High-voltage circuit fault positioning and automatic isolation system and method

Abstract

The invention discloses a high-voltage circuit fault positioning and automatic isolation system and method, and belongs to the technical field of power plant relay protection. According to the invention, a multi-mode sensing layer arranged at key nodes of high-voltage power transmission and transformation equipment of a power plant is used for cooperatively collecting high-frequency electrical, vibration and temperature signals; the physical-data fusion engine is used for carrying out accurate fault positioning by utilizing a graph neural network which combines a line topological structure and an electromagnetic wave propagation physical rule, and analyzing a fault root cause by utilizing a causal inference module; the risk assessment and isolation decision-making layer integrates a digital twin assessor and a distributed consensus module and generates an optimal isolation scheme; and on the basis of a security and auditing layer of an alliance chain, the non-tampering and traceability of data in the whole process are ensured. Rapid and accurate positioning and intelligent isolation of faults are realized, the automation level and reliability of power plant fault processing are improved, and meanwhile, the non-tampering and traceability of data in the whole processing process are ensured.

Description

Technical Field

[0001] This invention belongs to the field of power plant relay protection technology, specifically a high-voltage circuit fault location and automatic isolation system and method. Background Technology

[0002] High-voltage transmission and transformation equipment in power plants is a critical link in the transmission of electrical energy, and its operational reliability directly affects the safety, stability, and power transmission capacity of the power plant. When a fault occurs in a high-voltage circuit, quickly and accurately locating the fault point and implementing effective isolation are crucial to preventing the fault from escalating, minimizing the power outage area, and accelerating the restoration of power supply.

[0003] Existing high-voltage circuit fault location and isolation technologies mainly rely on the following methods:

[0004] Firstly, in fault location, traveling wave ranging and impedance methods are widely used. Traveling wave ranging calculates the fault distance by detecting the time difference between the arrival of the traveling wave generated by the fault at the measuring end, and its theoretical accuracy is relatively high. However, in practical applications, this method is affected by factors such as uncertain wave velocity, difficulty in accurately identifying the traveling wave front, and changes in the frequency characteristics of line parameters, resulting in errors in the location results. Especially for complex multi-branch lines or cable-overhead mixed lines, the reflection and refraction of the traveling wave signal makes fault point identification even more difficult. Impedance methods calculate the fault loop impedance based on the voltage and current after the fault, thereby estimating the fault distance. This method is simple in principle, but it is easily affected by factors such as system operating mode, transition resistance, and line asymmetry. In long lines or high-resistance grounding faults, the location accuracy decreases significantly.

[0005] Secondly, in terms of fault detection, traditional systems mainly rely on electrical quantity sensors (such as voltage transformers and current transformers) installed in substations. This detection method has monitoring blind spots and cannot effectively capture non-electrical fault signs occurring in the middle of the line (such as conductor vibration caused by external force damage and abnormal rise in insulator surface temperature), lacking collaborative judgment based on multi-dimensional information. Existing methods generally lack the ability to deeply identify the nature of faults. For cascading fault events triggered by a single initial fault, it is difficult to accurately locate the root fault point, leading to blind isolation strategy formulation and potentially expanding the scope of power outages.

[0006] In recent years, artificial intelligence technology, especially graph neural networks, has been introduced into the field of fault diagnosis. For example, the "Fault Diagnosis Method Based on Adaptive Graph Neural Network Multi-Source Data Fusion" (Publication No. CN117332818A) learns multi-graph structures through adaptive adjacency matrices to estimate fault types. However, most of these methods belong to a purely data-driven paradigm, and their model training and inference heavily rely on the completeness and quality of historical data, lacking clear physical constraints. When data is scarce or unseen fault scenarios occur, their generalization ability and the physical rationality of the results face challenges, and interpretability is often insufficient.

[0007] Therefore, there is an urgent need for a high-voltage circuit fault location and automatic isolation solution that can integrate multimodal information, accurate fault location and collaborative decision-making capabilities, and utilize physical mechanisms. Summary of the Invention

[0008] To address the above problems, this invention provides a high-voltage circuit fault location and automatic isolation system and method to solve the problems of inaccurate fault location results, insufficient utilization of physical mechanisms, and lack of multi-dimensional information coordination.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A high-voltage circuit fault location and automatic isolation system, comprising:

[0011] The multimodal sensing layer, located at the outgoing lines, line segmentation points, and important load access points of high-voltage transmission and transformation equipment in power plants, includes a high-frequency sampling unit, a distributed optical fiber sensing system, a synchronous phasor measurement unit, and edge computing nodes. The high-frequency sampling unit collects voltage and current signals from the high-voltage transmission and transformation equipment via electrical connections, with a sampling rate of 1MHz to 10MHz. The distributed optical fiber sensing system is deployed parallel to the high-voltage transmission and transformation equipment via optical fiber composite overhead ground wires or optical fibers laid along cable trenches, collecting vibration and temperature signals. The synchronous phasor measurement unit is connected to the high-voltage transmission and transformation equipment via voltage and current transformers, collecting three-phase voltage and current phasors. The edge computing nodes are connected to the high-frequency sampling unit, the distributed optical fiber sensing system, and the synchronous phasor measurement unit via a data bus, and are used to preprocess the collected voltage and current signals, vibration and temperature signals, and three-phase voltage and current phasors of the high-voltage circuit, outputting multimodal sensing layer data.

