Knowledge graph-AI fusion fault diagnosis method for operation and maintenance management of power distribution network

Abstract

The invention discloses a knowledge graph-AI fusion fault diagnosis method for operation and maintenance management of a power distribution network, and relates to the technical field of artificial intelligence. The problems of low initial diagnosis accuracy, insufficient diagnosis adaptation capability in different grounding forms, long fault positioning time and the like caused by external interference of zero-sequence current signals are solved, and the conditions that the operation and maintenance efficiency is reduced and the power supply loss is increased are finally avoided. The method comprises the following steps: acquiring multi-source data and constructing a dynamically updated power distribution network knowledge graph; multi-mode fault feature extraction is carried out; using an adaptive AI model to identify a fault type and performing primary judgment; and accurately positioning a fault section and analyzing a spreading path in combination with a knowledge graph. According to the method, the single-phase earth fault preliminary diagnosis accuracy is improved to reduce invalid troubleshooting of operation and maintenance, cross-scene diagnosis errors are reduced to adapt to diversified operation and maintenance scenes, average positioning time consumption is shortened, and the supporting capacity of fault diagnosis on operation and maintenance management of the power distribution network is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a knowledge graph-AI fusion fault diagnosis method for power distribution network operation and maintenance management. Background Technology

[0002] With the rapid development and intelligent upgrading of power systems, the distribution network, as a key link connecting users and the main grid, is crucial for urban power supply security and economic and social development. However, the distribution network structure is becoming increasingly complex, the operating environment is variable, and various faults, especially grounding faults, occur frequently, posing a serious threat to the reliability of the power grid. Therefore, efficient and accurate fault diagnosis technology for distribution systems has always been a core issue in the field of power system research and application.

[0003] As the scale of the power distribution network continues to expand, the topology changes frequently and dynamically due to load adjustments, temporary operations, etc., making the operation and maintenance scenarios increasingly complex. Fault diagnosis, as a core part of operation and maintenance management, faces many business-level challenges: In the diagnosis of single-phase grounding faults, the zero-sequence current signal is easily affected by the distributed capacitance of the line. Traditional diagnostic methods that rely on signal feature analysis are difficult to capture effective fault features, resulting in low initial diagnosis accuracy. Operation and maintenance personnel need to conduct repeated on-site inspections, which not only increases labor costs but also delays the opportunity to handle faults. In actual operation and maintenance of power distribution networks, there are various grounding methods such as grounding via arc suppression coil, ungrounded, and direct grounding. The fault characteristics differ significantly under different grounding methods. A single artificial intelligence model has insufficient cross-scenario adaptability, large diagnostic errors, and is prone to deviations in operation and maintenance decisions. For example, misjudging the fault type may lead to improper maintenance operations, increasing equipment wear and maintenance risks. Furthermore, modern distribution networks have numerous branch lines and frequent dynamic adjustments to their topology. Traditional fault location methods are slow to respond to topology changes and fail to identify fault propagation paths in a timely manner, resulting in long average location times. This not only prolongs power outage time and reduces power supply satisfaction for users but may also lead to wider equipment damage due to fault expansion, increasing economic losses in operation and maintenance. Existing technologies, such as those with announcement numbers CN104880646A and CN118244058A, disclose detectors or detection systems for cable faults, involving current / voltage signal measurement and analysis. However, they mostly focus on purely technical fault diagnosis and are not deeply integrated with the actual business needs of distribution network operation and maintenance management. This makes it difficult to support the precise and efficient implementation of operation and maintenance work and fails to fundamentally solve the problems of low operation and maintenance efficiency, high costs, and insufficient power supply reliability. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge graph-AI fusion fault diagnosis method for power distribution network operation and maintenance management, which solves the problems existing in the background technology.

