Power distribution network fault location system based on ai driving and power line sensor fusion

Abstract

The application belongs to the technical field of power distribution network fault diagnosis and positioning, and discloses a power distribution network fault positioning system based on AI driving and power line sensor fusion, which comprises: an offline cloud model optimization layer, which optimizes and prunes a random forest model through power line carrier communication constraints and power failure physical constraints to generate a lightweight pre-training model; an edge perception layer, which is used for real-time collection of power line electrical signals and generation of relative topological distance identification between nodes; an edge computing layer, which performs fault detection, feature extraction and local forward reasoning based on the collected electrical signals, and outputs fault pre-classification results; and a cloud positioning engine layer, which is used for receiving the relative topological distance identification to dynamically correct the real-time topological relationship, and based on the fault pre-classification results, fault traveling wave data and the corrected real-time topological relationship, solving to obtain the accurate geographic coordinates of the fault point. The application improves the generalization ability and reliability of fault pre-classification, simplifies the model volume and reduces the reasoning delay.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network fault diagnosis and location technology, specifically involving the integrated application of artificial intelligence optimization algorithms with power line sensing and power line carrier communication (PLC), and particularly involving a power distribution network fault location system based on AI-driven and power line sensing fusion, which is suitable for power distribution network fault pre-classification and high-precision location scenarios with edge computing and cloud collaboration. Background Technology

[0002] As the final link in the power system, the timeliness and accuracy of fault location in the distribution network directly affect the reliability of power supply. Existing distribution network fault location systems mainly rely on single electrical quantities (such as voltage and current) or traditional traveling wave location technology, which have the following technical shortcomings:

[0003] (1) Fault location is easily affected by dynamic changes in network topology, and the inaccuracy of traditional ledger data leads to large location errors;

[0004] (2) The application of machine learning algorithms such as random forest in fault classification is usually based on data statistical characteristics optimization only, without fully considering the low bandwidth and high packet loss characteristics of power line carrier communication in distribution network, resulting in excessive communication volume and unstable transmission when the model is sent; at the same time, without combining the strong physical constraints of power faults (such as fault transient attenuation law and impedance change threshold), the model output may show results that violate physical common sense and have insufficient generalization ability.

[0005] (3) Edge terminals have limited computing power and storage resources. Traditional random forest models have high redundancy, making it difficult to balance lightweight deployment and classification accuracy. Fault pre-classification has high latency and poor real-time performance.

[0006] (4) The lack of collaborative design between model optimization and communication transmission, the pursuit of classification accuracy while ignoring the communication constraints of the actual deployment environment of the distribution network, resulted in low practicality of the technical solution and poor engineering feasibility.

[0007] Therefore, how to construct a model optimization method that integrates communication constraints and physical constraints for the specific scenarios of power distribution networks, so as to achieve lightweight deployment at the edge, high-precision fault pre-classification and stable location, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] This invention addresses the technical shortcomings of existing power distribution network fault location systems by proposing a system based on AI-driven and power line sensing fusion. Specifically, it is a power distribution network fault pre-classification and location system based on a communication-physical dual-constraint model optimization. The specific objectives are as follows: 1) To construct a model optimization strategy adapted to the characteristics of power line carrier communication in power distribution networks, solving the stability issues of model optimization and delivery in low-bandwidth, high-packet-loss scenarios; 2) To integrate physical constraints of power faults with random forest optimization processes, ensuring that the model output conforms to the physical laws of faults and improving the generalization ability and reliability of fault pre-classification; 3) To design a hierarchical collaborative architecture of "cloud-based offline optimization - edge-based lightweight inference," achieving model size reduction and inference latency reduction while ensuring fault pre-classification accuracy; 4) To achieve synergy between fault pre-classification and accurate cloud-based location, combined with real-time topology correction, eliminating errors in ledger data and improving the overall accuracy and stability of power distribution network fault location.

[0009] This invention provides a power distribution network fault location system based on the fusion of AI-driven and power line sensing, comprising:

[0010] The offline cloud model optimization layer is used to optimize and prune the random forest model based on historical fault data of the distribution network, through power line carrier communication constraints and power fault physical constraints, to generate a lightweight pre-trained model.

[0011] The edge sensing layer is deployed at key nodes of the power distribution network to collect electrical signals from power lines in real time and generate relative topological distance identifiers between nodes.

[0012] The edge computing layer is deployed within the terminal device where the edge perception layer is located. It is used to load the lightweight pre-trained model and perform fault detection, feature extraction and local forward inference based on the collected electrical signals, and output fault pre-classification results.

[0013] The cloud-based positioning engine layer is used to receive the relative topological distance identifier to dynamically correct the real-time topological relationship, and to calculate the precise geographic coordinates of the fault point based on the fault pre-classification result, the fault traveling wave data collected by the edge perception layer and uploaded by the edge computing layer, and the corrected real-time topological relationship.