[0012] The physical-data fusion engine connects to the multimodal sensing layer via a communication network, receives data from the multimodal sensing layer, and includes a physical model library, a topology sensing graph neural network module, and a causal inference module. The input of the topology sensing graph neural network module is the line topology and multimodal sensing data, and the output is the probability distribution of fault location and its confidence level. The causal inference module is used to analyze causal relationships in multi-fault scenarios based on the transfer entropy algorithm and combined with Granger causality test.

[0013] The risk assessment and isolation decision layer connects to the physical-data fusion engine via a control network, receives the probability distribution of fault locations and causal relationships, and outputs isolation decisions. The risk assessment and isolation decision layer includes a digital twin evaluator, an isolation strategy library, and a distributed consensus module.

[0014] The security and auditability layer connects to all layers of the system through data interfaces. It adopts a distributed ledger based on a consortium blockchain to record event data throughout the fault handling process and provide auditing functions.

[0015] Furthermore, the edge computing node has a built-in clock that is synchronized with the system's main clock via the IEEE 1588 protocol and using GPS and BeiDou dual-redundant clock sources. The preprocessing operations of the edge computing node include: extracting the wavefront features of current and voltage traveling waves using the wavelet transform modulus maxima method, and extracting the frequency domain energy features of the vibration signal using short-time Fourier transform.

[0016] Furthermore, the topology-aware graph neural network module adopts a graph convolutional network structure, containing 6-8 convolutional layers, each containing 128-256 neurons, and its message passing function is defined as: ,in For node features, The physical distance between nodes. This is the time difference of arrival of the traveling wave, calculated based on the propagation speed of electromagnetic waves.

[0017] Furthermore, the causal inference module is specifically used to: sequence the fault events of each monitoring point according to time, calculate the transition entropy between each pair of events, and if the transition entropy exceeds the threshold determined by statistical testing and the Granger causality test P value is less than 0.05, then establish a directed causal edge to construct a fault causal directed graph, and identify the node with zero or minimum in-degree in the graph as the root fault point.

[0018] Furthermore, the digital twin evaluator employs a surrogate model based on a combination of dynamic mode decomposition and deep neural networks to perform rapid power flow calculations and risk assessments on candidate isolation schemes; the evaluation metrics include power outage load, equipment overload probability, and estimated recovery time.

[0019] Furthermore, the isolation strategy library stores specific rules for minimum cut set strategy, regional isolation strategy and selective isolation strategy; the distributed consensus module adopts a weighted voting mechanism, and the voting weight of each decision unit is dynamically calculated based on the load capacity under its jurisdiction and its centrality in the line topology, with the consensus threshold set to 80%.

[0020] Furthermore, the security and auditability layer adopts practical Byzantine fault tolerance as the consensus mechanism, with a data backup interval of 1 second. High-frequency event data is processed using a combination of aggregation compression and off-chain storage. The event data recorded by the layer includes at least the event trigger time, fault location results, causal inference graph, risk assessment results of each candidate solution, decision basis, and final execution results.

[0021] A method for fault location and automatic isolation in high-voltage circuits includes the following steps:

[0022] S1: Real-time acquisition of multimodal signals from high-voltage power transmission and transformation equipment through the multimodal sensing layer. When any feature of voltage change rate, current change rate, or vibration signal energy exceeds a set threshold, a fault detection event is triggered.

[0023] S2: Edge computing nodes preprocess the acquired signals, extract traveling wave features, vibration spectrum features, and temperature distribution features, and generate event summary data;

[0024] S3: The physical-data fusion engine receives multimodal sensing data, performs message passing and node state updates based on the topology sensing graph neural network, and outputs the probability distribution and confidence level of the fault location.

[0025] S4: The causal inference module analyzes the causal relationship of fault events based on transfer entropy and combined with Granger causality test, determines whether it is a single point fault or a multi-point fault, and locates the root cause.

[0026] S5: The digital twin evaluator uses fault location and root cause information to perform parallel simulations and evaluate the risk indicators of multiple isolation schemes.

[0027] S6: The distributed consensus module determines the optimal isolation scheme based on the risk assessment results through a weighted voting mechanism;

[0028] S7: The control system performs isolation operations while the safety and auditability layer records the complete fault handling process;

[0029] S8: Based on the real data from this processing result, the system performs offline incremental training and updates on the topology-aware graph neural network and the digital twin agent model.