[0005] To address the aforementioned technical problems, the present invention adopts the following technical solution: The present invention provides a knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management, comprising: collecting real-time operation data of the distribution network, distribution network topology data, and distribution network equipment attribute data, constructing and dynamically updating a distribution network knowledge graph; the distribution network knowledge graph includes entities, relationships, and attribute information, wherein the entities represent specific physical or logical components in the distribution network, the relationships represent the connection or interaction methods between entities, and the attributes represent the characteristic parameters of the entities; Multimodal fault feature extraction is performed on the real-time operation data of the power distribution network to generate a fault feature vector; the multimodal fault feature extraction includes noise suppression and time-frequency feature extraction of zero-sequence current sampling waveform data, harmonic analysis and transient feature extraction of three-phase voltage and three-phase current sampling waveform data, and the extracted features are fused. Based on the fault feature vector and the context information provided by the distribution network knowledge graph, an adaptive AI diagnostic model is used to identify the fault type and make a preliminary judgment on the faulty feeder; the adaptive AI diagnostic model selects the corresponding diagnostic sub-model for adaptive adjustment according to the current grounding method information in the distribution network knowledge graph. Based on the fault type identification results, the initial judgment results of the fault feeder, and the distribution network knowledge graph, the fault section is accurately located and the fault propagation path is analyzed, and a diagnostic report is generated.

[0006] The beneficial effects of this invention are as follows: By introducing advanced wavelet packet decomposition, empirical mode decomposition, and zero-sequence current signal processing technology that combines temporal convolutional networks with attention mechanisms, this invention effectively overcomes the interference of line distributed capacitance on zero-sequence current signals, achieves refined extraction and enhancement of fault features, significantly improves the accuracy of initial diagnosis of single-phase grounding faults, reduces ineffective on-site troubleshooting work by maintenance personnel, reduces manpower input and time costs, and provides reliable data support for maintenance decisions.

[0007] This invention constructs a dynamically updated distribution network knowledge graph, accurately models the distribution network topology, equipment attributes, and operating modes, and especially refines the descriptions of different grounding methods. Combined with an adaptive AI diagnostic model, it automatically selects the appropriate diagnostic sub-model based on the real-time grounding method information provided by the knowledge graph, which greatly improves the robustness and accuracy of cross-scenario diagnosis, avoids diagnostic errors caused by poor model adaptation, reduces equipment damage and maintenance risks caused by operational and maintenance decision deviations, and adapts to diverse operational and maintenance scenario requirements.

[0008] This invention combines knowledge graphs with graph neural network inference algorithms, utilizing a real-time updated distribution network knowledge graph for precise fault location and rapid analysis of fault propagation paths. Through graph traversal and graph inference algorithms, it achieves rapid response to dynamic adjustments in the distribution network topology, shortens the average location time, and significantly improves fault handling efficiency and power supply reliability. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Reference Figure 1 As shown, this invention provides a knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management, including: collecting real-time operation data of the distribution network, distribution network topology data, and distribution network equipment attribute data, constructing and dynamically updating a distribution network knowledge graph; the distribution network knowledge graph includes entities, relationships, and attribute information, wherein the entities represent specific physical or logical components in the distribution network, the relationships represent the connection or interaction between entities, and the attributes represent the characteristic parameters of the entities; In a specific embodiment of the present invention, the real-time operation data of the distribution network includes real-time sampled waveform data of three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence current at several monitoring points of the line, and also includes phasor data of three-phase voltage and three-phase current at key nodes. It should be noted that the collection of real-time operation data of the distribution network specifically includes: obtaining real-time sampled waveform data of three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence current at each monitoring point of the distribution network through a distributed intelligent fault indicator; The distributed intelligent fault indicator transmits the collected data to the data aggregation unit via a wireless communication module. The three-phase voltage and three-phase current phasor data of key nodes are obtained by a miniature synchronous phasor measurement unit. The miniature synchronous phasor measurement unit uses the Global Positioning System or the Beidou Satellite Navigation System for time synchronization. The key nodes specifically refer to important locations that play a core role in the distribution network, such as switches in substations, connection points between main lines and branch lines, and important sectional switches. Changes in the electrical quantities of these nodes are crucial for fault diagnosis.

