Method and device for identifying power distribution network fault based on artificial intelligence

Abstract

The invention discloses a power distribution network fault identification method and device based on artificial intelligence, and the method comprises the steps: obtaining the multi-source fault characterization information of a power distribution network, and the multi-source fault characterization information comprises telemetering data and telecommand data; performing expert database-based normalization processing on the multi-source fault representation information to generate a standardized fault data sample; matching and checking the standardized fault data sample to generate checked fault data; performing graph data conversion processing on the checked fault data according to a power distribution network graph model to generate fault graph data; and based on the fault graph data, reasoning the fault of the power distribution network through the fault reasoning model to obtain the speculated fault of the power distribution network. The multi-source information is obtained, the limitation of a single channel is broken, data differences are eliminated through expert database normalization, data accuracy is guaranteed through matching checking, graph data conversion is associated with topology, artificial intelligence reasoning gets rid of manual dependence, and efficiency and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a method and device for identifying power distribution network faults based on artificial intelligence. BACKGROUND

[0002] As a key link of the power system directly facing users, the safe and stable operation of the power distribution network is crucial to ensuring power supply reliability. In actual operation, the power distribution network is prone to faults due to various factors such as adverse weather (lightning, strong winds, etc.), equipment aging, external damage, etc. Once a fault occurs, it needs to be quickly and accurately identified in order to quickly restore power supply and reduce economic losses and social impacts caused by power outages.

[0003] Currently, power distribution network fault identification is mainly based on artificial experience, which mainly relies on the professional experience of dispatchers or maintenance personnel to collect fault information through telephone repair, on-site inspection, etc., and then further judges the fault location and type. For example, after receiving a user's power outage repair report, the staff will preliminarily speculate the possible fault section according to the repair range and their own familiarity with the line direction and equipment distribution, and then confirm the specific fault point through on-site investigation.

[0004] However, this identification method relying on artificial experience has subjectivity and limitations, and different personnel may have different judgments on the same fault information, which makes it difficult to ensure the consistency and accuracy of fault identification. Especially in the case of increasingly complex power distribution network structure and diverse fault types, relying solely on artificial experience often cannot quickly and comprehensively integrate various fault information, which may lead to fault misjudgment or omission, prolong fault handling time, and affect power supply recovery efficiency. SUMMARY

[0005] The embodiments of the present application provide a method and device for identifying power distribution network faults based on artificial intelligence to solve or alleviate the technical defects in the prior art.

[0006] In one aspect, the embodiments of the present application provide a method for identifying power distribution network faults based on artificial intelligence, which includes:

[0007] Step 1, acquiring multi-source fault characterization information of the power distribution network, the multi-source fault characterization information including telemetry data and telesign data;

[0008] Step 2, performing normalization processing on the multi-source fault characterization information based on an expert library to generate standardized fault data samples;

[0009] Step 3, performing matching and checking processing on the standardized fault data samples to generate checked fault data;

[0010] Step 4, according to the power distribution network diagram model, the fault data after checking is processed by graph data conversion to generate fault graph data;

[0011] Step 5, based on the fault graph data, the fault of the power distribution network is inferred by the fault inference model to obtain the inferred power distribution network fault.

[0012] On the other hand, the application also provides a device for identifying power distribution network fault based on artificial intelligence, comprising:

[0013] The acquisition unit is configured to acquire multi-source fault characterization information of the power distribution network, wherein the multi-source fault characterization information comprises telemetering data and remote signaling data.

[0014] The standardization unit is configured to perform normalization processing on the multi-source fault characterization information based on an expert library to generate standardized fault data samples.

[0015] The checking unit is configured to perform matching checking processing on the standardized fault data samples to generate fault data after checking.

[0016] The feature extraction unit is configured to perform graph data conversion processing on the fault data after checking according to a power distribution network diagram model to generate fault graph data.

[0017] The fault inference unit is configured to infer the fault of the power distribution network based on the fault graph data by a fault inference model to obtain an inferred power distribution network fault.

[0018] The application has the following advantages: by acquiring multi-source fault characterization information (telemetry data and remote signaling data), the limitations of relying on a single information channel such as telephone repair and on-site inspection for manual identification are broken, and fault-related data of the power distribution network can be comprehensively collected to provide a richer information basis for subsequent fault identification and reduce misjudgment and missed judgment caused by incomplete information. By normalizing multi-source information based on an expert library to generate standardized samples, the format difference and expression inconsistency of different source data are eliminated, the judgment interference caused by various data forms during manual processing is avoided, the fault information has a unified standard, the data usability is improved, and a consistent data basis is laid for subsequent processing. By matching and checking the standardized samples to generate checked data, the data is verified and corrected by technical means, and the error or abnormal data is eliminated, the judgment deviation problem caused by information error in manual experience judgment is solved, and the accuracy of the fault data used for analysis is ensured. By converting the checked data into fault graph data, the association of the fault data and the topological structure is realized in combination with the power distribution network graph model, the data can reflect the physical connection relationship of the power distribution network, the deficiency that it is difficult to comprehensively integrate various types of information in a complex network during manual identification is overcome, the fault analysis is more in line with the actual structure of the power distribution network. The fault reasoning is performed through the fault reasoning model, the complex data and topological association are processed by using an artificial intelligence algorithm, the dependence on manual experience is eliminated, the subjectivity and limitation are avoided, the fault can be quickly identified from a large amount of data, the complex power distribution network and various fault types are adapted, the efficiency and accuracy of fault identification are improved, and power supply is quickly restored. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0020] Figure 1 A flowchart of a power distribution network fault identification method based on artificial intelligence according to an embodiment of the present application.

[0021] Figure 2 A structural diagram of a power distribution network fault identification device based on artificial intelligence according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0023] As shown in the figure, the method for identifying power distribution network fault based on artificial intelligence according to an embodiment of the present application comprises the following steps: Figure 1

[0024] Step 1: Obtain multi-source fault characterization information of the power distribution network, wherein the multi-source fault characterization information comprises telemetering data and telesignaling data.

[0025] Step 2: Perform normalization processing on the multi-source fault characterization information based on an expert library to generate standardized fault data samples.

[0026] Step 3: Perform matching and checking processing on the standardized fault data samples to generate checked fault data.

[0027] Step 4: Perform graph data conversion processing on the checked fault data according to a power distribution network graph model to generate fault graph data.

[0028] Step 5: Based on the fault graph data, perform reasoning on the fault of the power distribution network through a fault reasoning model to obtain a speculated fault of the power distribution network.

[0029] Optionally, in step 1, the multi-source fault characterization information of the power distribution network comprises telemetering data and telesignaling data, and specifically comprises the following steps:

[0030] Step 11: Collect telemetering data and telesignaling data within a set time of fault occurrence from a dispatching automation system, a power distribution network automation system, a power utilization information collection system and a power distribution line online monitoring system.

[0031] Step 12: Perform feature extraction on the collected telemetering data and telesignaling data, extract the line name, substation name, circuit breaker name corresponding number, switch state 0-1 distribution and fault section number, obtain multi-source coupled fault characterization information labels and fault characterization metadata, and generate multi-source fault characterization information.

[0032] ​Preferably, in a specific application scenario, step 11 is implemented by first determining the time range in which data needs to be collected, i.e. a set time period before and after the fault occurs, such as 10 minutes before the fault occurs to 30 minutes after the fault occurs, which can cover the key data changes before and after the fault occurs. Then, data communication connections are established with the dispatch automation system, the distribution network automation system, the electricity consumption information collection system, and the distribution line online monitoring system, respectively, and the telemetry data and the remote signaling data in the time period are obtained through the standard data interfaces (such as IEC 61850 protocol interfaces) provided by each system. The telemetry data includes continuous change analog data such as current, voltage, and power, and the remote signaling data includes discrete state data such as switch opening and closing states and protection action signals, ensuring that the collected data comprehensively and accurately reflects the state of the distribution network at the time of the fault.

[0033] Preferably, in a specific application scenario, step 12 is implemented by using a method combining data cleaning and feature extraction. First, the collected telemetry data and remote signaling data are preprocessed to remove noise data and invalid data, such as removing abnormal values that are obviously outside the reasonable range and filling in missing short-time data. Then, a rule-based feature extraction algorithm is used to extract key information from the processed data according to the naming conventions and coding rules of the distribution network equipment: identify and extract the unique numbers corresponding to the line names, substation names, and circuit breaker names, which are important identifiers for distinguishing different equipment and areas; convert the opening and closing states of the switches to digital signals with 0-1 distribution, where 1 represents the closing state of the switch and 0 represents the opening state of the switch; determine the number corresponding to the fault section, which is determined according to the position of the fault signal and the topology of the distribution network. Through these operations, the extracted information is integrated into multi-source coupled fault representation information labels and fault representation metadata containing data sources, collection times, etc., to form multi-source fault representation information.

