] IV. Cross-modal feature association and inference unit: The cross-modal feature association and inference unit adopts a bidirectional cross-modal attention mechanism to perform deep feature interaction and fusion of four types of heterogeneous data: electrical quantities, images, environment, and text.

[0034] First, feature encoding is performed on each modal data: electrical quantity features: the 12.8kHz transient waveform is input into a 1D-CNN with convolution kernel sizes of 3, 5, and 7 to extract multi-scale temporal features. At the same time, wavelet packet decomposition is performed using db4 wavelet 3-layer decomposition to obtain the energy coefficients of each frequency band, forming a 256-dimensional electrical quantity feature vector.

[0035] Image features: For visible light and infrared images, an improved YOLOv8s target detection network with embedded EMA attention mechanism is used to detect targets such as foreign objects, insulator damage, and hot spots. The target category, location, and confidence level are output, and feature vectors of RoI regions are extracted to form a 128-dimensional image feature vector.

[0036] Environmental characteristics: Data such as wind speed, wind direction, temperature, humidity, and rainfall are normalized to form a 32-dimensional environmental feature vector.

[0037] Text features: The BERT model, finely tuned in the power field, is pre-trained on corpora such as power grid fault reports and operation tickets to transform the text into 768-dimensional semantic vectors, which are then reduced to 128-dimensional through a fully connected layer.

[0038] The bidirectional cross-modal attention mechanism uses electrical quantity features as the query matrix to calculate attention weights for image and text features, and simultaneously uses image features as the query matrix to calculate reverse attention for electrical quantity and text features. The results of the multiple attention calculations are concatenated and input into a two-layer fully connected network with 512-dimensional hidden layers, outputting a 256-dimensional fused feature vector.

[0039] V. Multi-scale power grid twin model: The multi-scale power grid twin model is constructed based on the power grid CIM model and actual equipment parameters, and uses the graph attention network (GAT) to dynamically model the power grid topology.

[0040] The model comprises a three-layer structure: Topology Layer: Employing a graph structure, nodes represent devices such as buses, switches, transformers, and distributed power sources, while edges represent lines or transformer windings. Real-time topology changes, such as switch opening and closing and line switching, are obtained from the SCADA system, and the adjacency matrix is ​​dynamically updated. Physical Layer: Contains detailed device parameters, such as line impedance per unit length and admittance, transformer short-circuit impedance and turns ratio, and the inverter control model for distributed power sources. Electromagnetic transient simulation employs an improved EMTP algorithm with adaptively adjusted time steps: 1 μs during fault periods and 50 μs during normal periods. Behavioral Layer: Stores historical operating data, typical fault cases, and equipment aging curves for model parameter calibration and fault mode matching.

[0041] GAT employs an 8-head attention mechanism, with each head outputting 64-dimensional features, which are then concatenated to obtain 512-dimensional node features. Node features include electrical quantities, equipment parameters, and environmental quantities. Edge weights are dynamically learned through the multi-head attention mechanism. The attention coefficient calculation formula is as follows: ,in, For nodes The multi-source fusion feature vector is composed of the fusion features output by the cross-modal feature association and inference unit. This vector integrates information from multiple sources such as electrical quantities, images, environment, and text. Nodes provided for digital twin models With nodes The electrical distance between them, which is calculated based on line impedance parameters, topology and real-time operating status, reflects the physical possibility of fault propagation in the power grid; This is a scale parameter used to adjust the intensity of the influence of physical distance on attention weights, and can be set according to the scale of the power grid or experience. For nodes The set of neighboring nodes. The 512-dimensional node features and cross-modal features output by GAT are correlated and added to the 256-dimensional fusion features output by the inference unit to obtain the final fusion feature containing spatial topological information.

[0042] VI. Fault Inference and Localization Unit: The fault inference and localization unit includes a candidate region generation sub-unit, a rapid simulation sub-unit, and a waveform matching sub-unit.

[0043] The candidate region generation sub-unit uses state estimation residual analysis to calculate the residuals between the measured values ​​of each node and the estimated values ​​of the twin model. Nodes with residuals exceeding a preset threshold and their adjacent lines are designated as candidate fault sections. The preset threshold is determined based on historical data statistics; in this embodiment, it is three times the standard deviation. The number of candidate sections does not exceed three, thus quickly narrowing down the location range.