[0013] The power distribution network topology data includes physical components, logical components, line connection relationships, switch status, and feeder information; The line connection relationship refers to the connection logic of each physical component in the distribution network, reflecting the transmission path of power flow. The switch status refers to the operating status of various switching devices in the distribution network, which directly affects the dynamic changes of the topology. The various switching devices include circuit breakers, disconnect switches, sectionalizing switches, etc. The feeder information refers to the basic parameters and identification information of feeders in the distribution network, which are used to distinguish different feeders and clarify their functions. The attribute data of the power distribution network equipment includes transformer parameters, cable type, line length, and grounding method; The transformer parameters refer to the technical parameters of the transformer in the distribution network, reflecting its operating characteristics. The cable type refers to the model and material characteristics of the distribution cable, which affects its electrical performance. The line length refers to the physical length of each line segment, used for subsequent fault location calculations. The grounding method refers to the grounding form of the neutral point, line, or equipment in the distribution network, which directly affects the fault characteristics.

[0014] In a specific embodiment of the present invention, the construction and dynamic updating of the distribution network knowledge graph specifically includes: defining a distribution network knowledge ontology model, which is constructed based on physical and logical components in the distribution network topology data, including substation entities, feeder entities, line entities, switch entities, load entities, and fault entities, as well as the connection relationships, inclusion relationships, membership relationships, and event relationships between the entities; The inclusion relationship is generated based on the feeder information and line connection relationship in the distribution network topology data; the membership relationship is generated based on the connection relationship between load and line in the distribution network topology data, such as the load belonging to the line it is connected to; and the event relationship is generated based on the association between the fault characteristic signal in the real-time operation data of the distribution network and the line location information in the topology data. For example, the connection relationship is such that 110kV substation A is connected to sectionalizing switch K1 through main line L1; the inclusion relationship is such that feeder F1 includes line segments 1-2 and 2-3; the affiliation relationship is such that the transformer load of residential area B belongs to line segment 3; and the time relationship is such that the zero-sequence current mutation is associated with line segment 4-5 of feeder F2.

[0015] The knowledge extraction module extracts entity, relationship and attribute information from distribution network topology data, distribution network equipment attribute information, geographic information system data, distribution automation system data, relay protection setting data and manually entered operation and maintenance records, and stores them in the graph database. Specifically, it extracts the connection relationship between entities and the switch status attribute from the distribution network topology data, and extracts the technical parameter attribute of entities from the distribution network equipment attribute data. The knowledge update module receives power distribution network dispatch instructions, switch operation signals, and temporary line connection information, and updates the switch status and line connection relationships in the graph database in real time to ensure that the power distribution network knowledge graph is consistent with the actual operating status of the power distribution network.

[0016] Multimodal fault feature extraction is performed on the real-time operation data of the power distribution network to generate a fault feature vector; the multimodal fault feature extraction includes noise suppression and time-frequency feature extraction of zero-sequence current sampling waveform data, harmonic analysis and transient feature extraction of three-phase voltage and three-phase current sampling waveform data, and the extracted features are fused. It should be added that the harmonic analysis and transient feature extraction of the three-phase voltage already includes feature integration of the zero-sequence voltage sampling waveform data, such as the amplitude change of the zero-sequence voltage and the change of the zero-sequence component when the three phases are unbalanced. Its features, together with the zero-sequence current and the three-phase current features, participate in the construction of the initial fault feature vector.