[0034] To this end, by collecting the telemetry data (such as continuous quantities of current, voltage, etc.) and remote signaling data (such as discrete quantities of switch state, protection action, etc.) within the set time of fault occurrence from multiple independent systems such as dispatch automation system, distribution network automation system, etc., the limitation of limited data coverage of a single system can be broken through, and the information obtained can ensure the complete reflection of the state change of the distribution network before and after the fault occurs. This multi-source data collection method provides rich and relevant raw materials for subsequent fault analysis, avoiding fault misjudgment caused by one-sided data. In addition, through the feature extraction algorithm, the unique number of the line, the transformer area and the circuit breaker is accurately extracted from the original data, the switch state is standardized to 0-1 distribution, and the fault section number is determined. This process not only eliminates the integration obstacles caused by the format differences of multi-source data, but also converts massive raw data into structured fault characterization information tags and metadata by focusing on core features (such as device identification, state signal, fault location identification). This processing changes the data from a scattered and disordered state to a standardized information that can be directly used for subsequent analysis, greatly improving the usability of the data and laying a high-quality data foundation for subsequent normalized processing and fault reasoning based on the expert library.

[0035] Optionally, step 2, the multi-source fault characterization information is subjected to normalized processing based on the expert library to generate standardized fault data samples, specifically including the following steps:

[0036] Step 21, multi-dimensional feature analysis and coupling processing are performed on the multi-source fault characterization information to obtain multi-source heterogeneous coupled fault data samples;

[0037] Step 22, based on the expert library, topological correlation reasoning and dynamic normalization processing are performed on the multi-source heterogeneous coupled fault data samples to generate standardized fault data samples.

[0038] Optionally, step 21, multi-dimensional feature analysis and coupling processing are performed on the multi-source fault characterization information to obtain multi-source heterogeneous coupled fault data samples, specifically including the following steps:

[0039] Step 211, time series feature analysis is performed on the telemetry data in the multi-source fault characterization information to extract dynamic features, and state feature analysis is performed on the remote signaling data to extract discrete features;

[0040] Step 212, based on the dynamic features and discrete features, a multi-dimensional feature set is generated;

[0041] Step 213, based on the topological relationship of the distribution network, a spatio-temporal correlation rule between the telemetry data and the remote signaling data in the multi-source fault characterization information is established to generate a feature correlation matrix;

[0042] Step 214, the multi-dimensional feature set and the feature correlation matrix are fused and encoded to generate multi-source heterogeneous coupled fault data samples.

[0043] Preferably, in a specific application scenario, step 211 is specifically implemented for telemetric data (such as continuous quantities of current, voltage, etc. changing over time) using a sliding window combined with time series analysis to analyze time series features. Specifically, using a long short-term memory network (LSTM) algorithm, by setting a reasonable time window (such as 5 minutes as a window), the trend changes (such as voltage drop amplitude, current mutation slope) and periodic features (such as waveform change rules before and after the fault) of the telemetric data in each window are captured, thereby extracting dynamic features that can reflect the dynamic changes of the data. For telesignaling data (such as switch opening and closing states, protection action signals, etc. discrete quantities), a feature analysis method based on state transition is used, and a hidden Markov model (HMM) is used to analyze the transition probability and duration between different states, and to extract discrete features such as switch state 0-1 jump time and protection action sequence, which can accurately reflect the state changes of the equipment.

[0044] To this end, in step 211, the depth value mining of different types of data in multi-source fault information is realized through targeted feature analysis methods. For telemetric data, the LSTM algorithm is used for time series feature analysis, which can effectively capture the dynamic trend of continuous quantities such as current and voltage changing over time (such as mutation rules before and after the fault), because the gating mechanism of LSTM can remember long-term dependencies, avoiding feature distortion caused by loss of long-range data association in traditional time series analysis; for telesignaling data, the hidden Markov model is used to extract discrete features, which can accurately depict the state transition rules of discrete quantities such as switch state jump and protection action sequence, and it solves the feature extraction problem of discrete signals caused by randomness through the probability modeling of hidden states and observation states. This classification analysis method not only retains the dynamic change characteristics of telemetric data, but also highlights the state correlation features of telesignaling data, laying a high-quality foundation for subsequent multi-dimensional feature integration and avoiding feature interference when different types of data are mixed together.

[0045] Preferably, in a specific application scenario, step 212 is specifically implemented, based on the extracted dynamic features and discrete features, using feature splicing and dimension expansion to generate a multi-dimensional feature set. First, arrange the dynamic features in chronological order to form a continuous feature vector; arrange the discrete features by device type to form a discrete feature vector. Then, use a feature fusion algorithm to splice the two types of feature vectors according to the time dimension and device association dimension, expand the dimension of the features, so that the generated multi-dimensional feature set contains not only the dynamic change information of telemetric data, but also the state features of telesignaling data, realizing the preliminary integration of different types of features.

[0046] To this end, in step 212, by constructing a multi-dimensional feature set, the systematic integration and value release of different types of fault features are realized. The dynamic characteristics of telemetry data (such as the time sequence change trend of current and voltage) and the discrete characteristics of remote data (such as the 0-1 jump of switch state, the sequence of protection action) respectively reflect different aspects of power distribution network fault - dynamic characteristics embody the process change of fault, discrete characteristics embody the state response of equipment. Integrating these two types of features into a multi-dimensional feature set breaks the limitations of single feature dimension, and changes the feature information from scattered and isolated state to mutually complementary and correlated whole. This integration not only preserves the original properties of each type of feature, but also enables the feature set to comprehensively cover the dynamic evolution and static state of the fault through dimension expansion, providing a complete feature basis for subsequent establishment of spatio-temporal association rules of telemetry and remote data, avoiding the association analysis bias caused by feature fragmentation, and improving the accuracy of subsequent data coupling processing.

[0047] Preferably, in a specific application scenario, step 213 is specifically implemented to establish the association between telemetry data and remote data based on the topology relationship of the power distribution network by constructing a spatio-temporal association rule mining algorithm. First, according to the topology information of the power distribution network such as line connection and equipment layout, the physical connection relationship between different devices (such as circuit breaker and line, transformer and transformer area) is determined. Then, using an association rule mining algorithm (such as Apriori algorithm), the appearance frequency and sequence of telemetry data (such as current surge of a line) and remote data (such as tripping action of the circuit breaker of the line) in the same time period are analyzed, the spatio-temporal association rules (such as the probability of circuit breaker tripping within 10 seconds after current surge) are mined, and these rules are represented in matrix form to generate a feature association matrix, and the element value in the matrix represents the association strength of the corresponding telemetry feature and remote feature.

[0048] To this end, in step 213, by establishing the spatio-temporal association rules of telemetry data and remote data, the natural barriers between the two types of data are broken, and the internal logical association of multi-source information is realized. The occurrence of power distribution network fault is often accompanied by the coordinated change of telemetry data (such as current surge) and remote data (such as circuit breaker tripping), and this change is constrained by the topology relationship (such as the causal relationship between the state change of devices on the same line). Using an association rule mining algorithm (such as Apriori algorithm) combined with the topology structure of the power distribution network, the corresponding relationship between telemetry features and remote state of different devices in the same time dimension (such as the tripping of the upstream circuit breaker within 10 seconds after the current surge of a line) and the association rules of data in the same space position (such as the same transformer area) can be accurately captured. The generated feature association matrix quantifies the association strength, making the originally independent telemetry and remote data form a relationship chain that mutually confirms each other, providing a logical basis for subsequent data fusion, avoiding fault feature misjudgment caused by data isolation, and significantly improving the collaborative analysis capability of multi-source data.

[0049] Preferably, in a specific application scenario, when step 214 is specifically implemented, a method combining feature embedding and matrix fusion is adopted to fuse and encode the multi-dimensional feature set and the feature association matrix. First, the word embedding algorithm is used to convert various features in the multi-dimensional feature set into low-dimensional dense vectors, so that the features are easier to calculate and associate. Then, the feature association matrix is ​​used as a weight matrix to perform weighted adjustment on the converted feature vectors to strengthen the connection between features with high correlation and weaken the influence of irrelevant features. Finally, through matrix splicing and flattening processing, the adjusted feature vector and the feature association matrix are fused into a unified high-dimensional feature vector, which contains multi-dimensional feature information and reflects the correlation relationship between features, thereby generating multi-source heterogeneous coupled fault data samples.