[0030] Furthermore, the specific process of fault location described in S3 is as follows:

[0031] S31: Construct a circuit topology diagram, where nodes represent the installation locations of sensing devices and edges represent electrical connections;

[0032] S32: Extract node feature vectors, including voltage surge amplitude, current traveling wave arrival time, vibration signal energy, and temperature change rate;

[0033] S33: Message passing is performed through a graph neural network, and the message function includes the calculation of electromagnetic wave propagation delay;

[0034] S34: The output layer generates a probability distribution of fault locations, and the confidence level is calculated based on a Bayesian inference method.

[0035] Furthermore, the specific process of risk assessment in S5 is as follows: S51: Generate a set of candidate isolation schemes based on the fault location results; S52: The digital twin system performs fast power flow calculations for each isolation scheme; S53: Calculate the risk indicators for each scheme, including the power outage load range of 0MW~200MW, equipment overload probability of 0%~100%, and recovery time of 0s~300s; S54: Sort the schemes based on a multi-objective optimization algorithm, with the objective function being min[α·power outage load + β·equipment overload probability + γ·recovery time], where α, β, and γ are weight coefficients set according to the operation strategy; Before the isolation scheme is executed, a safety constraint check is performed, including power flow constraint check, stability constraint check, protection coordination constraint check, and operation sequence constraint check.

[0036] The beneficial effects of this invention are as follows: By integrating multiple physical quantities such as high-frequency electrical traveling waves, synchronous phasors, vibration, and temperature, a three-dimensional fault perception network is constructed. The mutual verification between electrical and physical quantities overcomes the insufficient sensitivity of single electrical quantity detection when dealing with complex faults such as high-resistance grounding and intermittent arcing. The topology-sensing graph neural network embeds physical laws such as line topology and electromagnetic wave propagation delay into the message passing function, making the model's inference process constrained by physical laws. The resulting fault probability distribution is more reliable and has higher confidence than the single result of traditional traveling wave or impedance methods. The causal inference module, through transfer entropy analysis, can identify causal chains in multiple faults from a time series perspective, accurately pinpointing the "root fault point" that triggers subsequent chain reactions. Targeting the fault characteristics of high-voltage transmission and transformation equipment in power plants, it achieves rapid and accurate location and intelligent isolation, improving the automation level and reliability of power plant fault handling, while ensuring the immutability and traceability of data throughout the entire processing process. Attached Figure Description

[0037] Figure 1 A block diagram of a high-voltage circuit fault location and automatic isolation system;

[0038] Figure 2 This is a flowchart for fault location and automatic isolation in high-voltage circuits. Detailed Implementation

[0039] To enable those skilled in the art to better understand the technical solution, the present invention will be described in detail below with reference to embodiments. The description in this part is only exemplary and explanatory, and should not be used to limit the scope of protection of the present invention in any way.

[0040] Example 1

[0041] A specific implementation of a high-voltage circuit fault location and automatic isolation system is disclosed. The system is configured on a 280-kilometer-long 500kV high-voltage outgoing line of a power plant, with eight monitoring points along the line, spaced 40 kilometers apart. Simultaneously, a system with the same architecture is configured on the 220kV outgoing line side of the power plant for collaborative monitoring.

[0042] The hardware configuration of the multimodal sensing layer is as follows: the sampling frequency of the high-frequency sampling unit is set to 10MHz; the voltage sampling module uses a 2000:1 capacitive voltage divider; and the current sampling module uses a 4000:1 Rogowski coil. The distributed fiber optic sensing system uses a distributed acoustic sensing module of model APSensingN4385A, with a spatial resolution of 10 meters and a sampling frequency of 2kHz. The synchronization phasor measurement unit uses a DT-8000 series device. The edge computing nodes use Huawei Atlas800 inference servers, equipped with a 16-core CPU and 32GB of memory. Clock synchronization enhancement design: the edge computing nodes have a built-in high-stability temperature-controlled crystal oscillator, using GPS and BeiDou dual-redundant clock sources, with a synchronization deviation tolerance of 1 microsecond.

[0043] The system operation process is as follows:

[0044] When a single-phase ground fault occurs 105 kilometers from the substation, the high-frequency sampling unit detects a voltage change rate exceeding 580 kV / ms and a current change rate exceeding 8.5 kA / ms. The distributed fiber optic sensing system detects a vibration signal energy value of 35 dB at the corresponding location. The time synchronization module of the edge computing node aligns the timestamps of the data from each monitoring point.

[0045] Edge computing nodes employ 8-level wavelet decomposition using the db4 wavelet basis function to extract the arrival time of the traveling wavefront. The arrival times of the wavefronts at each monitoring point are as follows: 75 microseconds for monitoring point 1, 110 microseconds for monitoring point 2, 145 microseconds for monitoring point 3, 180 microseconds for monitoring point 4, 215 microseconds for monitoring point 5, and 250 microseconds for monitoring point 6. Vibration signal analysis utilizes a Hanning window for short-time Fourier transform with a window length of 100 milliseconds and an overlap rate of 50%.