[0017] In a specific embodiment of the present invention, the noise suppression and time-frequency feature extraction of the zero-sequence current sampling waveform data specifically includes: performing wavelet packet decomposition on the zero-sequence current sampling waveform data; selecting wavelet basis functions and decomposition levels from a pre-set zero-sequence current-wavelet parameter mapping table in the data warehouse based on the frequency range and noise characteristics of the zero-sequence current in the distribution network; and performing adaptive threshold processing or empirical mode decomposition on the high-frequency coefficients after wavelet packet decomposition to filter out high-frequency random noise and power frequency interference. It should be noted that wavelet packet decomposition is a technique for finely dividing a signal into different frequency bands. It can decompose a zero-sequence current signal into multiple low-frequency effective components and high-frequency noise components. After decomposition, the high-frequency coefficients mainly correspond to noise signals. Through adaptive thresholding or empirical mode decomposition, high-frequency random noise and power frequency interference can be effectively filtered out, while retaining fault characteristic signals in the low-frequency band. The adaptive thresholding specifically involves automatically adjusting the threshold according to the noise energy and eliminating high-frequency coefficients below the threshold. The empirical mode decomposition specifically involves decomposing high-frequency noise into intrinsic mode functions and removing them. The specific application methods are already quite mature in the existing technology and will not be elaborated here.

[0018] It should also be noted that the zero-sequence current-wavelet parameter mapping table specifically refers to the mapping relationship between the frequency range and noise characteristics of the zero-sequence current and the wavelet basis function and the number of decomposition layers. Specifically, it is based on the analysis of the frequency range and noise characteristics of a large number of zero-sequence current samples, and the corresponding relationship is established through pre-training such as statistical analysis and machine learning fitting.

[0019] The denoised zero-sequence current signal is subjected to short-time Fourier transform or continuous wavelet transform to generate a time-spectrum graph of the zero-sequence current, and time-domain, frequency-domain, and time-frequency-domain features are extracted. The time-domain features include peak value, root mean square value, and waveform factor. The frequency-domain features include dominant frequency component and harmonic content. The time-frequency-domain features include energy distribution.

[0020] This invention introduces advanced wavelet packet decomposition, empirical mode decomposition, and time convolutional network combined with attention mechanism zero-sequence current signal processing technology. It effectively overcomes the interference of line distributed capacitance on zero-sequence current signal, realizes refined extraction and enhancement of fault features, significantly improves the accuracy of initial diagnosis of single-phase grounding faults, reduces the ineffective on-site investigation work of operation and maintenance personnel, reduces manpower input and time costs, and provides reliable data support for operation and maintenance decision-making.

[0021] In a specific embodiment of the present invention, the harmonic analysis and transient feature extraction of the three-phase voltage and three-phase current sampled waveform data specifically includes: performing a fast Fourier transform on the three-phase voltage and three-phase current sampled waveform data to extract the amplitude and phase of each harmonic component; It should be noted that the three-phase voltage and three-phase current are dominated by the 50Hz fundamental frequency during normal operation. However, a large number of harmonics will be generated under fault conditions. The amplitude and phase changes of these harmonics are directly related to the fault type. Therefore, harmonic analysis of the three-phase voltage and three-phase current is required.

[0022] Identify transient overvoltage and transient overcurrent events in the three-phase voltage and three-phase current sampled waveform data, and extract the start time, duration, amplitude, steepness, and energy of the transient waveform.

[0023] It should be noted that at the moment a fault occurs, the three-phase voltage and three-phase current will experience non-periodic transient changes. These transient characteristics are the key to distinguishing fault types. Therefore, it is necessary to extract transient characteristics from the three-phase voltage and three-phase current.

[0024] In a specific embodiment of the present invention, the step of fusing the extracted features specifically includes: splicing the time-domain, frequency-domain, and time-frequency-domain features of the zero-sequence current with the harmonic features and transient features of the three-phase voltage and current to form an initial fault feature vector; The initial fault feature vector is nonlinearly mapped and dimensionality reduced using a feature fusion network based on deep learning to generate the final fault feature vector.