[0050] To this end, step 214 achieves the organic integration of multi-source heterogeneous data through fusion coding, allowing the scattered feature information to form a holistic and correlated fault description. The multi-dimensional feature set covers the dynamic features of telemetry data and the discrete features of telesignaling data, but lacks the quantitative correlation between features. While the feature correlation matrix characterizes this correlation, it does not contain specific eigenvalues. A feature embedding algorithm is used to convert the multi-dimensional feature set into a low-dimensional dense vector, solving the problem of high-dimensional feature redundancy. At the same time, through weight adjustment of the feature correlation matrix, the influence of key correlation features (such as the strong correlation between current mutation and circuit breaker operation during faults) is strengthened. The resulting multi-source heterogeneous coupled fault data sample not only retains the original information of each dimensional feature, but also reflects the spatiotemporal correlation patterns between features through fusion coding of the correlation matrix, transforming the data from "fragmented" to "structured." This provides complete and closely correlated input for subsequent normalization processing and fault reasoning based on the expert library, avoiding analytical bias caused by data fragmentation and improving the accuracy of fault identification.

[0051] Optionally, step 22, based on the expert database, performs topological association reasoning and dynamic normalization processing on the multi-source heterogeneous coupled fault data samples to generate standardized fault data samples, specifically including the following steps:

[0052] Step 221: construct a dynamic network tree diagram based on the distribution network model in the expert database, capture the upstream and downstream connection relationships between the measuring points through the real-time topology perception module, map the characteristics of each measuring point in the multi-source heterogeneous coupled fault data sample to the corresponding node of the tree diagram, and obtain a topological correlation feature distribution diagram;

[0053] Step 222: Based on the historical fault logic rules in the expert database, the correlation of each node feature in the topological association feature distribution graph is calculated, and the corresponding correlation is weighted according to the size of the correlation through the attention mechanism and weighted fusion is performed to obtain a weighted fusion feature set;

[0054] Step 223, according to the fault type, adaptively select the normalization template, and convert the weighted fusion feature set into a unified format string containing topology identification, state code, and time sequence label to generate a standardized fault data sample.

[0055] Preferably, in a specific application scenario, when step 221 is implemented, first, the graph model data of the power distribution network is called from the expert library, which contains the topology information of the power distribution network, such as the line direction and the connection relationship of the equipment. Based on this, a hierarchical network construction algorithm is used to generate a dynamic network tree diagram. The root node of the tree diagram corresponds to the power supply point, and each level of the child node corresponds to the substation, the line section, the distribution transformer, the transformer area, and other equipment or areas in turn. At the same time, through the real-time topology perception module (using the depth-first search algorithm), the physical connection relationship of the power distribution network is traversed, and the upstream and downstream connection relationship between each measuring point (such as circuit breaker, fault indicator) is captured in real time, for example, it is determined that a certain circuit breaker is the upstream control equipment of a certain line section. Subsequently, the features (such as current mutation feature, switch state feature) of each measuring point in the multi-source heterogeneous coupled fault data sample are accurately mapped to the corresponding nodes of the dynamic network tree diagram according to their corresponding physical positions, forming a topology-related feature distribution diagram that intuitively reflects the association between the features of each measuring point and the topology position. This diagram can clearly show the distribution of fault features in the topology structure of the power distribution network.

[0056] To this end, step 221 achieves accurate binding of fault data and physical topology of the power distribution network by constructing a dynamic network tree diagram and mapping measuring point features, giving abstract fault features a clear spatial positioning. The power distribution network graph model in the expert library contains complete device connection relationships and hierarchical structures, and the dynamic network tree diagram generated based on this can intuitively reflect the real-time topology form of the power distribution network. The real-time topology perception module uses the depth-first search algorithm to capture the upstream and downstream relationships of the measuring points, ensuring the accuracy of the topology structure. After mapping the measuring point features in the multi-source heterogeneous coupled fault data sample to the corresponding nodes of the tree diagram, the topology-related feature distribution diagram is formed, which can clearly show which nodes (devices) have fault features, where these nodes are located in the topology structure, and what upstream and downstream relationships exist between them, thereby associating the originally isolated fault feature data with the specific physical structure of the power distribution network, providing a spatial dimension reference for subsequent fault reasoning based on topology relationships, and avoiding fault positioning deviation caused by disconnection between features and topology.

[0057] Preferably, in a specific application scenario, step 222 is specifically implemented to extract historical fault logic rules from the expert library, which are rules of association between fault characteristics summarized based on a large number of historical fault cases, for example, “when the upstream circuit breaker of a line section trips and the downstream fault indicator sends a fault signal, the probability of failure of the line section is extremely high”. Based on these rules, the Pearson correlation coefficient algorithm is used to calculate the correlation between each node feature in the topological correlation feature distribution, and the correlation strength between different measurement point features is quantified. Then, the attention mechanism (Bahdanau attention algorithm) is introduced, and each node feature is given a corresponding weight according to its correlation degree, the higher the correlation degree, the greater the weight, for example, the circuit breaker state feature and the fault indicator feature directly associated with the fault section will be given a higher weight. Finally, the node features are weighted and fused according to the weight, and the dispersed features are integrated into a comprehensive weighted fusion feature set, so that the key fault features are highlighted and the influence of secondary features is weakened.

[0058] To this end, in step 222, the accurate selection and efficient fusion of fault features are realized by combining expert experience and intelligent algorithms. The historical fault logic rules in the expert library are the crystallization of knowledge refined based on a large number of actual cases, providing a reliable priori basis for feature correlation analysis; the Pearson correlation coefficient algorithm is used to calculate the correlation between features, which can objectively quantify the correlation strength between different node features and avoid the subjectivity of manual judgment. And the Bahdanau attention mechanism is introduced to give the feature weight, which can dynamically highlight the key features closely related to the fault (such as the switch state and current change of the upstream and downstream of the fault section) and weaken the interference of irrelevant features, so that the fused feature set is more focused on the essence of the fault. This “rule-guided + data-driven” weighted fusion method not only retains the deep understanding of the fault rules by experts, but also realizes the adaptive adjustment of the importance of features through algorithms, solving the problems of multi-source feature redundancy and key information flooding, and providing high-quality core feature input for subsequent normalization processing.

[0059] Preferably, in a specific application scenario, the expert library is pre-provided with normalization templates for different fault types (such as short-circuit fault, ground fault, broken line fault, etc.) when step 223 is specifically implemented, and each template specifies the format, field meaning and coding rules of the data corresponding to the fault type. Using a rule-based adaptive matching algorithm, the fault type is determined according to the key features (such as current waveform features, switch action sequence) extracted from the weighted fusion feature set, for example, when a sudden increase in current is detected and the duration is short, and the protection device acts quickly, it is determined as a short-circuit fault, and the corresponding short-circuit fault normalization template in the expert library is automatically matched. Then, according to the format requirements of the template, the topology-related information in the weighted fusion feature set is converted into topology identifiers (such as standardized coding of line ID, device number), the device state information is converted into a unified state code (such as "10" for closing circuit breaker and "01" for opening), and the time sequence features are converted into time sequence labels containing time stamps. Finally, these information is integrated into a string in a unified format to generate a standardized fault data sample, ensuring that fault data of different sources and types have consistent format and meaning.

[0060] To this end, in step 223, through fault type adaptive matching and unified format conversion, the standardization convergence of multi-source heterogeneous data is achieved, providing a consistent input basis for subsequent fault reasoning. The pre-provided normalization templates in the expert library define a unified data format specification for different fault types (such as short-circuit, ground fault, etc.), which is a standardized framework based on historical fault handling experience. Using a rule-based adaptive matching algorithm, the corresponding template can be accurately matched according to the key features (such as current waveform, switch action time sequence) in the weighted fusion feature set, ensuring the adaptability of normalization processing and fault type. After converting the feature set into a unified format string containing topology identifier, state code, and time sequence label, not only the differences in format and coding of multi-source data (such as differences in switch state description by different systems) are eliminated, but also the topology association, state change, and time sequence characteristics of the fault are preserved through structured labels, making the originally chaotic multi-source data into standardized samples with unified semantics. This standardization process solves the "information island" problem of multi-source data, ensures that the subsequent fault reasoning model can stably receive input with consistent format and clear meaning, avoids model parsing errors caused by chaotic data format, and significantly improves the stability and reliability of the fault identification process.