[0046] The physics-data fusion engine constructs a circuit topology graph containing 8 nodes and 7 edges. Node feature vectors include: voltage surge amplitude, current traveling wave arrival time, vibration signal energy, and temperature change rate. The graph neural network uses 8 convolutional layers, each with 512 neurons, a learning rate of 0.001, and 1000 training cycles. The electromagnetic wave propagation speed in the message passing function is set to 2.98 × 10^8 m / s. The MLP message passing function has a three-layer fully connected network with 256 hidden neurons using the ReLU activation function, which effectively introduces nonlinearity and alleviates the gradient vanishing problem in deep networks. The graph neural network is trained using the Adam optimizer, and the loss function combines cross-entropy loss and a physical consistency regularization term.

[0047] The topology-aware graph neural network outputs the probability distribution of the fault location, showing a probability of 94.2% at 105.3 km. The causal inference module calculates the transfer entropy between nodes, and the maximum value does not exceed the threshold of 0.15 bits, indicating a single point of failure.

[0048] The digital twin evaluator for risk assessment and isolation decision-making generates isolation schemes: The digital twin evaluator adopts a surrogate model based on a combination of dynamic mode decomposition and deep neural networks. The specific architecture of the surrogate model is as follows: the 5000-dimensional system is reduced to 100 dimensions through DMD, and the DNN contains 4 hidden layers (512 / 256 / 128 / 64 neurons); the surrogate model is trained using the Adam optimizer; Scheme 1 has a power outage load of 65MW, an overload probability of 8%, and a recovery time of 240 seconds; Scheme 2 has a power outage load of 45MW, an overload probability of 15%, and a recovery time of 180 seconds; Scheme 3 has a power outage load of 80MW, an overload probability of 5%, and a recovery time of 300 seconds. The digital twin evaluator uses the dynamic mode decomposition method to reduce the system order from 5000 dimensions to 100 dimensions, and the simulation time is controlled within 120 milliseconds.

[0049] The distributed consensus module employs a weighted voting mechanism, with each node's weight allocated based on its managed load capacity: 0.5 for the 500kV master station, 0.3 for the 220kV interconnection switch, and 0.2 for the segment switch. Scheme 2 achieved a consensus score of 0.82, exceeding the threshold of 0.8, and was therefore determined as the optimal isolation scheme.

[0050] The security and auditability layer utilizes the Fabric consortium blockchain platform, with 6 ledger nodes and a practical Byzantine fault-tolerant consensus mechanism, generating a new block every second. Recorded data includes the time of failure, location of failure, type of failure, and isolation scheme employed. In this embodiment, the recorded data includes the failure time (May 20, 2024, 14:30:25:358), the failure location (105.3 km), the failure type (single-phase grounding), the isolation scheme used (number 2), and the execution result (successful).

[0051] Example 2

[0052] This embodiment provides a method for locating and automatically isolating faults in high-voltage circuits.

[0053] S1: Real-time acquisition of multimodal signals from high-voltage power transmission and transformation equipment through the multimodal sensing layer. When any feature in the voltage change rate, current change rate, or vibration signal energy exceeds the adaptively set threshold, a fault detection event is triggered.

[0054] S2: Edge computing nodes preprocess the acquired signals, extract traveling wave features, vibration spectrum features, and temperature distribution features, and generate event summary data;

[0055] S3: The physical-data fusion engine receives multimodal sensing data, performs message passing and node state updates based on the topology sensing graph neural network, and outputs the probability distribution and confidence level of the fault location.

[0056] S4: The causal inference module analyzes the causal relationship of fault events based on transfer entropy and Granger causality test, determines whether it is a single point fault or a multi-point fault, and locates the root cause.

[0057] S5: The digital twin evaluator uses fault location and root cause information to perform parallel simulations and evaluate the risk indicators of multiple isolation schemes.

[0058] S6: The distributed consensus module determines the optimal isolation scheme based on the risk assessment results through a weighted voting mechanism;

[0059] S7: The control system performs isolation operations while the safety and auditability layer records the complete fault handling process;

[0060] S8: Based on the real data from this processing result, the system performs offline incremental training and updates on the topology-aware graph neural network and the digital twin agent model.

[0061] Step S1 includes at least steps S110-S120:

[0062] S110. Real-time acquisition of multimodal signals from high-voltage power transmission and transformation equipment via a multimodal sensing layer. Specifically, the multimodal sensing layer is installed at the outgoing lines, line segmentation points, and important load access points of the high-voltage power transmission and transformation equipment in the power plant, and includes a high-frequency sampling unit, a distributed optical fiber sensing system, a synchronous phasor measurement unit, and an edge computing node. The high-frequency sampling unit acquires voltage and current signals from the high-voltage power transmission and transformation equipment through electrical connections, with a sampling rate of 10MHz. The distributed optical fiber sensing system is deployed in parallel with the high-voltage power transmission and transformation equipment via optical fiber composite overhead ground wires or optical fibers laid along cable trenches, acquiring vibration and temperature signals, with a vibration signal sampling frequency of 2kHz. The synchronous phasor measurement unit is connected to the high-voltage power transmission and transformation equipment through voltage transformers and current transformers, acquiring three-phase voltage and current phasors. The edge computing node is connected to the high-frequency sampling unit, the distributed optical fiber sensing system, and the synchronous phasor measurement unit via a data bus, and has a built-in clock, achieving synchronization with the system master clock through the IEEE 1588 protocol.