[0025] It should be noted that the deep learning-based feature fusion network is pre-trained. Specifically, historical fault data of the distribution network is collected, and the time-frequency features of zero-sequence current, harmonics and transient features of three-phase voltage and current are extracted and concatenated into an initial fault feature vector. Corresponding fault type labels, such as single-phase grounding and phase-to-phase short circuit, are also added. After data normalization, the data is divided into training, validation, and test sets. The network adopts an architecture of an input layer, multiple fully connected hidden layers, and an output layer. The loss function employs a joint strategy: reconstruction loss (using the decoder to restore the initial features) ensures the preservation of key information, while classification loss uses cross-entropy to associate fault labels, enhancing feature discriminative power. Iteration is performed using the Adam optimizer, updating weights through backpropagation, combined with early stopping strategies such as stopping when the validation set accuracy decreases and L2 regularization (limiting excessive weights to avoid overfitting). Training is complete when the final feature vector on the validation set achieves a classification accuracy exceeding the accuracy threshold, and the correlation between features after dimensionality reduction is lower than the correlation coefficient threshold. At this point, the trained deep learning-based feature fusion network can transform the new initial vector into a concise, highly discriminative final feature vector.

[0026] Based on the fault feature vector and the context information provided by the distribution network knowledge graph, an adaptive AI diagnostic model is used to identify the fault type and make a preliminary judgment on the faulty feeder; the adaptive AI diagnostic model selects the corresponding diagnostic sub-model for adaptive adjustment according to the current grounding method information in the distribution network knowledge graph. In a specific embodiment of the present invention, the step of using an adaptive AI diagnostic model to identify fault types and make preliminary judgments on fault feeders specifically includes: the adaptive AI diagnostic model includes multiple trained diagnostic sub-models, each diagnostic sub-model corresponding to a specified distribution network grounding method, the grounding method including grounding through an arc suppression coil, no grounding, and direct grounding; It should be noted that the training of the diagnostic sub-model is carried out separately for different grounding methods. Specifically, historical fault data is collected according to the grounding method: for systems grounded via arc suppression coils, data such as the zero-sequence current resonance characteristics and damped oscillation waveforms during single-phase grounding are collected; for ungrounded systems, data such as small zero-sequence current and unbalanced three-phase voltage during single-phase grounding are collected; for directly grounded systems, data such as large current surges and high harmonic content during short-circuit faults are collected. Simultaneously, the corresponding fault types are labeled for each type of data, such as single-phase grounding and phase-to-phase short circuit. Features are extracted separately for each type of data, and training and validation sets are created. The diagnostic sub-model uses the aforementioned LSTM, GRU, or a combination thereof, incorporating an attention mechanism to enhance key features such as the 100Hz resonant frequency of the arc suppression coil system and the transient current steepness of the directly grounded system. During training, for each sub-model of each grounding method, the cross-entropy loss is minimized using the Adam optimizer with the corresponding dataset, iteratively adjusting the network weights so that the model learns the mapping relationship between fault characteristics and types under that grounding method. For example, in an ungrounded system: a rise in zero-sequence voltage + a small zero-sequence current corresponds to single-phase grounding.

[0027] The distribution network knowledge graph provides information on the grounding method of each feeder in the current distribution network and the association information between the feeder and the fault characteristics. The association information reflects the attribution relationship between each dimension of the final fault characteristic vector and the corresponding feeder. The final fault feature vector contains fault features of multiple feeders in the distribution network, and the fault features of each feeder are distinguished through the association information. Based on the grounding method information provided by the power distribution network knowledge graph, feature sub-vectors corresponding to each feeder are extracted from the final fault feature vector. Diagnostic sub-models matching the grounding method of each feeder are selected as input to obtain the fault type identification result and the initial judgment result of the fault feeder for each feeder. The fault type identification result includes single-phase grounding fault, phase-to-phase short circuit fault, or open circuit fault. The initial judgment result of the fault feeder indicates the feeder range where the fault may occur.