[0014] The offline cloud model optimization layer, edge perception layer, edge computing layer and cloud positioning engine layer interact with each other via a power line carrier communication link.

[0015] Preferably, the offline cloud model optimization layer includes:

[0016] The power fault sample library module is used to store historical fault data labeled with fault categories and physical characteristics, including fault transient attenuation coefficient, high-frequency impedance change value and fault duration.

[0017] The power edge-adaptive physical constraint pruning random forest optimization module is used to optimize the voting weights of the decision tree under power line carrier communication constraints using the particle swarm optimization algorithm, and prune the decision tree according to the accuracy criterion and physical compliance criterion to generate a lightweight pre-trained model.

[0018] The model distribution module is used to distribute the lightweight pre-trained model to the edge computing layer via a power line carrier communication link.

[0019] Preferably, when the power edge-adaptive physical constraint pruning random forest optimization module optimizes the voting weights of the decision tree for power line carrier communication constraints using the particle swarm optimization algorithm, the specific configuration is as follows:

[0020] The communication environment is simulated using a power line carrier standard channel model for distribution networks, with channel packet loss rate as the optimization criterion.

[0021] The inertia weight in the particle swarm optimization algorithm is adaptively adjusted according to the channel packet loss rate. When the channel packet loss rate is higher than a preset threshold, the first inertia weight is used, and when the channel packet loss rate is not higher than the preset threshold, the second inertia weight is used.

[0022] A sparse particle transport strategy is adopted. After a preset number of iterations, only the particles with the highest objective function values ​​are retained to participate in subsequent iterations, so as to reduce the amount of communication during the optimization process.

[0023] The optimization objective of the particle swarm optimization algorithm is a weighted sum of model classification accuracy, physical compliance, and model sparse entropy.

[0024] Preferably, when the power edge-adaptive physical constraint pruning random forest optimization module prunes the decision tree using accuracy and physical compliance criteria, the specific configuration is as follows:

[0025] The accuracy criterion is that the classification accuracy of a single decision tree is not lower than the average accuracy of the forest.

[0026] The physical compliance criterion is that the physical compliance degree output by a single decision tree is not lower than a preset physical compliance threshold.

[0027] The physical compliance is determined based on the ratio of the number of samples that violate physical constraints to the total number of verification samples in the verification sample set. The determination of the violation of physical constraints is based on the mapping relationship between the fault category and the range of physical parameters.

[0028] Decision trees that simultaneously satisfy both the accuracy criterion and the physical compliance criterion are retained, while decision trees that do not satisfy either criterion are pruned, thereby generating a lightweight pre-trained model.

[0029] Preferably, the preset physical compliance threshold is 0.95; the formula for calculating the physical compliance degree is:

[0030]

[0031] in, The number of verification samples that violate physical constraints. The total number of validation samples is represented by Phys, which represents the physical compliance score.

[0032] Preferably, the determination of physical constraint violation is based on the mapping relationship between fault category and physical parameter range. When the physical parameter range corresponding to the fault category output by the lightweight pre-trained model does not match the measured physical feature quantity in the verification sample, it is determined to be a violation of physical constraint.

[0033] The lightweight pre-trained model uses a weighted voting mechanism to output fault classification results. Specifically, the weighted voting mechanism is as follows: based on the voting weights of each decision tree obtained after optimization by the particle swarm optimization algorithm, the classification outputs of each decision tree are weighted and summed, and the fault category corresponding to the maximum value in the weighted summation result is taken as the fault pre-classification result output by the model.

[0034] Preferably, the edge-aware layer includes:

[0035] The signal coupling and acquisition unit is directly connected to the power line and is used to acquire power frequency electrical quantities and fault traveling wave signals in real time.

[0036] A cross-band cognitive communication unit is used to adaptively select the operating frequency based on orthogonal frequency division multiplexing technology to construct a power line carrier communication link;

[0037] The carrier ranging unit is used to obtain the electrical distance between this terminal and adjacent terminals based on the time of arrival principle, and generate the relative topological distance identifier between the nodes.

[0038] Preferably, the edge computing layer includes:

[0039] The mutation event detection submodule is used to analyze the acquired voltage and current signals using a cumulative sum algorithm, capture signal mutation points, and trigger high-speed waveform recording.

[0040] The impact feature extraction submodule is used to extract the absolute impact height, impact width, fall velocity, and multi-point fall degree of the fault waveform, and generate a standardized fault feature vector.

[0041] The lightweight fault pre-classification submodule is used to load the lightweight pre-trained model, perform local forward inference on the standardized fault feature vector, and output the fault type using the weighted voting mechanism corresponding to the lightweight pre-trained model.

[0042] Preferably, the cloud positioning engine layer includes:

[0043] The real-time topology construction module is used to receive the relative topology distance identifier and dynamically correct the real-time topology relationship of the distribution network in combination with the graph theory tree search algorithm;

[0044] The multi-terminal traveling wave positioning module is used to output the geographical coordinates of the fault point by combining the time difference of the fault traveling wave arriving at each edge sensing layer terminal with the path distance in the real-time topology relationship and through the multi-terminal traveling wave ranging equation set.