[0063] S120. When any characteristic of the voltage change rate, current change rate, or vibration signal energy exceeds a set threshold, a fault detection event is triggered. The set threshold is an adaptive threshold based on the real-time operating status of the system. Specifically, the system calculates the mean and standard deviation of the sliding window (data from the most recent hour) of voltage, current, and vibration signal characteristics at each monitoring point every 15 minutes, and dynamically sets the threshold to "current mean + K times the standard deviation" (K is 3~5). This calculation is performed on the edge computing node, and the detection threshold is adaptively adjusted according to changes in operating conditions such as load level and environmental noise.

[0064] Step S2 includes at least steps S210-S230:

[0065] S210, the edge computing node preprocesses the acquired signals. Specifically, it uses wavelet transform modulus maxima to extract the wavefront features of current and voltage traveling waves. The wavelet transform uses the sym4 wavelet basis and has 8 decomposition levels. It uses short-time Fourier transform to extract the frequency domain energy features of the vibration signal. The window function of the short-time Fourier transform is the Kaiser window with a window length of 512 points and an overlap rate of 50%.

[0066] S220. Extract traveling wave features, vibration spectrum features, and temperature distribution features from the preprocessed signal to generate event summary data; specifically, the traveling wave features include the arrival time and amplitude of the traveling wave front, the vibration spectrum features include the frequency domain energy distribution and the dominant frequency component, and the temperature distribution features include the temperature gradient and the abnormal temperature rise region.

[0067] S230. Send the event summary data to the physical-data fusion engine via the communication network. The event summary data includes timestamps, monitoring point IDs, feature vectors, and confidence scores. The data format adopts the JSON standard, and the transmission protocol adopts MQTT.

[0068] Step S3 includes at least steps S310-S320:

[0069] The S310 physical-data fusion engine receives multimodal sensing data and performs message passing and node state updates based on a topology sensing graph neural network. Specifically, it constructs a line topology graph with the installation locations of sensing devices as nodes and electrical connections as edges. The node feature vectors include voltage surge amplitude, current traveling wave arrival time, vibration signal energy, and temperature change rate. The graph neural network adopts a graph convolutional network structure, containing 8 convolutional layers, each containing 512 neurons. The multilayer perceptron (MLP) used in its message passing function is a typical three-layer fully connected structure, consisting of: an input layer (dimension determined by the concatenated features). The computational layers are: dynamic computation, hidden layer (256 neurons, using ReLU activation function), and output layer (all nodes have the same feature dimension, no activation function is used). The message passing function is defined as follows: ,in For node features, The physical distance between nodes. The arrival time difference of the traveling wave is calculated based on the electromagnetic wave propagation speed, which is set to 2.98 × 10^8 m / s.

[0070] S320, Output the fault location probability distribution and confidence level; Specifically, the fault location probability distribution is generated through a Softmax layer, the confidence level is calculated based on the Bayesian inference method, the prior probability is set to 0.5, and the confidence threshold is set to 95%.

[0071] Step S4 includes at least steps S410-S420:

[0072] S410, the causal inference module analyzes the causal relationship of fault events based on transfer entropy and Granger causality tests. Specifically, the fault events at each monitoring point are time-series-based, and the transfer entropy between each pair of events is calculated. The time series window length for calculating the transfer entropy is 150 points, and the lag step is 2. The method for determining the statistical test threshold is as follows: before the system is put into operation, based on a large amount of historical normal operating conditions and known single-point fault scenarios, the transfer entropy values ​​between each pair of monitoring points are calculated to form a background distribution; the 99.5th percentile of this background distribution is taken as the threshold (typically 0.15 bits) to ensure that multiple causal chain judgments are not falsely triggered under normal fluctuations and single-point faults.

[0073] S420. Determine whether it is a single-point fault or a multi-point fault, and locate the root cause. Specifically, if the transfer entropy exceeds the threshold of 0.15 bits determined by statistical testing and the Granger test passes the significance level, then establish a directed causal edge to construct a fault causal directed graph, and identify the node with zero or minimum in-degree in the graph as the root fault point.

[0074] Step S5 includes at least steps S510-S540:

[0075] The S510 and digital twin evaluator generate a set of candidate isolation schemes based on fault location and root cause information. Specifically, the candidate isolation schemes are generated based on the line topology and fault location, including sectionalizing switch operation, tie switch switching, and load transfer schemes.