[0044] The rapid simulation subunit sets hypothetical fault points within candidate sections in a multi-scale power grid twin model with a preset step size. In this embodiment, the step size is 50m, and the fault types include single-phase grounding, phase-to-phase short circuit, and open circuit. A model reduction technique is used to reduce the full-order electromagnetic transient model to a lower-order equivalent model for rapid electromagnetic transient simulation. This embodiment employs the vector matching method to reduce the high-order transfer function to a low-order rational function approximation. The simulation time for a single run is controlled within 50 milliseconds, generating simulated voltage and current waveforms for each node.

[0045] The waveform matching subunit calculates the matching degree between the simulated and measured waveforms. Before the matching degree calculation, adaptive time window truncation is performed on both the simulated and measured waveforms, and the time window is determined based on the fault initiation point detection algorithm. The fault initiation point detection algorithm uses the wavelet transform modulus maxima method to perform wavelet decomposition on the measured waveform, find the modulus maxima point, and determine the arrival time T0 of the traveling wave. The time window truncation takes waveform data for 5 ms, starting from T0 and taking 0.5 ms forward and 4.5 ms backward as the starting point.

[0046] The matching degree is calculated using a combination of the Dynamic Time Warping (DTW) algorithm and cross-correlation analysis, and the formula is as follows: ,in, For the simulation waveform sequence, This is a measured waveform sequence. For dynamic time-normalized distance, The maximum value of DTW distance across all simulation sequences. It is a cross-correlation function. , For signal power, The weighting coefficient ranges from 0.5 to 0.8; it is dynamically adjusted according to the fault type, and is set to 0.65 in this embodiment. The fault point with the highest matching degree is selected as the final accurate fault point, with a positioning error ≤ 50m.

[0047] For complex power grid scenarios with multiple power supply terminals, the waveform matching subunit calculates the matching degree of the current waveform at each terminal and uses a weighted voting mechanism to comprehensively determine the location of the fault point. The weight of each terminal is adaptively allocated according to its electrical distance from the fault point, with higher weights for closer electrical distances.

[0048] VII. Decision Execution Module: The decision execution module includes a fault diagnosis and handling unit, with a built-in large language model and power fault knowledge graph.

[0049] The large language model uses ChatGLM3-6B as its foundation and is fine-tuned on a power sector corpus. This corpus contains 100,000 historical fault reports, equipment manuals, and dispatching procedures, enabling the model to understand power grid terminology and fault causal relationships. The model supports a 32k context length.

[0050] The power fault knowledge graph is built on the Neo4j graph database and includes equipment nodes, fault type nodes, cause nodes, and handling measure nodes, as well as the relationships between them. It has stored over 5,000 historical fault cases, with relationships including occurrence, cause, handling, and relatedness.

[0051] The fault diagnosis and handling unit's workflow is as follows: Structured information such as fusion characteristics, fault location, real-time weather, and equipment status, along with text alarm information, are combined to construct natural language prompts. Using graph retrieval-enhanced generation (RAG) to retrieve the top-5 cases most similar to the current fault characteristics from the knowledge graph, the case content is injected into the large language model as context. The large language model performs causal reasoning, outputting a fault cause analysis, such as "B-phase insulator pollution flashover may be caused by haze," a confidence level of 92%, and handling suggestions, such as "immediately clean the insulator and check adjacent insulators." When the confidence level is higher than 95% and the handling suggestion is a routine operation such as reclosing or disconnecting distributed power sources, the system automatically generates control commands in IEC61850 Goose message format and sends them to the power grid smart terminal for execution; otherwise, a diagnostic report is pushed to the dispatcher for confirmation.

[0052] 8. Continuous learning mechanism: The digital twin space module is also used to update the network parameters of the multi-scale power grid twin model by using the waveform data, location results, diagnostic conclusions and handling effects generated in each fault handling process as new samples and adopting an incremental learning algorithm based on elastic weight consolidation EWC.