[0045] Preferably, the system is configured to perform collaborative work in the following phases:

[0046] Offline optimization phase: The offline cloud model optimization layer is based on the power fault sample library, uses the power line carrier standard channel model to complete particle swarm weight optimization, and performs physical constraint pruning according to the mapping relationship between fault category and physical parameter range to generate a lightweight pre-trained model, which is then sent to the edge computing layer through the power line carrier communication link;

[0047] During normal operation: The edge sensing layer collects electrical signals from power lines in real time and generates relative topological distance identifiers between nodes, which are then uploaded to the cloud positioning engine layer for dynamic correction of real-time topological relationships;

[0048] Fault Triggering Phase: The edge computing layer identifies signal abrupt changes and triggers high-speed fault waveform recording, extracts key features of the fault waveform, and generates a standardized fault feature vector;

[0049] Edge pre-classification stage: The edge computing layer loads the lightweight pre-trained model, performs local forward inference on the standardized fault feature vector, outputs the fault pre-classification result through weighted voting, and encapsulates the fault pre-classification result and fault waveform data and uploads it to the cloud positioning engine layer.

[0050] In the cloud-based precise positioning stage: the cloud positioning engine layer corrects the real-time topological relationship based on the relative topological distance identifier, and integrates the real-time topological path distance with the arrival time difference of the fault traveling wave to solve the traveling wave ranging equation and output the precise geographical coordinates of the fault point.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. Dual-constraint collaborative optimization enhances model robustness and classification accuracy: This invention, for the first time, integrates power line carrier communication constraints and power fault physical constraints into the entire random forest optimization process. Through inertial weight adjustment adapted to packet loss rate, sparse particle transmission strategy, and compliance pruning guided by fault category-physical parameter mapping table, the model is adapted to both the distribution network communication environment and the physical laws of faults. Experimental verification shows that on actual distribution network fault datasets, the fault pre-classification accuracy is improved by 3.2%~5.1% compared to traditional random forests, and the proportion of outputs violating physical constraints is reduced to below 8%.

[0053] 2. Extremely streamlined model, adapted for edge embedded deployment: Through physical constraint dual-criteria pruning (single tree accuracy ≥ forest average accuracy + physical compliance ≥ 0.95), the model size is reduced by 50%~60% compared to the initial 100-tree random forest, and the edge inference latency is shortened by more than 40%. It is fully adapted to the computing power and storage constraints of power distribution network edge terminals, and solves the core contradiction of "lightweight and high precision".

[0054] 3. Significantly reduced communication volume, adapting to the low bandwidth characteristics of PLC: The sparse particle transmission strategy (selecting the top 10% of the best particles every 10 iterations) reduces the communication volume of the model optimization process by 58%~67%, effectively solving the problem of limited bandwidth of power line carrier communication in power distribution networks and improving the stability and efficiency of model transmission.

[0055] 4. The layered architecture has a clear logic and strong engineering feasibility: complex model training and dual-constraint optimization are completed centrally in the cloud, while the edge only performs forward inference, avoiding the problem of insufficient computing power at the edge terminal; the functions of each layer of modules are clear, the data flow is smooth, and all key parameters are determined based on the actual measured data of the power distribution network and industry standards. The technical solution fits the actual engineering situation and has no design that violates physical common sense or is impossible to implement.

[0056] 5. Stable positioning accuracy and outstanding practicality: The fault pre-classification results provide prior information for cloud positioning. Combined with real-time topology correction and multi-terminal traveling wave positioning, it effectively eliminates positioning errors caused by inaccurate ledger data. The fault point positioning accuracy is improved by more than 15% compared with traditional methods. Moreover, it can maintain stable performance even in harsh communication environments. It is suitable for actual deployment scenarios where distribution network terminals cannot access the public network. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the system architecture of a power distribution network fault location system based on the fusion of AI-driven and power line sensing according to an embodiment of the present invention.

[0058] Figure 2 This is a flowchart of the offline model optimization process according to an embodiment of the present invention. Detailed Implementation

[0059] The following detailed implementation of the distribution network fault location system based on AI-driven and power line sensing fusion of the present invention is described in detail with reference to specific embodiments. The embodiments of the present invention are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0060] This invention addresses the technical problems of existing distribution network fault location systems, such as reliance on single electrical quantities, susceptibility to network topology changes, difficulty in balancing lightweight design and classification accuracy in edge-end fault pre-classification, and the fact that traditional random forest optimization is typically based on statistical data characteristics and does not fully consider the low bandwidth and high packet loss characteristics of power line carrier communication in distribution networks, as well as the strong physical constraints of faults. For these application scenarios, a model optimization method integrating both communication and physical constraints is proposed. It employs a layered collaborative architecture of "cloud-based offline customized optimization, power line carrier adaptive transmission, and lightweight edge-end inference" to improve fault location stability, reduce edge computing load, and achieve reliable physical location of distribution network fault points. The functions and data flows of each module in this system are fully disclosed in the specification and accompanying drawings, and it does not contain any technical solutions that are impossible to implement or violate physical principles.