[0076] The S520 and digital twin system perform rapid power flow calculations and risk assessments for each isolation scheme. Specifically, a surrogate model based on dynamic mode decomposition and deep neural networks is used to reduce the system order from 5000 dimensions to 100 dimensions, and the simulation time is controlled within 120 milliseconds. The evaluation indicators include the power outage load range of 0MW~200MW, the equipment overload probability of 0%~100%, and the recovery time of 0s~300s.

[0077] The construction and training process of this proxy model is as follows:

[0078] Data preparation and DMD dimensionality reduction: Based on historical operating data and simulation data, a high-dimensional system state sequence matrix X (containing variables such as node voltage and branch power) is constructed. The Dynamic Mode Decomposition (DMD) algorithm is used to decompose X, extract the dominant dynamic modes of the system, and reduce the system order from 5000 dimensions to 100 dimensions, resulting in a low-dimensional representation matrix Z.

[0079] DNN Model Training: A deep neural network is trained using a low-dimensional representation Z and candidate isolation schemes (such as combinations of switching actions) as input, and risk assessment results obtained from simulation calculations (such as power outage load, equipment overload probability, and recovery time) as output labels. The specific structure of this DNN is as follows: input layer (100+ scheme feature dimensions), 4 fully connected hidden layers (512, 256, and 64 neurons respectively, all using the ReLU activation function), and output layer (3 neurons, corresponding to three risk assessment indicators, using linear activation functions).

[0080] Proxy Model Deployment: After training, this DNN model can be used as a proxy model for the digital twin. During risk assessment, the low-dimensional representation of the current system state (DMD) and the scheme to be evaluated are input into the DNN, and risk assessment metrics can be output within 120 milliseconds.

[0081] S530. The schemes are ranked based on a multi-objective optimization algorithm. The objective function is min[α·power failure load + β·equipment overload probability + γ·recovery time], where α, β, and γ are weight coefficients set according to the operation strategy. α is set to 0.5, β is set to 0.3, and γ is set to 0.2.

[0082] S540. Before implementing the isolation scheme, a safety constraint check is performed, including power flow constraint check that the node voltage deviation does not exceed 3%, stability constraint check that the power angle difference does not exceed 20 degrees, protection coordination constraint check that the action time difference is greater than 0.1 seconds, and operation sequence constraint check that the switch operation interval is greater than 2 seconds.

[0083] Step S6 includes at least steps S610-S620:

[0084] S610, the distributed consensus module determines the optimal isolation scheme based on the risk assessment results through a weighted voting mechanism; specifically, the voting weight of each decision-making unit is dynamically calculated based on the load capacity under its jurisdiction and its centrality in the line topology, with load capacity accounting for 80% and topology centrality accounting for 20%.

[0085] S620. The consensus threshold is set at 80%. When a scheme achieves a consensus level exceeding the threshold, it is determined to be the optimal isolation scheme. Specifically, the voting process adopts a multi-round negotiation mechanism, with a maximum of 3 rounds.

[0086] Step S7 includes at least steps S710-S720:

[0087] S710 The control system performs isolation operations. Specifically, it sends operation instructions to the switchgear through the control network. The instructions include the switch number, operation type, and execution time. The operation results are fed back to the system.

[0088] The S720 security and auditability layer records the complete fault handling process. Specifically, it adopts a distributed ledger based on a consortium blockchain, uses Byzantine fault tolerance as the consensus mechanism, and has a data backup interval of 1 second. The recorded event data includes at least the event trigger time, fault location results, causal inference graph, risk assessment results of each candidate solution, decision basis, and final execution result. To cope with high-frequency fault events, the system adopts the following optimization schemes: Data compression and aggregation: Within a 1-second backup interval, multiple related events (such as the same fault event at multiple monitoring points) are aggregated into a single transaction and uploaded to the chain, and a lossless compression algorithm is used to reduce the data volume; Layered storage: Only the key hashes and metadata of the event are stored on the chain, while the complete data is stored in an off-chain database, and its immutability is ensured through hash values; Scalable architecture: The blockchain network adopts a modular design, supporting the linear improvement of system throughput by adding accounting nodes or introducing sharding technology.

[0089] Step S8 includes at least steps S810-S830:

[0090] S810. Collect complete data after fault handling, perform data standardization and augmentation to obtain training dataset; specifically, collect raw sampling data, feature data and processing data, standardize using z-score method, and augment data by adding Gaussian noise, with the noise standard deviation set to 5% of the measured value.

[0091] S820. An incremental learning algorithm is used to train the topology-aware graph neural network and the digital twin agent model offline. Specifically, the training parameters of the topology-aware graph neural network are set to a learning rate of 0.0001, a batch size of 32, and a training cycle of 500 times. The digital twin agent model is updated using a transfer learning strategy.

[0092] S830. The updated model is validated using cross-validation, and the entire model update process is recorded through a security and auditability layer. Specifically, k-fold cross-validation (k=5) is used, and the validation indicators include positioning accuracy, false alarm rate, false alarm rate, and decision accuracy. The performance indicators of the new model are required to decrease by no more than 5% during the 24-hour trial operation period.