[0053] Elastic Weight Consolidation (EWC) introduces a parameter change penalty term into the loss function by calculating a diagonal matrix representing the importance of model parameters to historical tasks. Specifically, after training on historical tasks, the Fisher information matrix F is calculated, representing the importance of each parameter to the historical task.

[0054] A meta-learning mechanism constructs a shared feature extractor across fault types, enabling the model to rapidly generalize to novel fault patterns from a small number of new fault samples. Employing the MAML (Model-Agnostic Meta-Learning) algorithm, a large number of meta-tasks are constructed during the meta-training phase. Each meta-task contains support sets and query sets for several fault types. Through rapid adaptation in the inner loop and optimization of meta-parameters in the outer loop, initial model parameters with rapid adaptability to novel faults are trained. When a new fault appears, fine-tuning adaptation is only required with 1-5 samples.

[0055] Adversarial training strategies generate adversarial examples during incremental learning to enhance the model's robustness to noisy data and anomalous waveforms. The Fast Gradient Sign Method (FGSM) is used to generate adversarial examples, which are added to the training dataset, enabling the model to learn feature representations that are insensitive to small perturbations.

[0056] The dynamic updating of the temporal knowledge graph transforms new fault cases and their complete processing procedures into a temporal knowledge graph structure, employing a Graph Convolutional Network (GCN) to extract the embedded representations of new cases. Link prediction and entity alignment techniques automatically discover potential semantic associations between new cases and existing fault types, equipment attributes, and meteorological conditions. Link prediction uses the TransE model, embedding entities and relations into a low-dimensional vector space and evaluating the authenticity of triples. Entity alignment uses embedding-based similarity calculation, automatically establishing associations between new cases and the most similar existing entities. The dynamic updating of the knowledge graph forms an adaptive, continuous learning loop where model parameters and the knowledge base evolve collaboratively.

[0057] IX. Examples: To verify the technical effectiveness of the present invention, a prototype system was deployed in a real 220kV / 10kV hybrid distribution network environment, and three typical fault scenarios were selected for detailed testing. Each example fully records the parameters and intermediate results throughout the entire process from fault occurrence to system output.

[0058] Example 1: Lightning strike causing insulator flashover fault: Scenario description: A lightning strike occurred near tower 23 of a 110kV overhead line, causing flashover of phase B insulators and forming a single-phase ground fault. The fault occurred during a thunderstorm. Environmental sensors recorded an instantaneous wind speed of 12m / s and rainfall of 15mm / h. The lightning location system detected that the lightning strike point was approximately 50m away from the line.

[0059] Input parameters: Electrical quantities: PMU sampling rate 12.8kHz, after the fault, 10 cycles of data (approximately 200ms) are recorded. The peak current of phase B suddenly increases from the normal 300A to 2.8kA, the voltage drops to 30% of the normal value, and the zero-sequence current is significant; Image: The visible light camera is blurred due to rain and fog, and the infrared thermal imager shows that the temperature of the phase B insulator string of tower 23 rises abnormally, increasing by 8℃ within 1 minute after the fault; Environment: Wind speed 12m / s, rainfall 15mm / h, lightning location information includes timestamp, coordinates, and lightning current amplitude of 45kA; Text: The dispatch system issues a 110kV line grounding alarm, and the lightning location system pushes a lightning strike alarm at the same time.

[0060] System operation flow: 1. Data synchronization and preprocessing: After GPS synchronization, all data timestamps are aligned. Electrical quantity waveforms were analyzed using IQR anomaly detection, and no outliers were found. Image data was defogging and enhanced using a dark channel prior algorithm to extract insulator region features due to rain and fog. Cubic spline interpolation was used to unify electrical quantities, image features, and environmental data to a 1ms time axis. Lightning location information was encoded as text data into a 128-dimensional semantic vector using a power BERT model.

[0061] 2. Cross-modal feature fusion: Electrical waveforms are processed using 1D-CNN to extract temporal features, outputting 256 dimensions. Wavelet packet decomposition yields energy coefficients in eight frequency bands. Infrared images are processed using YOLOv8s to detect temperature features in the insulator region, including maximum temperature and temperature rise rate, outputting 64 dimensions. Environmental data such as wind speed, rainfall, and lightning parameters are used as 32-dimensional features. Text alarms are encoded using BERT to obtain 128 dimensions. All features are input into a bidirectional cross-modal attention module, which calculates attention between the image and text using electrical features as the query pair, while simultaneously performing inverse calculation, ultimately fusing them into a 256-dimensional feature vector.