[0061] Combined with appendix Figures 1-2 As shown, this invention provides a distribution network fault location system based on the fusion of AI-driven and power line sensing, specifically a distribution network fault pre-classification and location system based on a communication-physical dual-constraint model optimization. The system consists of an offline cloud model optimization layer 110, an edge sensing layer 120, an edge computing layer 130, and a cloud positioning engine layer 140. Each layer interacts with the model via a power line carrier communication link, making it suitable for deployment scenarios where distribution network terminals cannot access the public network. Specifically, the offline cloud model optimization layer 110, the edge sensing layer 120, the edge computing layer 130, and the cloud positioning engine layer 140 interact via a power line carrier communication link. Details are as follows:

[0062] Offline cloud model optimization layer 110 is used to optimize and prune the random forest model based on historical fault data of the distribution network through power line carrier communication constraints and power fault physical constraints, and generate a lightweight pre-trained model.

[0063] Step 1: Construct offline cloud model optimization layer 110

[0064] The offline cloud model optimization layer 110 is deployed on the cloud main station to complete the offline training, constraint optimization and model simplification of the fault classification model, and outputs a lightweight pre-trained model to the downstream edge computing layer 130, including a power fault sample library module, a power edge-adaptive physical constraint pruning random forest optimization module and a model distribution module.

[0065] Step 1.1, Power Fault Sample Library Module

[0066] The power fault sample library module stores historical fault data of the distribution network labeled with fault categories and physical characteristics, covering transient faults, permanent faults, and high-resistance grounding fault samples. The samples are divided into a training set and a hold-out validation set, with validation samples taken from the hold-out validation set of the fault sample library (the sample size accounts for 20% of the total samples). The sample labels include fault category labels, as well as three types of physical characteristics: fault transient attenuation coefficient, high-frequency impedance change value, and fault duration.

[0067] The range of physical parameters is determined based on the on-site measured statistical values ​​of the distribution network and DL / T 1870-2018 "Technical Specification for Power System Grid-Source Coordination". The mapping relationship between fault categories and physical parameters is shown in Table 1 below:

[0068] Table 1

[0069] Fault Category Transient decay coefficient Impedance mutation threshold Fault duration Transient failure 0.6~0.9 ≥50Ω 10ms~100ms Permanent failure 0.1~0.4 ≥50Ω >100ms High-resistance grounding 0.4~0.6 20Ω~50Ω 10ms~5s

[0070] Step 1.2: Power Edge Adaptive Physical Constraint Pruned Random Forest (PE-APCP-RF) Optimization Module

[0071] The power edge-adaptive physical constraint-pruned random forest optimization module is based on a basic random forest. It uses a particle swarm optimization algorithm to optimize the voting weights of the decision tree under power line carrier communication constraints, and prunes the decision tree according to accuracy and physical compliance criteria to generate a lightweight pre-trained model adapted to power distribution network scenarios. The optimization process is as follows:

[0072] Step 1.2.1: Initial Model Construction

[0073] Generate a basic random forest containing 100 decision trees.

[0074] Step 1.2.2: Adaptive Sparse Particle Swarm Optimization Weight Configuration for Power Line Carrier Channel

[0075] Furthermore, when the power edge-adaptive physical constraint pruning random forest optimization module optimizes the voting weights of the decision tree for power line carrier communication constraints using the particle swarm optimization algorithm, the specific configuration is as follows: A standard channel model of the power line carrier in the distribution network is used to simulate the communication environment, with the channel packet loss rate as the optimization criterion; the inertia weights in the particle swarm optimization algorithm are adaptively adjusted according to the channel packet loss rate; when the channel packet loss rate is higher than a preset threshold, the first inertia weight is used, and when the channel packet loss rate is not higher than the preset threshold, the second inertia weight is used; a sparse particle transmission strategy is adopted, where after each preset number of iterations, only the particles with the highest objective function values ​​are retained to participate in subsequent iterations, in order to reduce the communication volume during the optimization process; the optimization objective of the particle swarm optimization algorithm is the weighted sum of model classification accuracy, physical compliance, and model sparse entropy. Specifically:

[0076] (1) Power line carrier communication channel model: The standard channel model of power line carrier in distribution network is adopted, which includes multipath fading characteristics and Bernoulli random packet loss model. The packet loss rate covers the gradient range of 0% to 30%; 15% is the typical packet loss rate threshold for severe communication scenarios in distribution network (based on the statistics of 120 days of on-site communication data of three distribution network areas of a power supply company from 2022 to 2024, the scenarios with a packet loss rate of more than 15% accounted for 82.5% of all abnormal communication events).