[0093] Example 3

[0094] Based on Example 1, the system operating environment is the same as in Example 1. The main description is of the system's self-optimization process after fault handling.

[0095] After completing the fault handling in Example 1, the system initiates the online learning mechanism. The learning process consists of three stages: data preparation, model training, and model validation.

[0096] During the data preparation phase, complete data for this failure event was collected, including 512GB of raw sampling data, 150MB of feature data, and 25MB of processing data. Data standardization was performed using the z-score method, scaling all features to a distribution with a mean of 0 and a standard deviation of 1.

[0097] The model training phase employed an incremental learning algorithm, conducted during nighttime hours when system load was below 30%. The training parameters for the topology-aware graph neural network were set as follows: learning rate 0.0001, batch size 32, and training epochs 500. The loss function used a weighted average squared error loss of 0.7 and a physical constraint loss of 0.3. The physical constraint loss included energy conservation constraints and propagation delay constraints. The graph neural network training employed the Adam optimizer, with a learning rate set to 0.0001 and fine-tuned parameters.

[0098] The surrogate model update for the digital twin evaluator employs a transfer learning strategy, preserving the basic network structure and fine-tuning only the last three fully connected layers. Training data augmentation utilizes Gaussian noise with a noise standard deviation set to 5% of the measured values. The surrogate model is trained using the Adam optimizer with a learning rate of 0.0005.

[0099] The model validation phase employs k-fold cross-validation, with k set to 5. Validation metrics include positioning accuracy, false alarm rate, false negative rate, and decision accuracy. The requirements are: positioning error less than 0.5 km, false alarm rate less than 1.5%, false negative rate less than 0.8%, and decision accuracy greater than 96%.

[0100] After the update, the system showed the following improvements in subsequent tests: the average positioning accuracy increased from 92.5% to 95.8%, the fault identification time was shortened from 95 milliseconds to 78 milliseconds, and the accuracy of isolation scheme evaluation increased from 88.3% to 93.6%.

[0101] The security and auditability layer records the entire model update process, including training data hashes, model parameter versions, validation results, and deployment time. The blockchain record shows the model update completion time and version number.

[0102] The system is equipped with a rollback mechanism. If any performance metric of the new model drops by more than 5% during the 24-hour trial period, it will automatically revert to the previous stable version. In this embodiment, the new model passed all tests and was put into normal operation.

[0103] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the present invention. These examples are merely for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or variations without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, variations, or combinations, or the direct application of the concept and technical solution of the present invention to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A high voltage circuit fault location and automatic isolation system, characterized by, The application relates to a power grid fault diagnosis system based on multi-modal perception, physical-data fusion and risk assessment, which comprises a multi-modal perception layer, a physical-data fusion engine and a risk assessment and isolation decision layer. The multi-modal perception layer is arranged at an outgoing terminal of a high-voltage power transmission and transformation device, a line section point and an important load access point, and comprises a high-frequency sampling unit, a distributed optical fiber sensing system, a synchronous phasor measurement unit and an edge computing node; the high-frequency sampling unit collects voltage and current signals of the high-voltage power transmission and transformation device through electrical connection, and the sampling rate is 1MHz-10MHz; the distributed optical fiber sensing system is arranged in parallel with the high-voltage power transmission and transformation device through an optical fiber composite overhead ground wire or an optical fiber laid along a cable trench, and vibration signals and temperature signals are collected; the synchronous phasor measurement unit is connected with the high-voltage power transmission and transformation device through a voltage transformer and a current transformer, and three-phase voltage and current phasors are collected; the edge computing node is connected with the high-frequency sampling unit, the distributed optical fiber sensing system and the synchronous phasor measurement unit through a data bus, and is used for pre-processing the collected voltage and current signals, vibration signals and temperature signals and three-phase voltage and current phasors, and outputting multi-modal perception layer data. The physical-data fusion engine is connected with the multi-modal perception layer through a communication network, receives the multi-modal perception layer data, and comprises a physical model library, a topological perception graph neural network module and a causal inference module; the input of the topological perception graph neural network module is a line topological structure and multi-modal perception data, and the output is a fault position probability distribution and a confidence degree thereof; The causal inference module is used for analyzing a causal relationship in a multi-fault scenario based on a transfer entropy algorithm and in combination with a Granger causality test; The risk assessment and isolation decision layer is connected with the physical-data fusion engine through a control network, receives the fault position probability distribution and the causal relationship, and outputs an isolation decision; the risk assessment and isolation decision layer comprises a digital twin evaluator, an isolation strategy library and a distributed consensus module; A security and audit layer is connected with all layers of the system through a data interface, adopts a distributed ledger based on a consortium chain, and is used for recording event data of a whole process of fault processing and providing an audit function.