[0028] This invention constructs a dynamically updated distribution network knowledge graph, accurately models the distribution network topology, equipment attributes, and operating modes, and especially refines the descriptions of different grounding methods. Combined with an adaptive AI diagnostic model, it automatically selects the appropriate diagnostic sub-model based on the real-time grounding method information provided by the knowledge graph, which greatly improves the robustness and accuracy of cross-scenario diagnosis, avoids diagnostic errors caused by poor model adaptation, reduces equipment damage and maintenance risks caused by operational and maintenance decision deviations, and adapts to diverse operational and maintenance scenario requirements.

[0029] This invention combines knowledge graphs with graph neural network inference algorithms, utilizing a real-time updated distribution network knowledge graph for precise fault location and rapid analysis of fault propagation paths. Through graph traversal and graph inference algorithms, it achieves rapid response to dynamic adjustments in the distribution network topology, shortens the average location time, and significantly improves fault handling efficiency and power supply reliability.

[0030] In a specific embodiment of the present invention, the adaptive AI diagnostic model is constructed based on a long short-term memory network, a gated recurrent unit network, or a combination of both, and utilizes an attention mechanism to enhance the ability to capture key features.

[0031] Specifically, the introduction of an attention mechanism enables the adaptive AI diagnostic model to automatically assign dynamic weights to different features when processing multimodal fault feature vectors such as zero-sequence current time-frequency features and three-phase harmonic features. That is, features that play a key role in fault diagnosis, such as the transient current steepness at the moment of fault and specific harmonic components, will be given higher attention, while secondary features, such as small amplitude changes in normal fluctuations, will be weakened. This enhances the model's ability to capture core fault features and reduces interference from irrelevant information.

[0032] Based on the fault type identification results, the initial judgment results of the fault feeder, and the distribution network knowledge graph, the fault section is accurately located and the fault propagation path is analyzed, and a diagnostic report is generated.

[0033] In a specific embodiment of the present invention, the step of performing precise fault location and fault propagation path analysis specifically includes: based on the initial judgment result of the fault feeder, obtaining the topology information and equipment status information of the relevant feeder segment from the power distribution network knowledge graph; Using a graph neural network-based inference algorithm, combined with alarm information from fault indicators and line current direction information, fault sections are identified on the distribution network knowledge graph to determine the most likely line section where the fault occurred. Based on the topology and historical fault patterns of the power distribution network knowledge graph, depth-first search or breadth-first search algorithms are used to predict the potential power outage range and downstream equipment that may be affected by the fault, thus forming a fault propagation path.

[0034] In a specific embodiment of the present invention, the precise location of the fault section is specifically combined with at least one of the following location principles: based on impedance method location, using the precise impedance parameters of each line segment stored in the distribution network knowledge graph, combined with the voltage and current measurement values ​​before and after the fault point, to calculate the electrical distance from the fault point to the measurement point. The fault location is based on the traveling wave method, which utilizes the propagation characteristics of the traveling wave signal generated when the fault occurs. By calculating the time difference of the traveling wave reaching different measurement points, the location of the fault point can be accurately calculated. The fault location information is located by utilizing the fault isolation and recovery operation information provided by the feeder automation system deployed in the distribution network.

[0035] The network models involved in this invention, such as Long Short-Term Memory (LSTM) networks, gated recurrent unit (GRU) networks, and graph neural networks, as well as techniques commonly used in fault diagnosis, such as wavelet packet decomposition, short-time Fourier transform, cross-entropy loss function, and Adam optimizer, are all common knowledge in the field. The basic architecture and training logic of these models, such as iterative optimization, regularization to prevent overfitting, and feature processing methods, are well-documented in existing literature and publicly available materials related to power distribution automation and fault diagnosis. Those skilled in the art can understand and apply these techniques based on common knowledge. Although some details are not elaborated upon, combining the known technologies with the core processes disclosed in this invention, such as feature fusion logic and the matching mechanism between sub-models and grounding methods, is sufficient to achieve the technical solution of this invention without affecting its implementation and reproduction.