[0077] (2) Particle Swarm Optimization (PSO) Particle Definition: A particle represents the weight vector of a decision tree, and each dimension of the vector corresponds to the voting weight of a decision tree.

[0078] (3) Particle swarm update formula:

[0079] Speed ​​update formula:

[0080] Position update formula:

[0081] Symbol definition: For the first The first particle per iteration speed; Inertial weights; , For learning factors; , A random number within the interval [0,1]; For the first The best position in the history of each particle; For the first The first particle Next iteration position; The optimal position for the entire population;

[0082] (4) Inertia weight adaptive strategy: When the channel packet loss rate > 15%, When the channel packet loss rate is ≤15%, (Based on comparative experiments with preset packet loss rates ranging from 0% to 30%, it was determined that when the packet loss rate is >15%) Comparison (The accuracy on the validation set was improved by 3.2% to 5.1%). The number of particles is 50, and the maximum number of iterations is 100.

[0083] (5) Sparse particle transport strategy: Particle selection is performed once every 10 iterations, and the optimal particle is the one with the highest objective function value. Only the best particles are transmitted for subsequent iterations (experiments have verified that at this screening interval and ratio, the communication volume is reduced). Model accuracy loss );

[0084] (6) Optimize the objective function:

[0085] ;

[0086] Where: Acc is the model classification accuracy; Phys is the physical compliance; Sparse is the model sparse entropy.

[0087] Physical compliance calculation formula: ;

[0088] in: The number of verification samples that violate physical constraints; This represents the total number of validation samples;

[0089] Physical constraint determination rules: Model outputs fault category → Matches corresponding physical parameter range → Compares and verifies the measured physical quantity of the sample → If the measured value exceeds the corresponding range, it is determined to be a violation of constraint.

[0090] The criterion for determining a violation of physical constraints is the mapping relationship between the fault category and the range of physical parameters. When the range of physical parameters corresponding to the fault category output by the lightweight pre-trained model does not match the measured physical feature quantity in the validation sample, it is determined to be a violation of physical constraints.

[0091] Formula for calculating sparse entropy in a model: Sparse ;

[0092] in: For the first The voting weights of each decision tree;

[0093] (7) Model Weighted Voting Rule: The lightweight pre-trained model uses a weighted voting mechanism to output fault classification results. Specifically, the weighted voting mechanism involves: based on the voting weights of each decision tree obtained after optimization using the particle swarm optimization algorithm, a weighted sum is calculated on the classification outputs of each decision tree, and the fault category corresponding to the maximum value in the weighted sum is taken as the final fault pre-classification result output by the model. Specifically, the formula for the weighted voting mechanism is: .

[0094] in, Output the fault category for the model; For the first The classification output of each decision tree; For the first The voting weights of each decision tree.

[0095] Step 1.2.3: Physical constraint dual-criteria pruning

[0096] The power edge-adaptive physical constraint pruning random forest optimization module prunes decision trees using accuracy and physical compliance criteria. Specifically, the accuracy criterion is that the classification accuracy of a single decision tree is not lower than the forest's average accuracy; the physical compliance criterion is that the physical compliance score of a single decision tree output is not lower than a preset physical compliance threshold (preferably 0.95). The physical compliance score is determined based on the ratio of the number of samples violating physical constraints in the validation sample set to the total number of validation samples. The determination of physical constraint violation is based on the mapping relationship between fault categories and physical parameter ranges. Decision trees that simultaneously meet both the accuracy and physical compliance criteria are retained, while decision trees that do not meet either criterion are pruned, thereby generating a lightweight pre-trained model. Details are as follows:

[0097] Pruning follows a precision criterion and a physical compliance criterion: the precision criterion is that the classification accuracy of a single decision tree is no less than the forest average accuracy; the physical compliance criterion is the physical compliance of the output of a single decision tree. (Based on the distribution statistics of samples violating physical constraints in the fault sample library, setting this threshold can make the filtering rate of violation samples ≥92%); Decision trees that meet both criteria are retained, and decision trees that do not meet the criteria are pruned; On the test set constructed based on the historical fault data of a certain distribution network from 2022 to 2024 (including 1200 samples of 3 types of faults), the optimized model volume is reduced by about 50%~60% compared with the original random forest of 100 trees. This proportion fluctuates by ±5% depending on the initial forest size and the setting of the pruning threshold.

[0098] Step 1.3 Model Distribution Module

[0099] The model delivery module distributes the optimized, lightweight pre-trained model to the edge computing layer terminal via a power line carrier communication link. The edge terminal only loads the model and performs inference, without participating in model training and optimization calculations. In the communication scenario corresponding to the above test set, the communication volume during the model delivery process is reduced by approximately 50% to 70% compared to not using a sparse transmission strategy. This value fluctuates depending on the feature dimension of the fault samples and the simplification of the model.