[0061] Optionally, step 3, the standardized fault data sample is subjected to matching and checking processing to generate checked fault data, specifically including the following steps:

[0062] Step 31, based on the topology logic reasoning module, the correlation of each string in the standardized fault data sample is calculated, and the fault data of the most upstream measuring point is determined as the reference string;

[0063] Step 32, each string in the standardized fault data sample is matched with the reference string layer by layer to mark the normal data matched successfully and the abnormal data matched unsuccessfully;

[0064] Step 33, an artificial checking mechanism is started for the marked abnormal data, the historical fault cases in the expert library and the real-time topology diagram are combined to modify the abnormal data, the normal data and the modified data are integrated to generate the checked fault data.

[0065] Optionally, step 31, based on the topology logical reasoning module, the correlation of each string in the standardized fault data sample is calculated, the fault data of the most upstream measuring point is determined as the reference string, and the step specifically includes the following steps:

[0066] Step 311, based on the network tree diagram constructed by the power distribution network diagram module, the number information of each measuring point in the standardized fault data sample is analyzed by the topology logical reasoning module to determine the upstream and downstream topology relationship between the measuring points, and a topology correlation matrix is obtained.

[0067] Step 312, according to the topology correlation matrix, the correlation of the fault data string of each measuring point is calculated to select the string with the largest correlation from it, and the fault data of the most upstream measuring point corresponding to the string is determined as the reference string.

[0068] Preferably, in a specific application scenario, when step 311 is specifically implemented, based on the network tree diagram (the root node is the power supply point, and the nodes are power substations, lines, and transformer areas, etc. from top to bottom) constructed by the power distribution network diagram module, the topology logical reasoning module calls the depth-first search algorithm to analyze the measuring point number information (such as line ID and equipment code) contained in each string in the standardized fault data sample. By traversing the node hierarchical relationship of the tree diagram, the position of each measuring point in the topology structure is determined, and the upstream and downstream relationship between each measuring point is determined (for example, the circuit breaker of a main line is the upstream measuring point of the branch line switch), and the relationship is quantitatively represented in the form of a matrix to generate a topology correlation matrix-the numerical value of the elements in the matrix reflects the correlation between the corresponding two measuring points in the topology, which provides a logical basis for the topology level for subsequent calculation of string correlation.

[0069] To this end, in step 311, a topological correlation matrix is established to provide a structured topological logical basis for the correlation analysis of fault data. The fault propagation and data correlation of the distribution network are strictly constrained by its physical topological structure. The fault data of different measuring points are not isolated, but follow the connection relationship of upstream and downstream. The topological logical reasoning module uses the depth-first search algorithm to analyze the measuring point number and determine the upstream and downstream relationship, which can accurately capture the logical correlation implied by the physical connection, for example, it is clear that the fault data of a certain line should be closely related to the status data of the upstream circuit breaker. The generated topological correlation matrix presents this correlation in a quantitative form, which converts the abstract topological relationship into a calculable matrix element, provides a clear topological basis for the calculation of the relevance of each measuring point fault data string in the subsequent steps, avoids blind data comparison that is independent of the physical connection, ensures that the calculation result of the relevance conforms to the actual operation logic of the distribution network, and thus lays a foundation for the accurate selection of the reference string.

[0070] Preferably, in a specific application scenario, in step 312, the cosine similarity algorithm is used to calculate the relevance between the fault data strings of each measuring point according to the topological correlation matrix. Each string contains information such as the topological identifier, status code and time sequence label of the measuring point. The cosine similarity algorithm quantifies the matching degree of different strings in content and topological logic by comparing the vector space angle of these information (the smaller the angle, the higher the relevance). In the calculation process, combined with the upstream and downstream relationship determined in step 311, the measuring point at the most upstream position (such as the circuit breaker closest to the power supply side) is focused on. When the fault data string of a certain upstream measuring point has the maximum relevance with most other strings, it is indicated that it is less affected by downstream faults and can reflect the most original fault state, so this string is determined as the reference string as the reference standard for subsequent matching and checking.

[0071] To this end, in step 312, the reliability and representativeness of the reference string are ensured by combining topological correlation and string relevance analysis. The topological correlation matrix has clearly defined the upstream and downstream logical relationship between measuring points, and the cosine similarity algorithm can accurately quantify the matching degree of each fault data string in content (topological identifier, status code, time sequence label). The combination of the two makes the relevance calculation conform to the physical connection law of the distribution network and can reflect the consistency of the data content. The fault data of the measuring point at the most upstream position is selected as the reference because the upstream measuring point is less affected by downstream faults, and its data is closer to the initial state of the fault. When the string of this measuring point has the maximum relevance with most other strings, it is indicated that it can stably reflect the overall trend of the fault, and thus as a reference can reduce the deviation of subsequent matching and checking, and provides a reliable reference for accurate identification of abnormal data.

[0072] Optionally, step 32, each string in the standardized fault data sample is matched with the reference string layer by layer to mark the normal data matched successfully and the abnormal data matched unsuccessfully, specifically including the following steps:

[0073] Step 321, based on the hierarchical relationship of the power distribution network topology, each string in the standardized fault data sample is arranged in turn according to the branch level of the network tree diagram to form a hierarchical string sequence;

[0074] Step 322, each string in the hierarchical string sequence is verified for topological logical consistency with the reference string, if the measurement point state and fault signal represented by the string are not contradictory to the reference string in the topological relationship, it is marked as normal data, otherwise it is marked as abnormal data.

[0075] Preferably, in a specific application scenario, when step 321 is implemented, based on the hierarchical relationship of the power distribution network topology (such as the hierarchical structure of power points-substations-main lines-branch lines-district areas), the breadth-first search algorithm is used to traverse the network tree diagram, and each string in the standardized fault data sample is corresponded to different branch levels of the tree diagram in the order from the root node to the leaf node. For example, the strings corresponding to the root node (power point) are arranged first, and then the strings of the first branch (main line), second branch (branch line), and third branch (district area) are arranged in turn to form a hierarchical string sequence. This arrangement method follows the topological logic of power distribution network fault propagation (fault usually spreads from upstream to downstream), providing a sequence basis in line with physical laws for subsequent layer-by-layer matching, ensuring that the matching process is consistent with the actual fault evolution path.

[0076] Therefore, in step 321, the strings are arranged hierarchically to make the matching process of the fault data highly consistent with the physical logic of power distribution network fault propagation. The topological structure of the power distribution network presents a clear hierarchical relationship (such as the gradual branching from power points to district areas), and the occurrence and impact of faults usually spread along this hierarchy from upstream to downstream. The breadth-first search algorithm is used to arrange the strings according to the branch level of the network tree diagram, so that the data sequence naturally follows the fault evolution law of "upstream response first and downstream spread later", for example, the fault data of the main line is arranged before the branch line. This arrangement method provides a sequence framework in line with physical logic for subsequent layer-by-layer matching, ensuring that each string can be compared with the reference string at the corresponding topological level, avoiding invalid matching across levels due to chaotic order, making the matching process more efficient and more consistent with the actual fault propagation path, and laying a structural foundation for accurately distinguishing normal data from abnormal data.

[0077] Preferably, in a specific application scenario, step 322 is specifically implemented by using a rule-based logical matching algorithm to perform topological logical consistency verification between each string in the hierarchical string sequence and the reference string. During the verification, the topological identifier (such as the correspondence between the upstream and downstream measurement point numbers), the state code (such as whether the switch opening and closing state conforms to the topological constraint, if the upstream circuit breaker is in the opening state in the reference string, the switch state of the corresponding line in the downstream string should be consistent with the power-off logic), and the time sequence label (such as whether the fault signal occurrence time conforms to the upstream and downstream propagation order) of the two are mainly compared. If a string has no logical contradiction with the reference string in the above three aspects (such as the fault signal of the downstream measurement point is later than the upstream and the state matches), it is marked as normal data; if there is a contradiction (such as the upstream circuit breaker is opened but the downstream still shows the live state), it is marked as abnormal data. This process verifies the topological logic binding between each string and the reference string, effectively identifies the inconsistency in the data, and provides a clear target for subsequent correction.