[0062] 3. Topology Feature Injection: The power grid topology map includes 12 nodes on the line where tower 23 is located. The GAT model calculates the features of each node, with the attention weight of tower 23 being the highest at 0.28. The 256-dimensional feature vectors mentioned above are added to the node features output by GAT to obtain the final fused features.

[0063] 4. Candidate region generation: The residual analysis of the state estimation showed that the residual of the line on both sides of tower 23 exceeded the threshold by 3 times the standard deviation. The candidate region was locked between towers 22 and 24, about 1.5km away.

[0064] 5. Fault Simulation: In the digital twin model, a B-phase grounding fault was set every 50m in the candidate area, considering a lightning channel resistance of 5Ω, for a total of 31 simulation points. Using model order reduction technology, the single simulation time was 42ms, and the total simulation time was approximately 1.3 seconds. The matching degree between the simulated waveform and the measured waveform was calculated using the DTW+ cross-correlation formula, with λ equal to 0.65. The results showed that the highest matching degree, M = 0.96, was achieved at 25m downstream of tower 23, corresponding to 23+25m.

[0065] 6. Diagnostic Decision: The system integrates features, location results, lightning information, and infrared temperature rise data to construct a diagnostic framework. Five historical lightning strike fault cases are retrieved from the knowledge graph. The large language model outputs: The fault cause is a lightning strike leading to a flashover of the B-phase insulator with a 98% confidence level. It is recommended to perform zero-value testing on the insulator and check the grounding device. The system has successfully reclosed the circuit breaker and restored power supply. The system automatically generates a diagnostic report and pushes it to the maintenance center.

[0066] Output results: Location error: The actual fault point is located 30m downstream of tower 23. On-site inspection confirmed that the error is less than 30m, which meets the requirement of less than 50m; Diagnosis time: A total of 2.1 seconds from the occurrence of the fault to the generation of the report; Accuracy: Completely consistent with manual analysis.

[0067] Example 2: Cable Joint Overheating Leading to Breakdown Fault: Scenario Description: A 10kV cable line's intermediate joint experienced insulation aging due to long-term overload, leading to localized overheating and a single-phase ground fault. Infrared thermography detected an abnormal joint temperature, 15°C higher than normal, within one week prior to the fault, but this was not addressed promptly. At the time of the fault, the load current was 800A, the rated current was 600A, and the ambient temperature was 32°C.

[0068] Input parameters: Electrical quantities: FTU sampling rate 12.8kHz, during the fault, the A-phase current suddenly increases to 4.2kA, the voltage drops to 20%, and there is a significant DC component and the third harmonic content reaches 12%; Image: The weekly inspection record of the infrared thermal imager shows that the temperature of the joint at the fault point is consistently high. The last image before the fault shows that the surface temperature of the joint is 98℃, which is normal at 60℃. There are no obvious abnormalities in the visible light image; Environment: Ambient temperature 32℃, humidity 65%, no wind; Text: The dispatch system received a 10kV line grounding alarm. At the same time, the equipment ledger shows that the joint has been in operation for 8 years and the last maintenance date was 3 years ago.

[0069] System operation process: 1. Data synchronization and preprocessing: Electrical quantity data are synchronized for 5 cycles before and after the fault. The highest and average temperatures of the joint area are extracted from the infrared image, and the temperature rise rate is calculated. The temperature rise rate is approximately 2°C per day in the week before the fault. Text data alarms are added to the ledger and encoded using BERT.

[0070] 2. Cross-modal feature fusion: Electrical quantity features are extracted using 1D-CNN to extract waveform features; wavelet packet decomposition shows a significant 18% proportion of third harmonic energy. Infrared images are used to extract 64-dimensional connector temperature gradient features. Environmental features (temperature and humidity) are normalized to 2-dimensional. Text features are 128-dimensional. Cross-modal attention fusion outputs 256-dimensional features, with image temperature features having a relatively high attention weight (0.35) on electrical quantity features.