[0100] Edge sensing layer 120 is deployed at key nodes of the power distribution network to collect power line electrical signals in real time and generate relative topological distance identifiers between nodes;

[0101] Step 2: Construct the edge-aware layer 120

[0102] The edge sensing layer 120 is a multimodal power line sensing integrated terminal, deployed at key nodes of the power distribution network such as substation outgoing lines, ring network cabinets, branch boxes, and distribution transformers. It includes a signal coupling and acquisition unit, a cross-band cognitive communication unit, and a carrier ranging unit.

[0103] Step 2.1, Signal Coupling and Acquisition Unit

[0104] The signal coupling and acquisition unit is directly connected to the power line to collect power frequency electrical quantities and fault traveling wave signals in real time. Based on the impedance testing principle, it senses the changes in the high-frequency impedance characteristics of the line, providing raw data support for fault feature extraction and pre-classification in the edge computing layer.

[0105] Step 2.2, Cross-band cognitive communication unit

[0106] The cross-band cognitive communication unit is based on orthogonal frequency division multiplexing technology. Its typical operating frequency band is 150kHz~12MHz. It adaptively selects the optimal operating frequency to build a stable power line carrier communication link, enabling model distribution, topology information uploading, fault pre-classification results and waveform data transmission between the edge and the cloud.

[0107] Step 2.3, Carrier Ranging Unit

[0108] The carrier ranging unit performs ranging calculations based on the time of arrival principle to obtain the electrical distance between this terminal and adjacent terminals; it is used to generate relative topological distance identifiers between nodes, providing real-time topological structure references for cloud-based traveling wave positioning; the fault accurate ranging function is completed by the multi-terminal traveling wave positioning module of the cloud positioning engine layer.

[0109] The edge computing layer 130 is deployed in the terminal device where the edge perception layer 120 is located. It is used to load the lightweight pre-trained model and perform fault detection, feature extraction and local forward inference based on the collected electrical signals, and output the fault pre-classification result.

[0110] Step 3: Construct the edge computing layer 130

[0111] The edge computing layer 130 is an embedded fault feature processing module, which is embedded in the multimodal power line sensing integrated terminal and includes a sudden event detection submodule, an impact feature extraction submodule, and a lightweight fault pre-classification submodule.

[0112] Step 3.1, Mutation Event Detection Submodule

[0113] The mutation event detection submodule uses the Cumulative Sum (CUSUM) algorithm to monitor the real-time acquired voltage and current signals online and calculate the cumulative deviation of the signal sample value relative to the mean value under normal conditions. When the cumulative sum of the cumulative deviation exceeds a preset threshold, it is determined to be a signal mutation point caused by a fault, and then a high-speed waveform recording operation is triggered to record the complete fault waveform data before and after the mutation point, providing data support for subsequent feature extraction.

[0114] Step 3.2, Impact Feature Extraction Submodule

[0115] After a fault is triggered, the impact feature extraction submodule extracts four key features of the fault waveform: absolute impact height, impact width, fall speed, and multi-point fall degree, and generates a standardized fault feature vector to provide input data for lightweight fault pre-classification.

[0116] Step 3.3, Lightweight Fault Pre-classification Submodule

[0117] The lightweight fault pre-classification submodule loads the offline optimized Power Edge Adaptive Physical Constraint Pruned Random Forest (PE-APCP-RF) pre-trained model, performs local forward inference on the standardized fault feature vector, and outputs the fault type (transient fault, permanent fault, high-resistance grounding) using the weighted voting mechanism corresponding to the lightweight pre-trained model. The prediction results and fault waveform data are encapsulated and uploaded to the cloud positioning engine layer to provide prior information on fault type for accurate cloud positioning.

[0118] The cloud positioning engine layer 140 is used to receive the relative topological distance identifier to dynamically correct the real-time topological relationship, and calculate the precise geographic coordinates of the fault point based on the fault pre-classification result, the fault traveling wave data collected by the edge perception layer and uploaded by the edge computing layer, and the corrected real-time topological relationship.

[0119] Step 4: Construct the cloud-based positioning engine layer 140

[0120] The cloud-based positioning engine layer 140 is deployed on the main station system and is a topology-ranging fusion positioning platform, including a real-time topology construction module and a multi-terminal traveling wave positioning module.

[0121] Step 4.1, Real-time Topology Construction Module

[0122] The real-time topology construction module receives the relative topological distance identifiers between nodes uploaded by each edge terminal. Using each terminal as a node and the relative topological distance as the edge weight, it constructs an initial topology graph of the distribution network. The initial topology graph is traversed and verified using the minimum spanning tree algorithm or breadth-first search algorithm in graph theory. It identifies and corrects topological connection deviations caused by errors in the ledger data, and dynamically generates the real-time topological relationship of the distribution network "station-line-transformer", providing accurate path distance references for multi-terminal traveling wave positioning.