[0078] Therefore, step 322 realizes accurate screening of the authenticity of fault data through topological logical consistency verification, ensuring that the subsequent analysis is based on reliable core data. In the distribution network, the state and fault signal of the upstream and downstream measurement points have strict logical constraints (such as after the upstream circuit breaker is opened, the downstream line should be de-energized), and the rule-based logical matching algorithm can effectively identify abnormal situations that violate such constraints by comparing the topological identifier, the state code, and the time sequence label. For example, if the reference string shows that the upstream circuit breaker has been opened, but the downstream string marks the line as live, this topological logical contradiction will be directly marked as abnormal data. This verification method not only utilizes the inherent rules of the distribution network topology, but also realizes efficient comparison through a standardized string format, avoiding the interference of false data or false signals on fault analysis, locking the abnormal target that needs to be focused on for subsequent manual checking, and improving the accuracy and efficiency of fault data processing.

[0079] Optionally, step 33, starting a manual checking mechanism for the marked abnormal data, combining historical fault cases in the expert library with real-time topology map to modify the abnormal data, integrating normal data and modified data to generate checked fault data, specifically including the following steps:

[0080] Step 331, calling a historical fault case matching the power-off type of the abnormal data from the expert library, extracting the fault feature association rules in the case as a reference for modifying the abnormal data;

[0081] Step 332, combining the real-time distribution network topology map, locating the physical location and upstream and downstream connection relationship of the measurement point corresponding to the abnormal data, and judging the false alarm / misreporting type of the abnormal data and correcting it by analyzing the remote signaling / telemetering data correlation of the measurement point and the neighborhood measurement points;

[0082] Step 333, the marked normal data and the corrected abnormal data are associated and integrated according to the measurement point number, a fault data set containing complete topological association information is constructed, and checked fault data is generated.

[0083] Preferably, in a specific application scenario, when step 331 is implemented, first, the abnormal data is analyzed to determine the type of power failure (such as regional power failure, line power failure, etc.) to which it belongs, and then a keyword-based search algorithm is used to retrieve historical fault cases from the expert library that are highly matched with the power failure type. These cases contain fault characteristics, processing methods, and data correlation rules in similar scenarios, and the fault characteristic correlation rules (such as the correspondence between abnormal data and actual faults in a specific topology, the correction method of common false data) are extracted through case-based reasoning technology (CBR). These rules serve as a reference for the correction of abnormal data, ensuring that the correction process is supported by mature experience and avoiding blind adjustments.

[0084] Therefore, in step 331, the experience accumulation of historical fault cases in the expert library provides a reliable reference standard and logical basis for the correction of abnormal data. Although power distribution network failures are diverse, the fault characteristics and data anomaly patterns of similar power failure types (such as line short circuit and substation power failure) often have similarities, and the fault characteristic correlation rules extracted from historical cases are a summary of these common rules. By using a keyword matching algorithm to accurately retrieve matching historical cases and using case-based reasoning technology to extract rules, technicians can avoid subjective speculation when correcting abnormal data and instead work based on proven practical experience, such as referring to the rule that "false fault signals usually do not match neighboring measurement point data" in similar cases to determine the nature of the current abnormal data. This approach not only improves the accuracy of abnormal data correction but also reduces the risk of correction errors due to a lack of reference, laying a high-quality foundation for subsequent data integration.

[0085] Preferably, in a specific application scenario, when step 332 is implemented, the number of the corresponding measurement point of the abnormal data is mapped to the real-time power distribution network topology map using a geographic information system (GIS), accurately locating the physical location of the measurement point (such as the tower number of a certain line) and its upstream and downstream connected devices (such as adjacent circuit breakers and fault indicators). An association rule mining algorithm (such as the FP-growth algorithm) is used to analyze the correlation between the measurement point and its neighboring measurement points in terms of remote signaling data (such as switch status) and remote measurement data (such as current and voltage), for example, to determine whether the "fault signal" shown in the abnormal data is consistent with the power failure status of the neighboring measurement points. If there is no fault signal in the neighborhood, it may be a false alarm, and if there is a fault signal in the neighborhood but not at the measurement point, it may be a missed alarm. After determining the type of abnormality, the data is corrected based on the rules extracted in step 331, such as deleting false signals and supplementing missing signals.

[0086] To this end, in step 332, by combining the spatial positioning of the real-time topology graph model with the correlation analysis of the neighborhood data, the accurate correction of abnormal data is realized, and it is ensured that the correction result conforms to the actual operation state of the power distribution network. The real-time topology graph model can accurately locate the physical position of the abnormal data corresponding to the measuring point and the upstream and downstream connection relationship, providing a spatial reference for analyzing the rationality of the data. The correlation rule mining algorithm analyzes the correlation of the remote signaling and telemetry data of the measuring point and the neighborhood measuring point, which can verify the authenticity of the abnormality (such as false alarm or missed alarm) from the data level. This double-checking method of "spatial positioning + data correlation" not only avoids blind correction that is divorced from the actual topology, but also improves the accuracy of abnormal type judgment through the mutual verification of neighborhood data, so that the corrected abnormal data can truly reflect the fault state and provide reliable input for subsequent data integration.

[0087] Preferably, in a specific application scenario, when step 333 is specifically implemented, an integration algorithm based on primary key correlation is used to associate and integrate the marked normal data and the corrected abnormal data by taking the measuring point number as the unique identifier. In the integration process, the data of the same measuring point number is quickly matched through the hash mapping technology, ensuring that the information of each measuring point is consistent in terms of time sequence and topology relationship (such as the time sequence data of the same measuring point is sorted by time, and the upstream and downstream measuring point data is associated by topology level). Finally, a fault data set containing complete topology correlation information (such as upstream and downstream connection relationship, data correspondence) of all measuring points is constructed, i.e. the checked fault data, which can accurately reflect the fault state of the power distribution network and provide a reliable data basis for subsequent fault reasoning.

[0088] To this end, in step 333, through systematic association and integration, the integrity and consistency of the fault data are realized, providing a comprehensive and reliable data source for subsequent fault analysis. The association and integration are performed by taking the measuring point number as the unique identifier, and the hash mapping technology is used to ensure the efficiency and accuracy of data matching, avoiding the confusion of data from different measuring points. In the integration process, the original effectiveness of the normal data is retained, and the real information of the corrected abnormal data is also included. At the same time, through the analysis of the topology correlation information, the entire data set forms an organic whole - the data of the upstream and downstream measuring points are echoed according to the physical connection relationship, and the time sequence data of the same measuring point are arranged in order according to the time sequence. This integration method eliminates the redundancy and contradiction between the data, and constructs a data set that can completely reflect the fault state of the power distribution network, ensuring that the subsequent fault location, cause analysis and other work based on this data can be carried out based on comprehensive and consistent information, improving the overall efficiency and accuracy of fault handling.

[0089] Optionally, in step 4, the checked fault data is subjected to graph data conversion processing according to the power distribution network graph model to generate fault graph data, specifically including the following steps:

[0090] Step 41, according to the power distribution network diagram model, the circuit breaker, sectionalizing switch, fault indicator and other devices in the checked fault data are abstracted as nodes of the graph data, the line sections between the devices are abstracted as edges of the graph data, and the topological structure of the graph is determined;

[0091] Step 42, based on the checked fault data, a node feature matrix is constructed, each element in the matrix corresponding to a feature of a node;

[0092] Step 43, an adjacency matrix is constructed according to the connection relationship between the nodes, the matrix elements are 0 or 1, where 1 indicates that there is a topological connection relationship between the corresponding nodes, and 0 indicates that there is not, and the node feature matrix and the adjacency matrix are fused to generate the fault graph data.

[0093] Optionally, in step 42, based on the checked fault data, a node feature matrix is constructed, each element in the matrix corresponding to a feature of a node, specifically: the node feature matrix is composed of node feature vectors represented by N measuring points, and each node feature vector is a 0-1 discrete sequence. If the element value is 1, it indicates that the measuring point sends a fault remote signaling signal, or the corresponding measuring point directly connected to the measuring point sends a fault remote signaling signal. If the value is 0, it indicates that no fault remote signaling signal is sent.