2. A high voltage circuit fault location and automatic isolation system according to claim 1, characterized in that, An internal clock of the edge computing node is synchronized with a system master clock through an IEEE1588 protocol and by adopting GPA and a Beidou dual-redundancy clock source; The preprocessing operation of the edge computing node comprises the following steps: wave head characteristics of current and voltage traveling waves are extracted by adopting a wavelet transform modulus maximum value method, and frequency domain energy characteristics of vibration signals are extracted by adopting a short-time Fourier transform.

3. A high voltage circuit fault locating and automatic isolating system according to claim 1, characterized in that, The topology-aware graph neural network module adopts a graph convolution network structure, contains 6-8 convolution layers, each convolution layer contains 128-256 neurons, and the message passing function is defined as: wherein is the node feature, is the physical distance between nodes, is the difference in arrival time of the traveling wave calculated based on the electromagnetic wave propagation speed.

4. The high voltage circuit fault locating and automatic isolating system of claim 1, wherein, The causal inference module is based on a causal discovery algorithm of transfer entropy, is used for sequencing fault events of monitoring points according to time sequences, calculates transfer entropy between two events, establishes a directed causal edge if the transfer entropy exceeds a threshold value determined through a statistical test and a Granger causality test is passed, constructs a fault causal directed graph, and identifies a node with zero or minimum in-degree in the graph as a root fault point.

5. The system of claim 1, wherein, The digital twin evaluator adopts an agent model based on dynamic modal decomposition and a deep neural network, is used for power flow calculation and risk assessment of a candidate isolation scheme, and evaluation indexes include outage load, device overload probability and expected recovery time.

6. The high voltage circuit fault locating and automatic isolating system of claim 1, wherein, The isolation policy library stores specific rules of the minimal cut set policy, the regional isolation policy and the selective isolation policy; the distributed consensus module adopts a weight voting mechanism, the voting weight of each decision unit is dynamically calculated according to the load capacity under its jurisdiction and the centrality in the line topology structure, and the consensus threshold is set to 80%.

7. The system of claim 1, wherein the system further comprises a plurality of switches, each of the plurality of switches being connected to one of the plurality of circuit breakers. The security and audit layer adopts practical Byzantine fault tolerance as the consensus mechanism, the data backup interval is 1 second, and the high-frequency event data is processed by combining aggregation compression and off-chain storage; the event data recorded by the layer includes event trigger time, fault location result, causal inference graph, risk assessment result of each candidate scheme, decision basis and final execution result.

8. A method applied to a high-voltage circuit fault location and automatic isolation system, characterized in that, The method comprises the following steps: S1: Real-time acquisition of multi-modal signals of high-voltage power transmission and transformation equipment through a multi-modal perception layer, triggering a fault detection event when any feature in voltage rate of change, current rate of change or vibration signal energy exceeds a set threshold; S2: The edge computing node pre-processes the collected signals, extracts traveling wave features, vibration spectrum features and temperature distribution features, and generates event summary data; S3: The physical-data fusion engine receives multi-modal perception data, performs message passing and node state updating based on a topology perception graph neural network, and outputs fault location probability distribution and confidence; S4: The causal inference module analyzes the causal relationship of the fault event based on transfer entropy analysis and combined with Granger causality test, judges whether it is a single-point fault or a multi-point fault, and locates the root cause; S5: The digital twin evaluator evaluates the risk indicators of multiple isolation schemes in parallel based on fault location and root cause information; S6: The distributed consensus module determines the optimal isolation scheme through a weight voting mechanism based on the risk assessment results; S7: The control system executes the isolation operation, while the security and audit layer records the complete fault handling process; S8: The system updates the topology perception graph neural network and the digital twin agent model based on the real data of the handling results.

9. The method of claim 8, wherein, The specific process of fault location in S3 is as follows: S31: Constructing a line topology graph, wherein the nodes represent the installation positions of sensing devices, and the edges represent electrical connection relationships; S32: Extracting node feature vectors, including voltage mutation amplitude, current traveling wave arrival time, vibration signal energy and temperature change rate; S33: Message passing through a graph neural network, the message function including electromagnetic wave propagation delay calculation; S34: The output layer generates fault location probability distribution, and the confidence is calculated based on the Bayesian inference method.

10. The method of claim 8, wherein, The specific process of risk assessment in S5 is as follows: S51: Generating a candidate isolation scheme set based on the fault location result; S52: The digital twin system performs fast power flow calculation on each isolation scheme; S53: Calculating the risk indicators of each scheme, including outage load range 0MW~50MW, device overload probability 0%~100%, and recovery time 0s~300s; S54: The scheme is sorted based on a multi-objective optimization algorithm, and the objective function is min [α· outage load + β· equipment overload probability + γ· recovery time], wherein α, β, and γ are weight coefficients set according to the operation strategy; before the isolation scheme is executed, a safety constraint check is performed, including a power flow constraint check, a stability constraint check, a protection coordination constraint check, and an operation sequence constraint check.

Images