[0123] Step 4.2 Multi-terminal traveling wave positioning module

[0124] The multi-terminal traveling wave positioning module is based on the time difference of the fault traveling wave arriving at each terminal, combined with the path distance of the real-time topology, and substitutes it into the standard multi-terminal traveling wave ranging equation set to complete the joint solution, eliminate the error of the ledger data, and output the geographical coordinates of the fault point.

[0125] Standard multi-terminal traveling wave ranging equation: ;

[0126] in: , For the first , Terminal coordinates; The coordinates of the fault point; The propagation speed of the fault traveling wave; , For the fault traveling wave to reach the first , The time of each terminal.

[0127] Preferably, the system configuration in this embodiment of the invention is to perform the following stages of collaborative work:

[0128] (1) Offline optimization stage: The offline cloud model optimization layer 110 is based on the power fault sample library and the reserved validation set. It uses the power line carrier standard channel model to complete the particle swarm weight optimization. It performs physical constraint pruning according to the fault category-physical parameter mapping relationship to generate a lightweight power edge-adaptive physical constraint pruning random forest model, which is then sent to the edge computing layer 130 through the power line carrier communication link.

[0129] (2) Normal operation phase: The edge perception layer 120 collects power line electrical signals in real time, and the carrier ranging unit generates relative topological distance identifiers between nodes, which are then uploaded to the cloud positioning engine layer 140 for dynamic correction of real-time topological relationships.

[0130] (3) Fault Triggering Stage: The sudden event detection submodule of the edge computing layer 130 identifies signal sudden changes and triggers high-speed fault waveform recording. The impact feature extraction submodule extracts key features of the fault waveform and generates a standardized fault feature vector.

[0131] (4) Edge pre-classification stage: The lightweight fault pre-classification sub-module loads the lightweight pre-trained model, performs local forward inference on the standardized fault feature vector, outputs the fault pre-classification result through the weighted voting mechanism, and uploads the fault pre-classification result and fault waveform data to the cloud positioning engine layer 140 after encapsulation.

[0132] (5) Cloud-based precise positioning stage: The cloud positioning engine layer 140 real-time topology construction module corrects the topology relationship based on the relative topology distance identifier. The multi-terminal traveling wave positioning module integrates the real-time topology path distance and the arrival time difference of the fault traveling wave, solves the traveling wave ranging equation, and outputs the precise geographic coordinates of the fault point.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0134] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A distribution network fault location system based on the fusion of AI-driven and power line sensing, characterized in that, include: The offline cloud model optimization layer is used to optimize and prune the random forest model based on historical fault data of the distribution network, through power line carrier communication constraints and power fault physical constraints, to generate a lightweight pre-trained model. The edge sensing layer is deployed at key nodes of the power distribution network to collect electrical signals from power lines in real time and generate relative topological distance identifiers between nodes. The edge computing layer is deployed within the terminal device where the edge perception layer is located. It is used to load the lightweight pre-trained model and perform fault detection, feature extraction and local forward inference based on the collected electrical signals, and output fault pre-classification results. The cloud-based positioning engine layer is used to receive the relative topological distance identifier to dynamically correct the real-time topological relationship, and to calculate the precise geographic coordinates of the fault point based on the fault pre-classification result, the fault traveling wave data collected by the edge perception layer and uploaded by the edge computing layer, and the corrected real-time topological relationship. The offline cloud model optimization layer, edge perception layer, edge computing layer and cloud positioning engine layer interact with each other via a power line carrier communication link.

2. The system according to claim 1, characterized in that, The offline cloud model optimization layer includes: The power fault sample library module is used to store historical fault data labeled with fault categories and physical characteristics, including fault transient attenuation coefficient, high-frequency impedance change value and fault duration. The power edge-adaptive physical constraint pruning random forest optimization module is used to optimize the voting weights of the decision tree under power line carrier communication constraints using the particle swarm optimization algorithm, and prune the decision tree according to the accuracy criterion and physical compliance criterion to generate a lightweight pre-trained model. The model distribution module is used to distribute the lightweight pre-trained model to the edge computing layer via a power line carrier communication link.

3. The system according to claim 2, characterized in that, When the power edge-adaptive physical constraint pruning random forest optimization module optimizes the voting weights of the decision tree for power line carrier communication constraints using the particle swarm optimization algorithm, the specific configuration is as follows: The communication environment is simulated using a power line carrier standard channel model for distribution networks, with channel packet loss rate as the optimization criterion. The inertia weight in the particle swarm optimization algorithm is adaptively adjusted according to the channel packet loss rate. When the channel packet loss rate is higher than a preset threshold, the first inertia weight is used, and when the channel packet loss rate is not higher than the preset threshold, the second inertia weight is used. A sparse particle transport strategy is adopted. After a preset number of iterations, only the particles with the highest objective function values ​​are retained to participate in subsequent iterations, so as to reduce the amount of communication during the optimization process. The optimization objective of the particle swarm optimization algorithm is a weighted sum of model classification accuracy, physical compliance, and model sparse entropy.