[0094] Preferably, in a specific application scenario, when step 41 is specifically implemented, the device layout and connection relationship contained in the power distribution network diagram model are taken as the benchmark, the graph theory abstraction algorithm is adopted, the circuit breaker, sectionalizing switch, fault indicator and other entity devices in the checked fault data are converted into nodes in the graph data, and each node is associated with an actual device through a unique identifier (such as a device number). At the same time, the line sections connecting these devices are converted into edges in the graph data, and the attributes of the edges can be attached to the basic parameters (such as length, model) of the lines. Through this abstraction, the physical topological structure of the power distribution network is converted into a "node-edge" graph structure, where the nodes represent the devices and the edges represent the physical connection between the devices, thereby determining the topological structure of the graph and converting the complex entity relationship of the power distribution network into a graph data model convenient for calculation and analysis, thereby providing a structural framework for the construction of the subsequent feature matrix and adjacency matrix.

[0095] To this end, in step 41, the power distribution network physical equipment and connection relationship are abstracted into a "node-edge" structure of graph data, realizing the mathematical and structured expression of the complex power distribution network topology. The power distribution network contains a large number of circuit breakers, switches and other equipment and complex line connections, and it is difficult to directly analyze the physical structure, while the graph modeling method can convert entity equipment into nodes and line sections into edges. This abstraction retains the topological correlation nature between devices (such as which devices are directly connected and how to connect). For example, the circuit breaker on a line and the adjacent section switch are connected by a line, which is represented in graph data as two nodes connected by an edge, clearly reflecting the physical connection relationship between the two. This conversion changes the topology of the power distribution network from an intuitive physical graph to a calculable graph structure, laying the foundation for subsequent use of graph algorithms (such as node feature extraction and adjacent relationship analysis) to process fault data, and simplifying the analysis difficulty of the correlation between devices, allowing technicians to focus more on the correlation between fault features and topology.

[0096] Preferably, in a specific application scenario, in step 42, based on the state information of each device in the checked fault data, a feature discretization coding algorithm is used to construct a node feature matrix. The number of rows of the matrix corresponds to the total number of device nodes (N) in the power distribution network, and the number of columns corresponds to the feature dimension that needs to be represented for each node (such as fault signal state, associated device signal state, etc.). For each node, its feature vector is coded as a 0-1 discrete sequence: if the node itself sends a fault telesignaling signal (such as a circuit breaker trip signal), or the device directly connected to it sends a fault telesignaling signal (such as the alarm signal of the adjacent fault indicator), the element value at the corresponding position in the feature vector is 1; if the above conditions are not met, the element value is 0. For example, a section switch node does not send a fault signal itself, but the fault indicator directly connected to it sends a signal, so the corresponding element in the node feature vector is recorded as 1, forming a complete node feature matrix, and realizing the quantitative representation of the fault-related state of each device.

[0097] To this end, in step 42, the fault telecommunication signal state of the node is converted into a 0-1 discrete sequence by using a feature discretization encoding algorithm. This makes the originally abstract device fault state information accurately quantified. For example, a fault indicator sends a fault signal, and the corresponding position element in the feature vector is 1, which clearly and intuitively indicates the device fault state, providing clear data basis for subsequent fault analysis. In addition, in step 42, not only the fault signal of the node itself is considered, but also the fault telecommunication signal of the node directly connected to it. In this way, the fault situation of the local area where the node is located can be fully reflected, and the associated fault information between devices is integrated. For example, although the sectionalizing switch itself does not send a fault signal, when the fault indicator connected to it sends a signal, the corresponding position of the sectionalizing switch node feature vector is also 1, so that the technical personnel can grasp the fault situation from a more macroscopic point of view. Furthermore, in step 42, the node feature matrix formed is a structured data form, which is naturally adapted to subsequent graph data processing and analysis algorithms (such as fusion with adjacency matrix, graph neural network analysis, etc.). This structured expression facilitates efficient processing and analysis by computer, improves the efficiency and accuracy of subsequent operations such as fault diagnosis and positioning, and provides strong support for intelligent analysis of power distribution network faults.

[0098] Optionally, in step 43, an adjacency matrix is constructed according to the connection relationship between nodes, the matrix elements are 0 or 1, where 1 indicates that there is a topological connection relationship between the corresponding nodes, and 0 indicates that there is no connection. The node feature matrix and the adjacency matrix are fused to generate fault graph data, specifically: Adjacency matrix A ∈ R N×N ,N is the total number of measurement points in the power distribution network, and element a ij = 1 indicates that the i-th node and the j-th node have a direct topological connection, and a N = 0 indicates no connection; the node feature matrix X = [x1, x2, …, x i ], wherein the feature vector x i1 of the i-th node is [x i2 , x iN , …, x T ] in , x in = 1 indicates that the i-th node itself or the n-th node directly connected to it sends a fault telecommunication signal, and x ij = 0 indicates that it does not send; the adjacency matrix and the node feature matrix are fused according to the topological association to form a graph structure data containing node features and node connection relationship as fault graph data.

[0099] Preferably, in a specific application scenario, when step 43 is implemented, first, a topological relationship mapping algorithm is used to construct an adjacency matrix. According to the actual physical connection between nodes in the power distribution network, a matrix of size $N×N$ is created (N is the total number of measurement points in the power distribution network), wherein each element a ijrepresents the connection relationship between the ith node and the jth node. All node pairs are traversed, and if the ith node and the jth node have a direct topological connection (such as being directly connected through a line), the corresponding element a ij and a ji (under the condition of an undirected graph) is assigned a value of 1; if there is no direct connection, it is assigned a value of 0. In this way, the adjacency matrix completely records the topological connection relationship between the nodes in the power distribution network. Then, based on the node feature matrix X that has been constructed, where the feature vector x i of each node reflects the fault telecommunication signal state of the node and its directly connected nodes. A matrix fusion algorithm is used to fuse the adjacency matrix and the node feature matrix according to the topological association. Specifically, the connection relationship of the nodes determined by the adjacency matrix is used as a guide to integrate the feature information of each node in the node feature matrix into the corresponding connection relationship. For example, if a ij = 1 in the adjacency matrix, indicating that nodes i and j are connected, the information in the feature vectors x i and x j of nodes i and j is integrated and associated to finally form a graph structure data that contains both the fault feature information of the nodes themselves and the topological connection relationship between the nodes. This data is the fault graph data, which provides a comprehensive and structured data basis for subsequent fault analysis and processing based on graph structure.

[0100] Therefore, in step 43, by fusing the node feature matrix and the adjacency matrix, fault graph data that combines device state information and topological association relationship is constructed, providing a complete and adaptive data basis for fault analysis based on graph structure. The adjacency matrix uses a topological relationship mapping algorithm to accurately record the connection relationship between nodes using 0-1 encoding, clearly presenting the physical topological structure of the power distribution network. The node feature matrix quantifies the fault state and associated signals of each node through 0-1 discrete sequences. The fusion of the two is not a simple concatenation, but an organic combination of device features and connection relationships based on topological association - for example, when two nodes are connected (adjacency matrix element is 1), their feature vectors will be associated with each other according to this connection relationship, so that the graph data contains not only the feature information of "which devices have fault signals", but also the topological information of "how these devices are connected". This fusion method perfectly adapts to the needs of graph data processing algorithms such as graph neural networks, allowing the algorithm to utilize both node features and topological relationships for reasoning, such as locating the fault source by analyzing the propagation path of fault signals between connected nodes, avoiding the analysis bias caused by relying solely on features or topology, and significantly improving the accuracy and depth of fault analysis.

[0101] Optionally, in step 5, based on the fault graph data, a fault reasoning model is used to reason about the fault of the power distribution network to obtain a suspected power distribution network fault, which specifically includes the following steps:

[0102] Step 51, based on the graph convolution layer in the fault reasoning model, topological correlation feature extraction and aggregation processing are performed on the fault graph data to generate a feature matrix containing multi-scale topological correlations;

[0103] Step 52, based on the fully connected hidden layer in the fault reasoning model, non-linear mapping and dimension compression processing are performed on the feature matrix containing multi-scale topological correlations of the graph convolution layer to generate a fault feature vector;

[0104] Step 53, based on the fully connected output layer in the fault reasoning model, multi-label binary classification processing is performed on the fault feature vector to generate a suspected power distribution network fault.

[0105] Optionally, in step 51, based on the graph convolution layer in the fault reasoning model, topological correlation feature extraction and aggregation processing are performed on the fault graph data to generate a feature matrix containing multi-scale topological correlations, specifically: the graph convolution layer aggregates node self-features and topological correlation features of 1st-order and 2nd-order neighborhood nodes layer by layer through multi-layer graph convolution operations to determine fault state information of a single measurement point and correlation rules between measurement points in different hierarchical topological structures and generate a multi-scale topological correlation feature matrix accordingly.