[0071] 3. Topology feature injection: In the power grid topology, this cable joint belongs to a node of a 10kV feeder. The GAT model calculation shows that the node has abnormal features, with an attention weight of 0.31.

[0072] 4. Candidate region generation: The state estimation residual analysis accurately locates the section where the joint is located. The cable section is 500m long and the joint is located at the midpoint.

[0073] 5. Fault Simulation: In the twin model, an A-phase grounding fault was set every 20m in the cable section, considering a time-varying arc resistance model, for a total of 25 simulation points. The matching degree between the simulated waveform and the measured waveform was calculated. The highest matching degree, M, was 0.94 at the joint location 250m from the beginning. Considering the aging factor of the joint, the contact resistance was increased in the simulation, further improving the matching degree to 0.96.

[0074] 6. Diagnostic Decision: The large language model, combined with fusion features, historical temperature data, and equipment records, retrieves cable joint overheating and breakdown cases from the knowledge graph. Output: The cause of the fault is long-term overload of the cable joint, leading to insulation aging and localized overheating, resulting in breakdown with a 95% confidence level. It is recommended to replace the joint and conduct infrared general testing on equipment in the same batch. The faulty section has been isolated and the load has been transferred to a backup line. The system automatically generates a maintenance work order containing precise location coordinates down to the joint's location.

[0075] Output results: Location error: The actual fault point is the joint, with an error of 0m; Diagnosis time: 3.5 seconds due to the need to retrieve historical temperature trend data; Accuracy: Consistent with post-disassembly inspection, confirming that the internal insulation of the joint has broken down.

[0076] Example 3: Complex Fault on Dual-Power Grid-Connected Side: Scenario Description: A phase-to-phase short circuit occurs between phases B and C in a distributed photovoltaic grid-connected line in an active distribution network. The fault point is located near the photovoltaic grid connection point. During the fault, the photovoltaic output fluctuates, and the power flow direction is complex, rendering traditional impedance methods ineffective for fault location.

[0077] Input parameters: Electrical quantities: The grid-connected point PMU records the three-phase voltage and current. The BC phase current suddenly increases and the voltage drops. At the same time, the output current of the photovoltaic inverter is distorted; Image: The drone inspection image shows that the line near the grid-connected point has tree branches overlapping. The visible light image shows no abnormalities in the infrared; Environment: Light wind 2m / s, sunny; Text: The dispatch system receives a short circuit alarm for the photovoltaic grid-connected line, and the photovoltaic monitoring system pushes the inverter protection action information.

[0078] System operation procedure: 1. Data synchronization and preprocessing: Align electrical quantity waveforms and extract 5 cycles before and after the fault. The confidence level of the tree branch target detected in the image is not less than 0.85, and the extracted location coordinates are approximately 100m away from the grid connection point. The text information includes photovoltaic power output data of 300kW before the fault, which is encoded and then input.

[0079] 2. Cross-modal feature fusion: Electrical quantity features are extracted using 1D-CNN, and wavelet packet decomposition reveals that the fault phase-to-phase current contains high-frequency components, possibly due to tree discharge. Image features include tree location and size. Cross-modal attention is applied, with image features receiving a higher attention weight (0.42) for electrical quantity features.

[0080] 3. Topology feature injection: The GAT model processes the topology graph containing photovoltaic grid-connected nodes. The grid-connected nodes have high attention weights and are closely connected with their neighboring nodes.

[0081] 4. Candidate region generation: State estimation residual analysis identifies 200m sections before and after the grid connection point.

[0082] 5. Fault Simulation: In the twin model, a phase-to-phase short circuit (BC-Phase Short Circuit) is set every 50m in the section, considering the inverter control response of the equivalent model with photovoltaic access. The simulation results match the measured waveforms, with the highest matching degree point located near the tree branch position 80m upstream of the grid connection point, where M equals 0.92. For multi-terminal power supply scenarios, the waveform matching subunit calculates the matching degree of the current waveform at each terminal separately, and a weighted voting mechanism is used for comprehensive judgment. The weights of each terminal are adaptively allocated according to the electrical distance, ultimately confirming the location of the fault point.