[0036] The thresholds involved in this invention, such as the accuracy threshold and correlation coefficient threshold of the feature fusion network, are specifically determined based on statistical analysis of a large amount of historical fault data and cross-validation. By repeatedly testing the impact of different thresholds on the diagnostic effect through the validation set, values ​​that can balance accuracy and generalization ability are selected to adapt to actual fault scenarios. This is a conventional setting in the field and can be reasonably determined by those skilled in the art.

[0037] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management, characterized in that, include: Real-time operation data, topology data, and equipment attribute data of the distribution network are collected to construct and dynamically update a distribution network knowledge graph. The distribution network knowledge graph includes entities, relationships, and attribute information. Entities represent specific physical or logical components in the distribution network, relationships represent the connection or interaction between entities, and attributes represent the characteristic parameters of entities. Multimodal fault features are extracted from the real-time operation data of the power distribution network to generate fault feature vectors; The multi-mode fault feature extraction includes noise suppression and time-frequency feature extraction of zero-sequence current sampling waveform data, harmonic analysis and transient feature extraction of three-phase voltage and three-phase current sampling waveform data, and fusion of the extracted features; Based on the fault feature vector and the context information provided by the distribution network knowledge graph, an adaptive AI diagnostic model is used to identify the fault type and make a preliminary judgment on the faulty feeder; the adaptive AI diagnostic model selects the corresponding diagnostic sub-model for adaptive adjustment according to the current grounding method information in the distribution network knowledge graph. The use of an adaptive AI diagnostic model for fault type identification and initial fault feeder judgment specifically includes: The adaptive AI diagnostic model includes multiple trained diagnostic sub-models, each corresponding to a specified power distribution network grounding method, which includes grounding via an arc suppression coil, no grounding, and direct grounding. The distribution network knowledge graph provides information on the grounding method of each feeder in the current distribution network and the association information between the feeder and the fault characteristics. The association information reflects the attribution relationship between each dimension of the final fault characteristic vector and the corresponding feeder. The final fault feature vector contains fault features of multiple feeders in the distribution network, and the fault features of each feeder are distinguished through the association information. Based on the grounding method information provided by the power distribution network knowledge graph, feature sub-vectors corresponding to each feeder are extracted from the final fault feature vector. Diagnostic sub-models matching the grounding method of each feeder are selected as input to obtain the fault type identification result and the initial judgment result of the fault feeder for each feeder. The fault type identification result includes single-phase grounding fault, phase-to-phase short circuit fault, or open circuit fault. The initial judgment result of the fault feeder indicates the feeder range where the fault may occur. Based on the fault type identification results, the initial judgment results of the fault feeder, and the distribution network knowledge graph, the fault section is accurately located and the fault propagation path is analyzed, and a diagnostic report is generated.

2. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The real-time operation data of the distribution network includes real-time sampled waveform data of three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence current at several monitoring points of the line, as well as phasor data of three-phase voltage and three-phase current at key nodes. The power distribution network topology data includes physical components, logical components, line connection relationships, switch status, and feeder information; The line connection relationship refers to the connection logic of each physical component in the distribution network, reflecting the transmission path of power flow. The switch status refers to the operating status of various switching devices in the distribution network, which directly affects the dynamic changes of the topology. The feeder information refers to the basic parameters and identification information of the feeders in the distribution network, used to distinguish different feeders and clarify their functions. The attribute data of the power distribution network equipment includes transformer parameters, cable type, line length, and grounding method; The transformer parameters refer to the technical parameters of the transformer in the distribution network, reflecting its operating characteristics. The cable type refers to the model and material characteristics of the distribution cable, which affects its electrical performance. The line length refers to the physical length of each line segment, used for subsequent fault location calculations. The grounding method refers to the grounding form of the neutral point, line, or equipment in the distribution network, which directly affects the fault characteristics.

3. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The construction and dynamic updating of the distribution network knowledge graph specifically includes: Define a knowledge ontology model for a distribution network. The ontology model is constructed based on the physical and logical components in the distribution network topology data, including substation entities, feeder entities, line entities, switch entities, load entities, and fault entities, as well as the connection relationships, inclusion relationships, membership relationships, and event relationships between the entities. The knowledge extraction module extracts entity, relationship and attribute information from distribution network topology data, distribution network equipment attribute information, geographic information system data, distribution automation system data, relay protection setting data and manually entered operation and maintenance records, and stores them in the graph database. The knowledge update module receives power distribution network dispatch instructions, switch operation signals, and temporary line connection information, and updates the switch status and line connection relationships in the graph database in real time to ensure that the power distribution network knowledge graph is consistent with the actual operating status of the power distribution network.

4. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The noise suppression and time-frequency feature extraction of the zero-sequence current sampling waveform data specifically includes: The zero-sequence current sampling waveform data is decomposed into wavelet packets. Based on the frequency range and noise characteristics of the zero-sequence current in the distribution network, the wavelet basis function and the number of decomposition levels are selected from the zero-sequence current-wavelet parameter mapping table preset in the data warehouse. The high-frequency coefficients after wavelet packet decomposition are subjected to adaptive threshold processing or empirical mode decomposition to filter out high-frequency random noise and power frequency interference. The denoised zero-sequence current signal is subjected to short-time Fourier transform or continuous wavelet transform to generate a time-spectrum graph of the zero-sequence current, and time-domain, frequency-domain, and time-frequency-domain features are extracted. The time-domain features include peak value, root mean square value, and waveform factor. The frequency-domain features include dominant frequency component and harmonic content. The time-frequency-domain features include energy distribution.

5. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The harmonic analysis and transient feature extraction of the sampled waveform data of the three-phase voltage and three-phase current specifically includes: The three-phase voltage and three-phase current sampled waveform data are subjected to Fast Fourier Transform to extract the amplitude and phase of each harmonic component; Identify transient overvoltage and transient overcurrent events in the three-phase voltage and three-phase current sampled waveform data, and extract the start time, duration, amplitude, steepness, and energy of the transient waveform.

6. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The process of fusing the extracted features specifically includes: The time-domain, frequency-domain, and time-frequency-domain characteristics of the zero-sequence current are concatenated with the harmonic and transient characteristics of the three-phase voltage and current to form an initial fault feature vector. The initial fault feature vector is nonlinearly mapped and dimensionality reduced using a feature fusion network based on deep learning to generate the final fault feature vector.

7. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The adaptive AI diagnostic model is constructed based on long short-term memory networks, gated recurrent unit networks, or a combination of both, and utilizes attention mechanisms to enhance the ability to capture key features.

8. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 1, characterized in that, The precise location of the faulty section and analysis of the fault propagation path specifically include: Based on the initial judgment of the faulty feeder, the topology information and equipment status information of the relevant feeder segment are obtained from the distribution network knowledge graph; Using a graph neural network-based inference algorithm, combined with alarm information from fault indicators and line current direction information, fault sections are identified on the distribution network knowledge graph to determine the most likely line section where the fault occurred. Based on the topology and historical fault patterns of the power distribution network knowledge graph, depth-first search or breadth-first search algorithms are used to predict the potential power outage range and downstream equipment that may be affected by the fault, thus forming a fault propagation path.

9. The knowledge graph-AI fusion fault diagnosis method for distribution network operation and maintenance management according to claim 8, characterized in that, The precise location of the fault section specifically combines at least one of the following location principles: Based on impedance method location, the electrical distance from the fault point to the measurement point is calculated by using the precise impedance parameters of each line segment stored in the distribution network knowledge graph and combining the voltage and current measurements before and after the fault point. The fault location is based on the traveling wave method, which utilizes the propagation characteristics of the traveling wave signal generated when the fault occurs. By calculating the time difference of the traveling wave reaching different measurement points, the location of the fault point can be accurately calculated. The fault location information is located by utilizing the fault isolation and recovery operation information provided by the feeder automation system deployed in the distribution network.

Images