[0106] Preferably, in a specific application scenario, in step 51, based on the graph convolution layer in the fault reasoning model, a graph convolution neural network (GCN) algorithm is used to process the fault graph data, and topological correlation features are extracted and aggregated through multi-layer graph convolution operations. The calculation of each layer of graph convolution follows the formula: In the formula, the physical meanings of the letters are as follows: H (l) represents the input feature matrix of the lth layer of graph convolution (each row corresponds to a feature vector of a node, and the elements reflect the fault telesignaling signal state of the node, etc.); is the adjacency matrix with self-loops (A is the adjacency matrix generated in step 43, I N is an N-order identity matrix, which ensures that the node can aggregate its own features); is the degree matrix of is used for normalization processing to avoid feature weight imbalance due to differences in the number of node connections); W (l) is the trainable weight matrix of the lth layer (used for linear transformation of aggregated features to extract more abstract correlation features); σ is an activation function (such as ReLU, which introduces nonlinearity to enhance the expression ability of the model); H (l+1) ​is the output feature matrix of the l+1 layer. Through multiple layers of such operations, the graph convolution layer can aggregate the 1-order neighborhood (directly connected nodes), 2-order neighborhood (nodes connected to the directly connected nodes), and other multi-scale topological correlation features layer by layer. For example, the first layer of graph convolution aggregates the features of a certain fault indicator and the circuit breaker directly connected thereto, the second layer further aggregates the features of other fault indicators connected to the circuit breaker, and finally a feature matrix containing multi-scale topological correlation is generated, which not only retains the fault state information of a single node, but also embodies the correlation between nodes in different hierarchical topological structures, thereby providing a rich feature basis for subsequent fault reasoning.

[0107] Therefore, in step 51, through multi-layer topological correlation feature extraction and aggregation of the graph convolution layer, multi-scale and deep-level capture of the power distribution network fault features is realized, thereby laying a foundation for accurate fault reasoning. The power distribution network fault has a significant topological propagation characteristic, i.e., a fault in a certain section will affect multiple measurement points upstream and downstream through line connection, and the features of a single node cannot fully reflect such correlation. The graph convolution layer uses the graph convolution neural network (GCN) algorithm to aggregate the features of the node itself and the 1-order and 2-order neighborhood nodes layer by layer through multi-layer convolution operation, for example, the first layer aggregates the state correlation of the directly connected circuit breaker and fault indicator, and the second layer further incorporates the features of the adjacent nodes of these devices, thereby forming a multi-scale topological correlation feature matrix. This aggregation method not only retains the fault state details of a single measurement point, but also captures the correlation between measurement points in different hierarchical topological structures (such as the linkage relationship between a line fault and upstream switch action and downstream load change), thereby breaking through the limitation of traditional feature extraction which only focuses on local information. Meanwhile, the introduction of the self-loop adjacency matrix and degree matrix normalization processing ensures the equalization of the weights of the features of nodes with different connection densities during aggregation, thereby avoiding the problem of excessive dominance of the features of some nodes due to a large number of connections, and making the extracted features more consistent with the actual physical logic of power distribution network fault propagation, thereby significantly improving the accuracy and robustness of subsequent fault reasoning.

[0108] Optionally, in step 52, the multi-scale topological correlation feature matrix contained in the graph convolution layer is subjected to nonlinear mapping and dimension compression processing based on the fully connected hidden layer in the fault reasoning model, thereby generating a fault feature vector, specifically: the fully connected hidden layer uses the ReLU activation function as a nonlinear mapping tool to operate on the multi-scale topological correlation feature matrix, thereby mapping the high-dimensional features to a low-dimensional space through a weight matrix while retaining key fault correlation features, so as to generate a fault feature vector.

[0109] Preferably, in a specific application scenario, step 52 is specifically implemented, based on the fully connected hidden layer in the fault reasoning model, the multi-scale topological correlation feature matrix output by the graph convolution layer is processed using a multi-layer perception (MLP) algorithm, and a fault feature vector is generated through nonlinear mapping and dimension compression. The core calculation follows the formula: F = ReLU(H·W + b), where the physical meaning of each letter is: H represents the feature matrix output by the graph convolution layer (containing multi-scale topological correlation features of N nodes, and each row corresponds to a high-dimensional feature vector of a node); W is a trainable weight matrix of the fully connected layer (dimension: "number of columns of H x target low-dimensional feature number", used to compress high-dimensional features to a low-dimensional space); b is a bias vector (used to adjust the feature distribution and enhance the model generalization ability); ReLU is an activation function (expression: ReLU(x) = max(0, x), which introduces nonlinearity by filtering negative features, solving the problem that linear models cannot capture complex fault correlations); F is the output fault feature vector (dimension is reduced and key fault correlation features are retained). For example, if the feature matrix output by the graph convolution layer has a dimension of 1000 columns (high-dimensional features), it is mapped to 128 columns through the weight matrix W, then F is a 128-dimensional feature vector, which not only removes redundant information, but also retains the nonlinear correlation between fault features (such as the complex mapping relationship between a node fault and multiple downstream node abnormalities), providing a simplified and effective feature input for subsequent fault classification.

[0110] To this end, in step 52, through nonlinear mapping and dimension compression of the fully connected hidden layer, the key fault features are retained while the reasoning efficiency and generalization ability of the model are improved. The fault features of the power distribution network often contain a large amount of redundant information (such as repeated signals reflecting the same fault at multiple measurement points), and the fully connected hidden layer uses a multi-layer perception (MLP) algorithm to map the high-dimensional feature matrix to a low-dimensional space through a weight matrix, effectively eliminating redundancy and reducing subsequent computational complexity - for example, compressing 1000-dimensional multi-scale topological features to 128-dimensional, which not only reduces the parameter size, but also speeds up the reasoning. At the same time, the ReLU activation function introduces nonlinearity (filters negative features) to solve the problem that linear models cannot capture complex fault correlations, such as accurately depicting the nonlinear mapping relationship between "a node fault and multiple downstream node abnormal states". This processing makes the generated fault feature vector not only simplified but also retains key correlation information, avoiding the risk of overfitting, so that the model can better adapt to fault patterns under different power distribution network topologies, providing high-quality input features for subsequent fault classification.

[0111] Optionally, step 53, based on the full connection output layer in the fault reasoning model, performs multi-label binary classification processing on the fault feature vector to generate the inferred power distribution network fault, specifically: the full connection output layer uses a Sigmoid activation function to perform mapping operation on the fault feature vector, and outputs a fault probability value between 0 and 1 for each line section in the power distribution network; compare the fault probability value of each section with the set threshold value, if the probability value is greater than or equal to the threshold value, it is determined as 1, indicating that there is a fault, and if it is less than the threshold value, it is determined as 0, indicating no fault, and finally a sequence composed of 0 and 1 is formed, each position corresponds to the fault state of a line section, to generate the inferred power distribution network fault.

[0112] Preferably, in a specific application scenario, when step 53 is specifically implemented, based on the full connection output layer in the fault reasoning model, a multi-label classification algorithm is used to process the fault feature vector output by the full connection hidden layer, and the inferred power distribution network fault is generated through multi-label binary classification. The core calculation follows the formula: P = Sigmoid(F·W o +b o ), where the physical meaning of each letter is: F represents the fault feature vector output by the full connection hidden layer (including the compressed key fault correlation feature); W o is the trainable weight matrix of the output layer (dimension: “dimension of F x total number of line sections”, used to map the feature vector to the fault probability space of each line section); b o is the output layer bias vector (used to calibrate the probability distribution and improve the classification accuracy); Sigmoid is the activation function (expression: map the linear output to the 0-1 interval to get the fault probability of each line section); P is the output probability vector (each element corresponds to the fault probability of a line section). Then, compare each element in the probability vector P with the set threshold value θ (such as 0.5), if P i ≥ θ, it is determined that the i-th line section has a fault (marked as 1); if P i < θ, it is determined that there is no fault (marked as 0), and finally a sequence composed of 0 and 1 is formed, each position corresponds to the fault state of a line section, to realize accurate inference of the power distribution network fault. For example, the probability value of a certain line section is 0.85 (greater than the threshold value 0.5), which is marked as 1, indicating that the section is inferred to have a fault. This way can handle the fault judgment of multiple line sections at the same time, and adapt to the multi-section fault scenario that may occur in the power distribution network.