[0083] 6. Diagnostic Decision: The large language model, combined with image and electrical features, outputs: The fault is caused by overlapping tree branches leading to a phase-to-phase short circuit (93% confidence level). It is recommended to clear the branches and prune nearby trees. Since the photovoltaic inverter has been operating correctly, no equipment maintenance is required. The system generates an inspection task sheet and assigns a drone to immediately proceed with the cleanup.

[0084] Output results: Positioning error: The actual location of the tree branch is 75m away from the grid connection point, with an error of 5m; Diagnosis time: 2.8 seconds; Accuracy: Consistent with on-site inspection.

[0085] Example 4: Continuous Learning Capability Verification: Scenario Description: After a new photovoltaic power station is connected to the grid, a novel fault mode emerges—voltage sag caused by inverter transient disturbances. This fault mode is not present in the historical training data. The system needs to quickly learn the new fault without forgetting old fault knowledge.

[0086] Experimental Design: The initial model was trained on 5000 historical fault cases, including traditional fault types such as lightning strikes, foreign object short circuits, and connector overheating. Twenty novel inverter transient disturbance fault samples were introduced, and the system was compared using three strategies: Strategy A: directly fine-tuning the model on the new samples; Strategy B: retraining with a mix of old and new samples; Strategy C: the EWC + meta-learning + adversarial training mechanism used in this invention.

[0087] Experimental Results: Strategy A achieved a 92% recognition rate for novel faults, but its recognition rate for traditional faults dropped from 95% to 76%, exhibiting severe catastrophic forgetting. Strategy B maintained a 94% recognition rate for traditional faults, but required a training time of 5 hours and storage of all historical data. Strategy C, using only 20 new samples for incremental learning, achieved an 89% recognition rate for novel faults and maintained a traditional fault recognition rate above 93%, with a training time of only 8 minutes. Adversarial training improved the model's robustness to noisy waveforms by 12%.

[0088] Meta-learning mechanism test: When only one novel fault sample is provided, policy C generalizes rapidly through meta-learning, achieving a recognition rate of 78%; when five samples are provided, the rate reaches 89%. This indicates that the meta-learning mechanism enables the model to learn quickly from a small number of samples.

[0089] Time-series knowledge graph update: After new cases are transformed into graph structures, their association with existing entities such as photovoltaic equipment, voltage sags, and inverters is automatically discovered through link prediction. After the knowledge graph adds nodes and edges, the diagnostic accuracy of similar faults in the future is improved by 15%.

[0090] Example 5: Robustness Verification of Multi-End Voting: Scenario Description: Based on the dual-power grid connection scenario in Example 3, anomalies in sensor data at one end are artificially simulated to verify the robustness of the two-layer voting mechanism.

[0091] Experimental Design: The tree branch splicing fault scenario from Example 3 was used, with the actual fault point located 80m upstream of the grid connection point. 10dB Gaussian noise was artificially injected into the PMU data at the grid connection point to simulate sensor interference. Three strategies were compared: Strategy A: Single-layer voting, i.e., weighted only by electrical distance; Strategy B: Double-layer voting without dynamic removal; Strategy C: The present invention's double-layer voting + dynamic removal + Bayesian posterior probability.

[0092] Experimental Results: In Strategy A, the matching degree at the grid connection point decreased due to noise, but because of its close electrical distance, its weight was high, leading to a shift in the overall matching degree, resulting in a final positioning error of 120m. In Strategy B, the confidence neural network detected a decrease in data quality at the grid connection point, outputting a confidence level w2=0.45, but it did not remove the data and it still participated in the voting, resulting in a positioning error of 45m. In Strategy C, the confidence neural network output w2=0.45, which was lower than the threshold of 0.6. The system automatically removed the data from that end and performed positioning based only on the data from the other end, resulting in a positioning error of 12m. Bayesian posterior probability calculation showed that the posterior probability of the result after removal was 0.92, which was much higher than the result of retaining the abnormal end (0.31).

[0093] Statistical analysis of multiple experiments showed that in 20 tests with different noise levels, Strategy C had an average positioning error of 18m and a maximum error of 35m; Strategy A had an average positioning error of 76m and a maximum error exceeding 150m. This verifies the effective suppression capability of the two-layer voting mechanism for single-end measurement errors.