4. The system according to claim 2, characterized in that, The power edge-adaptive physical constraint pruning random forest optimization module is specifically configured to prune the decision tree using accuracy and physical compliance criteria as follows: The accuracy criterion is that the classification accuracy of a single decision tree is not lower than the average accuracy of the forest. The physical compliance criterion is that the physical compliance degree output by a single decision tree is not lower than a preset physical compliance threshold. The physical compliance is determined based on the ratio of the number of samples that violate physical constraints to the total number of verification samples in the verification sample set. The determination of the violation of physical constraints is based on the mapping relationship between the fault category and the range of physical parameters. Decision trees that simultaneously satisfy both the accuracy criterion and the physical compliance criterion are retained, while decision trees that do not satisfy either criterion are pruned, thereby generating a lightweight pre-trained model.

5. The system according to claim 4, characterized in that, in, The preset physical compliance threshold is 0.95; the formula for calculating the physical compliance level is: in, The number of verification samples that violate physical constraints. The total number of validation samples is represented by Phys, which represents the physical compliance score.

6. The system according to claim 5, characterized in that, The criterion for determining a violation of physical constraints is the mapping relationship between fault categories and physical parameter ranges. When the physical parameter range corresponding to the fault category output by the lightweight pre-trained model does not match the measured physical quantity in the validation sample, it is determined to be a violation of physical constraints. The lightweight pre-trained model uses a weighted voting mechanism to output fault classification results. Specifically, the weighted voting mechanism is as follows: based on the voting weights of each decision tree obtained after optimization by the particle swarm optimization algorithm, the classification outputs of each decision tree are weighted and summed, and the fault category corresponding to the maximum value in the weighted summation result is taken as the fault pre-classification result output by the model.

7. The system according to claim 1, characterized in that, The edge-aware layer includes: The signal coupling and acquisition unit is directly connected to the power line and is used to acquire power frequency electrical quantities and fault traveling wave signals in real time. A cross-band cognitive communication unit is used to adaptively select the operating frequency based on orthogonal frequency division multiplexing technology to construct a power line carrier communication link; The carrier ranging unit is used to obtain the electrical distance between this terminal and adjacent terminals based on the time of arrival principle, and generate the relative topological distance identifier between the nodes.

8. The system according to claim 6, characterized in that, The edge computing layer includes: The mutation event detection submodule is used to analyze the acquired voltage and current signals using a cumulative sum algorithm, capture signal mutation points, and trigger high-speed waveform recording. The impact feature extraction submodule is used to extract the absolute impact height, impact width, fall velocity, and multi-point fall degree of the fault waveform, and generate a standardized fault feature vector. The lightweight fault pre-classification submodule is used to load the lightweight pre-trained model, perform local forward inference on the standardized fault feature vector, and output the fault type using the weighted voting mechanism corresponding to the lightweight pre-trained model.

9. The system according to claim 1, characterized in that, The cloud-based positioning engine layer includes: The real-time topology construction module is used to receive the relative topology distance identifier and dynamically correct the real-time topology relationship of the distribution network in combination with the graph theory tree search algorithm; The multi-terminal traveling wave positioning module is used to output the geographical coordinates of the fault point by combining the time difference of the fault traveling wave arriving at each edge sensing layer terminal with the path distance in the real-time topology relationship and through the multi-terminal traveling wave ranging equation set.

10. The system according to claim 6, characterized in that, The system is configured to perform collaborative work in the following phases: Offline optimization phase: The offline cloud model optimization layer is based on the power fault sample library, uses the power line carrier standard channel model to complete particle swarm weight optimization, and performs physical constraint pruning according to the mapping relationship between fault category and physical parameter range to generate a lightweight pre-trained model, which is then sent to the edge computing layer through the power line carrier communication link; During normal operation: The edge sensing layer collects electrical signals from power lines in real time and generates relative topological distance identifiers between nodes, which are then uploaded to the cloud positioning engine layer for dynamic correction of real-time topological relationships; Fault Triggering Phase: The edge computing layer identifies signal abrupt changes and triggers high-speed fault waveform recording, extracts key features of the fault waveform, and generates a standardized fault feature vector; Edge pre-classification stage: The edge computing layer loads the lightweight pre-trained model, performs local forward inference on the standardized fault feature vector, outputs the fault pre-classification result through a weighted voting mechanism, and encapsulates the fault pre-classification result and fault waveform data and uploads it to the cloud positioning engine layer. In the cloud-based precise positioning stage: the cloud positioning engine layer corrects the real-time topological relationship based on the relative topological distance identifier, and integrates the real-time topological path distance with the arrival time difference of the fault traveling wave to solve the traveling wave ranging equation and output the precise geographical coordinates of the fault point.

Images