[0113] To this end, in step 53, through the multi-label binary classification processing of the fully connected output layer, accurate and parallel judgment of the fault states of multiple line sections of the power distribution network is realized, which adapts to the multiple fault scenarios that may exist in the power distribution network. The fully connected output layer uses a Sigmoid activation function to map the fault feature vector into a fault probability value between 0 and 1. This probabilistic output not only intuitively reflects the confidence of the existence of a fault in each line section, but also realizes the conversion from probability to determined fault state (0 or 1) by setting a threshold (such as 0.5), taking into account the flexibility and certainty of reasoning. For example, for the simultaneous occurrence of line A short circuit and line B ground fault in the power distribution network, this step can output the fault states of the two sections simultaneously through multi-label classification, avoiding the limitation of traditional single-label classification which can only judge a single fault. At the same time, the trainable weight matrix and bias vector can adapt to the fault modes of different power distribution network topologies, making the model more general in identifying various faults (such as transient faults and permanent faults). The final 0-1 sequence clearly corresponds to the fault state of each line section, providing direct and clear fault location basis for operation and maintenance personnel, and improving the efficiency and accuracy of power distribution network fault handling.

[0114] As shown in Figure 2 The application also provides a device for identifying power distribution network faults based on artificial intelligence, which comprises:

[0115] An acquisition unit is configured to acquire multi-source fault characterization information of a power distribution network, wherein the multi-source fault characterization information comprises telemetry data and telesign data.

[0116] A standardization unit is configured to perform expert library-based normalization processing on the multi-source fault characterization information to generate standardized fault data samples.

[0117] A checking unit is configured to perform matching checking processing on the standardized fault data samples to generate checked fault data.

[0118] A feature extraction unit is configured to perform graph data conversion processing on the checked fault data according to a power distribution network graph model to generate fault graph data.

[0119] A fault reasoning unit is configured to perform reasoning on the fault of the power distribution network based on the fault graph data through a fault reasoning model to obtain a suspected fault of the power distribution network.

[0120] The specific implementation of each unit can be referred to the description of the above Figure 1 , which will not be repeated here.

[0121] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, it is intended that such additions and modifications be included within the scope of the application. It is the following claims, including any amendments thereto, which define the scope of the application.

[0122] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for identifying distribution network faults based on artificial intelligence, characterized in that: include: Step 1: Obtain multi-source fault characterization information of the distribution network, where the multi-source fault characterization information includes telemetry data and telesignaling data; Step 2: normalize the multi-source fault characterization information based on the expert database to generate standardized fault data samples; Step 3: Perform matching and verification processing on the standardized fault data samples to generate verified fault data; Step 4: Based on the distribution network diagram model, the verified fault data is converted into diagram data to generate fault diagram data; Step 5: Based on the fault graph data, the fault inference model is used to infer the fault of the distribution network and obtain the inferred distribution network fault.

2. The method according to claim 1, characterized in that Step 1: Obtain multi-source fault characterization information of the distribution network. The multi-source fault characterization information includes telemetry data and telesignaling data. Specifically, the steps include: Step 11: Collect telemetry data and telesignaling data within the set time period of the fault occurrence from the dispatching automation system, distribution network automation system, power consumption information collection system, and distribution line online monitoring system; Step 12: Extract features from the collected telemetry data and telesignaling data, extract the numbers corresponding to the line name, substation name, circuit breaker name, switch status 0-1 distribution and fault section number, obtain multi-source coupled fault characterization information labels and fault characterization metadata, and generate multi-source fault characterization information.

3. The method according to claim 1, characterized in that Step 2: Perform normalization processing on the multi-source fault characterization information based on the expert database to generate standardized fault data samples, which specifically includes the following steps: Step 21: Perform multi-dimensional feature analysis and coupling processing on the multi-source fault characterization information to obtain multi-source heterogeneous coupled fault data samples; Step 22: Based on the expert database, perform topological association reasoning and dynamic normalization processing on the multi-source heterogeneous coupled fault data samples to generate standardized fault data samples.

4. The method according to claim 1, wherein Step 3: Perform matching and verification processing on the standardized fault data samples to generate verified fault data, which specifically includes the following steps: Step 31: Based on the topological logic reasoning module, calculate the correlation of each character string in the standardized fault data sample, and determine the fault data of the most upstream measurement point as the benchmark character string; Step 32: Match each character string in the standardized fault data sample with the reference character string layer by layer to mark normal data with successful matching and abnormal data with failed matching; Step 33: Start a manual verification mechanism for the marked abnormal data, correct the abnormal data by combining historical fault cases in the expert database and the real-time topology model, and integrate the normal data and the corrected data to generate verified fault data.

5. The method according to claim 1, wherein Step 4: Based on the distribution network diagram model, the verified fault data is converted into diagram data to generate fault diagram data, which specifically includes the following steps: Step 41: Based on the distribution network graph model, devices such as circuit breakers, section switches, and fault indicators in the verified fault data are abstracted as nodes of the graph data, and line sections between the devices are abstracted as edges of the graph data to determine the topological structure of the graph; Step 42: Based on the verified fault data, a node feature matrix is ​​constructed, where each element in the matrix corresponds to a feature of a node; Step 43: Construct an adjacency matrix based on the connection relationship between nodes. The matrix elements are 0 or 1, where 1 indicates that there is a topological connection relationship between the corresponding nodes, and 0 indicates that there is no topological connection relationship. The node feature matrix and the adjacency matrix are fused to generate fault graph data.

6. The method according to claim 1, characterized in that Step 5: Based on the fault graph data, the fault inference model is used to infer the fault of the distribution network and obtain the inferred distribution network fault. The specific steps include: Step 51: Based on the graph convolution layer in the fault inference model, perform topological correlation feature extraction and aggregation processing on the fault graph data to generate a feature matrix containing multi-scale topological correlations; Step 52: Based on the fully connected hidden layer in the fault inference model, perform nonlinear mapping and dimension compression on the feature matrix containing multi-scale topological associations in the graph convolution layer to generate a fault feature vector; Step 53: Based on the fully connected output layer in the fault inference model, perform multi-label binary classification processing on the fault feature vector to generate an inferred distribution network fault.

7. The method according to claim 6, characterized in that Step 51: Based on the graph convolution layer in the fault inference model, the fault graph data is subjected to topological correlation feature extraction and aggregation processing to generate a feature matrix containing multi-scale topological correlations. Specifically, the graph convolution layer aggregates the node features and the topological correlation features of the first-order and second-order neighboring nodes layer by layer through multi-layer graph convolution operations to determine the fault status information of a single measuring point and the correlation rules between measuring points in different levels of topological structures, and generates a feature matrix of multi-scale topological correlations based on this.

8. The method according to claim 6, characterized in that Step 52: Based on the fully connected hidden layer in the fault inference model, nonlinear mapping and dimensionality compression are performed on the feature matrix of the graph convolution layer containing multi-scale topological associations to generate a fault feature vector. Specifically, the fully connected hidden layer uses the ReLU activation function as a nonlinear mapping tool to operate on the feature matrix containing multi-scale topological associations, and maps high-dimensional features to low-dimensional space through the weight matrix, while retaining key fault association features to generate a fault feature vector.

9. The method according to claim 6, characterized in that Step 53: Based on the fully connected output layer in the fault inference model, perform multi-label binary classification processing on the fault feature vector to generate an inferred distribution network fault. Specifically, the fully connected output layer uses the Sigmoid activation function to perform a mapping operation on the fault feature vector, and outputs a fault probability value between 0 and 1 for each line section in the distribution network; the fault probability value of each section is compared with the set threshold. When the probability value is greater than or equal to the threshold, it is judged as 1, indicating that a fault exists; when it is less than the threshold, it is judged as 0, indicating that there is no fault. The final sequence consists of 0 and 1, and each position corresponds to the fault state of a line section to generate the inferred distribution network fault.

10. A device for identifying distribution network faults based on artificial intelligence, characterized in that: include: An acquisition unit, configured to acquire multi-source fault characterization information of the distribution network, the multi-source fault characterization information including telemetry data and telesignaling data; A normalization unit, used to perform normalization processing on multi-source fault characterization information based on an expert library to generate standardized fault data samples; A verification unit, configured to perform matching verification processing on the standardized fault data samples to generate verified fault data; A feature extraction unit is used to perform graph data conversion processing on the verified fault data according to the distribution network graph model to generate fault graph data; The fault reasoning unit is used to reason about the fault of the distribution network based on the fault graph data through the fault reasoning model to obtain the inferred distribution network fault.

Images