[0094] 10. Overall Performance Statistics: During the 6-month trial operation, the system detected a total of 127 faults of various types. The positioning accuracy statistics for five typical scenarios are as follows:

[0095] Table 1. Overall Results Statistics

[0096] Fault type frequency Average positioning error Maximum error Diagnostic accuracy Lightning flashover 23 32 meters 72 meters 94.5% Overheating of the connector 15 18 meters 42 meters 98.0% Foreign object short circuit 31 26 meters 56 meters 95.2% Inverter disturbance 20 22 meters 48 meters 90.5% other 38 38 meters 78 meters 92.8% Total / Average 127 29.8 meters 78 meters 93.7%

[0097] Over 90% of the overall positioning errors did not exceed 50 meters, with an average diagnosis time of 3.5 seconds. A continuous learning mechanism improved the system's adaptability to new faults by 76%, and a two-layer voting mechanism improved positioning accuracy in multi-terminal scenarios by 62%. This invention, through multi-source data fusion, digital twin simulation, and continuous learning evolution, achieves accurate and rapid positioning and intelligent handling of power grid faults in complex environments.

[0098] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A multi-source sensing intelligent detection system for precise location of power grid faults, characterized in that, include: The physical space sensing module, deployed on the physical side of the power grid, includes various types of sensors and data interfaces, and is used to collect multi-dimensional heterogeneous data on power grid operation in real time. The data synchronization mapping module is used to perform spatiotemporal alignment processing on the multidimensional heterogeneous data and map it to the digital twin space; The digital twin space module includes a multi-scale power grid twin model, a cross-modal feature association and inference unit, and a fault inference and localization unit. The multi-scale power grid twin model is used to characterize the physical characteristics and operational behavior of the power grid. The cross-modal feature association and inference unit is used to receive mapped multi-source data and use a bidirectional cross-modal attention mechanism to perform feature interaction and fusion on the multi-source data, and output fused features. The fault inference and location unit is used to generate candidate fault sections based on the fusion features using state estimation residual analysis. In the multi-scale power grid twin model, hypothetical faults are set for the candidate fault sections and electromagnetic transient simulations are performed. The simulation waveforms are compared with real-time collected electrical quantity data to calculate the waveform matching degree. The precise fault point is determined based on the matching degree. The waveform matching degree calculation adopts a combination of dynamic time warping algorithm and cross-correlation analysis. The decision execution module includes a fault diagnosis and handling unit. This unit has a built-in large language model, which is used to perform causal reasoning based on the fusion features and the precise fault point, combined with a pre-built power fault knowledge graph, to generate fault cause analysis and handling suggestions, and output a diagnostic report or issue control commands. The digital twin space module is also used to feed back the data and results of each fault handling process to the multi-scale power grid twin model and the power fault knowledge graph, so as to realize the iterative evolution of model parameters and knowledge base.

2. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The physical space perception module constructs a multi-dimensional holographic perception network covering electrical, visual, environmental, and semantic dimensions. Among them, electrical quantity sensors are deployed at key nodes of the power grid and adopt a high sampling frequency that can completely record fault traveling waves. Visual sensors achieve top-down shooting of key equipment and lines through gimbal control and simultaneously acquire visible light and infrared images. Environmental sensors monitor micro-meteorological data in real time. The text data interface connects to the power dispatch data network to obtain unstructured text data.

3. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The data synchronization mapping module includes: a clock synchronization unit, which uses the second pulse of the satellite navigation system to synchronize the clocks of each sensor; an abnormal data preprocessing unit, which uses an outlier detection algorithm based on statistical distribution to identify outliers in electrical quantity data, and uses a sequence learning model to perform context completion and repair on the identified abnormal data; and an interpolation and alignment unit, which uses an interpolation method to uniformly interpolate multi-source data with different sampling rates to a high-precision time axis, and extracts feature vectors from image data before performing time alignment to generate a spatiotemporally consistent multimodal fusion data stream.

4. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The cross-modal feature association and inference unit employs a bidirectional cross-modal attention mechanism to perform feature interaction and fusion on multidimensional heterogeneous data. The bidirectional cross-modal attention mechanism uses electrical quantity features as queries to perform attention calculations on image features and text features, and simultaneously uses image features as queries to perform reverse attention calculations on electrical quantity features and text features. After fusing the results of multiple attention calculations, a fused feature vector of a preset dimension is output. The electrical quantity features are obtained by temporal feature extraction and frequency domain decomposition, the image features are obtained by object detection and the attention mechanism, and the text features are obtained by a pre-trained language model in the power domain.

5. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The multi-scale power grid twin model employs a graph attention network to dynamically model the power grid topology, treating power grid nodes as vertices and lines as edges. Node features include fused features and physical features. The graph attention network uses a multi-head attention mechanism, where each head can be configured with different scale parameters or electrical distance definitions. The features output by multiple heads are concatenated to obtain the final fused feature containing spatial topology information. The attention coefficients between nodes are calculated using the following formula: ,in, For nodes Multi-source fusion feature vector Nodes provided for digital twin models With nodes Electrical distance between them For scale parameters, For nodes The set of neighboring nodes.

6. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The fault simulation and localization unit includes: a candidate region generation subunit, used to generate candidate fault segments based on state estimation residuals, where the state estimation residuals refer to the deviation between measured values ​​and twin model estimated values; a simulation subunit, used to perform electromagnetic transient simulations on the candidate segments in the twin model to generate simulation waveforms; and a waveform matching subunit, used to calculate the matching degree between the simulation waveform and the measured waveform, employing a combination of dynamic time warping algorithm and cross-correlation analysis. The matching degree calculation formula is as follows: ,in, For the simulation waveform sequence, This is a measured waveform sequence. For dynamic time-normalized distance, The maximum value of DTW distance across all simulation sequences. It is a cross-correlation function. , For signal power, The weighting coefficients are used to balance morphological similarity and phase alignment; the fault point with the highest matching degree is selected as the final accurate fault point.

7. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 6, characterized in that: For multi-terminal power supply scenarios, the waveform matching subunit adopts a two-layer adaptive weighted voting mechanism to comprehensively determine the location of the fault point; the first layer weight is adaptively allocated based on electrical distance, using a function that is inversely proportional to the electrical distance; The second layer of weights is based on dynamic data quality assessment, and uses a confidence assessment unit to take data quality parameters as input and output confidence coefficients. After merging the two layers of weights, the matching degree of each end is weighted and averaged. When the confidence of a certain end is lower than the preset threshold, the end is automatically removed from the voting set. For the remaining ends, uncertainty estimation is introduced to calculate the posterior probability. The uncertainty estimation refers to the method of calculating the distribution of the probability of each candidate position through a probability model, and the position with the highest posterior probability is selected as the final fault point.

8. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The large language model built into the fault diagnosis and handling unit is a generative pre-trained model based on fine-tuning of power domain corpus; the power fault knowledge graph includes equipment, fault type, cause, handling measures nodes and their relationships, and stores historical fault cases.

9. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 8, characterized in that: The fault diagnosis and handling unit retrieves similar cases from the knowledge graph as context input to the large language model through graph retrieval enhancement generation. The large language model combines fusion features, fault location and retrieved cases to perform causal reasoning, generate fault cause analysis, confidence assessment and handling suggestions, and automatically generate control commands based on the handling suggestions to be issued to the execution agency.

10. The intelligent detection system for precise location of power grid faults based on multi-source sensing according to claim 1, characterized in that: The digital twin space module is also used to collaboratively update the multi-scale power grid twin model and the power fault knowledge graph by using waveform data, location results, diagnostic conclusions, and handling effects generated during each fault handling process as new samples. Specifically, for updating the multi-scale power grid twin model, a regularized incremental learning algorithm is used to update network parameters by applying penalty terms to important model parameters, and an adversarial training strategy is used to generate adversarial examples during the incremental learning process. For updating the power fault knowledge graph, new fault cases are transformed into a knowledge graph structure, and graph neural networks are used to extract embedding representations. The knowledge graph is dynamically updated through link prediction and entity alignment techniques. Furthermore, a meta-learning mechanism is introduced to construct a shared feature extractor across fault types.

Images