Power distribution network fault accurate positioning system and method based on multi-source data fusion

Abstract

The invention relates to the technical field of digital data processing, in particular to a power distribution network fault accurate positioning system and method based on multi-source data fusion. The positioning system comprises an Agent unit, a large model LLM unit, a thinking chain knowledge base and a multi-source database. The Agent unit comprises a thinking chain module, an SQL (Structured Query Language) retrieval module and a fault positioning module; the thinking chain module is in bidirectional data communication with the large model LLM unit and the thinking chain knowledge base; the SQL retrieval module is in bidirectional data communication with the multi-source database and the large model LLM unit; the fault positioning module is in bidirectional data communication with the large model LLM unit; the multi-source database is in one-way data communication with the thinking chain module, and the thinking chain module serves as a data receiving end; the scheme is not limited to the types of the data in the multi-source database, and fault positioning can be carried out on the obtained data and the thinking chain.

Description

Technical Field

[0001] This patent relates to the field of digital data processing technology, specifically to a real-time fault location system and method for power distribution networks based on multi-source data fusion. Background Technology

[0002] Artificial intelligence has begun to be applied to analyze data from power distribution networks to diagnose faults. However, currently, large-scale models are only used for simple fault location in relatively simple scenarios. For example, large-scale models achieve real-time fault location by analyzing electricity consumption data collected in the power distribution network in real time and identifying abnormal data. Existing technologies have two drawbacks:

[0003] (i) The fault location logic of existing technologies is pre-set by technical personnel, so it can only perform fault location according to the existing logic. However, once a new situation occurs, causing the existing logic's fault location method to become incompatible, effective fault location cannot be achieved.

[0004] (ii) The complexity of the distribution network determines that the data related to the distribution network is complex and multi-sourced. However, existing technologies cannot integrate multi-source data to comprehensively locate faults in the distribution network in real time. Summary of the Invention

[0005] To address the limitation of existing technologies that can only utilize a limited set of data by technicians for fault location analysis, this method can effectively utilize multi-source data by identifying a fault-related thought chain and interacting with multi-source data in a multi-source database. Based on the identified thought chain and the data from the required multi-source database, fault location can be performed.

[0006] In a first aspect, the present invention provides a precise fault location system for distribution networks based on multi-source data fusion, comprising:

[0007] Agent unit, large model LLM unit, mind chain knowledge base, multi-source database;

[0008] The Agent unit includes: a thinking chain module, an SQL retrieval module, and a fault location module; the thinking chain module is bidirectionally connected to the large model LLM unit and the thinking chain knowledge base; the SQL retrieval module is bidirectionally connected to the multi-source database and the large model LLM unit; the fault location module is bidirectionally connected to the large model LLM unit; the multi-source database is unidirectionally connected to the thinking chain module, with the thinking chain module acting as the data receiving end;

[0009] The thought chain module is used to receive multi-source data related to abnormal events in the distribution network, and based on the received multi-source data, retrieve relevant thought chain instances in the thought chain knowledge base, construct a prompt, send the prompt to the LLM, and request the LLM to return thought chain data.

[0010] The SQL retrieval module is used to extract text information of the steps to be retrieved from the thought chain returned by the LLM.

[0011] Based on the extracted text information corresponding to each step of the thought chain, SQL commands are generated to query the corresponding data in the multi-source database, and the query is performed in the multi-source database to return the multi-source data corresponding to each step of the thought chain.

[0012] The fault location module is used to receive the thought chain generated by LLM from the thought chain module and the multi-source data corresponding to each step of the thought chain from the SQL retrieval module, generate a prompt, and send it to LLM, thereby determining the fault location of the distribution network through LLM analysis.

[0013] Secondly, the present invention provides a real-time fault location method for power distribution networks based on multi-source data fusion, implemented on the aforementioned positioning system, comprising the following steps:

[0014] S100: Building a mind chain knowledge base;

[0015] S200: The agent acquires multi-source fault data related to distribution network faults;

[0016] S300: Retrieved multiple thought chain information related to the fault in this distribution network from the thought chain knowledge base;

[0017] S400: Using multi-source fault data and the matched multiple thought chain information, a prompt is constructed. This prompt is used to enable LLM to generate a thought chain for fault localization, resulting in the final thought chain for fault localization.

[0018] S500: Extracts text from the thought chain for database queries, and then uses Text2SQL technology to obtain data from multiple source databases;

[0019] S600: Combines the thought chain and data obtained from multiple source databases into a prompt, enabling the LLM to locate faults.

[0020] Preferably, each piece of data in the mind chain knowledge base in S100 includes two parts:

[0021] Multi-source data related to distribution network faults can be encoded by a model-based encoder to obtain word vector features or sentence vector features, which can be used as index information for retrieval and matching.

[0022] The thought process for fault location in the distribution network based on reasoning using this multi-source data.

[0023] Preferably, the specific process of constructing the data in the mind chain knowledge base is as follows:

[0024] S110: Obtain historical records of fault location performed by maintenance personnel;

[0025] S120: Based on historical records, identify historical faults and determine the step-by-step thought process of maintenance personnel from the occurrence of the fault to the completion of maintenance, the results of fault location, and the maintenance results. Based on this information, thought chain data related to this fault can be generated. The thought process is given according to the actual maintenance work process. The results of fault location include: locating the specific fault location, that is: what equipment / line in what location failed, and the cause of the fault.

[0026] Preferably, the following process follows S120:

[0027] S130: Using the data in the knowledge base built in S120 as examples, when a fault is known, combine the fault-related data and examples into a prompt and input it into the LLM;

[0028] S140: Enable LLM to generate a thought chain data;

[0029] S150: The generated thought chain data is used as a reference for troubleshooting this fault. If the generated thought chain data is found to be correct after on-site troubleshooting and the fault can be located according to this thought chain data, then the relevant data and thought chain data of this fault will be added to the thought chain knowledge base.

[0030] Preferably, the multi-source data in S100 is encoded by an encoder based on the BERT model.

[0031] Preferably, in S300, the thought chain module uses a BERT-based encoder to encode the multi-source fault data to obtain corresponding word vectors or sentence vector features. These vector features are then used to perform matching in the thought chain knowledge base to obtain several thought chain information links showing similarity to power distribution network faults.

[0032] In this solution, a thought chain knowledge base is first constructed. Then, when a distribution network fault occurs, multi-source fault data related to the fault is acquired, and multiple thought chain information associated with the fault are retrieved from the thought chain knowledge base. Using the multi-source fault data and the matched thought chain information, an LLM (Local Management Model) is used to generate the final thought chain for fault location. Next, text for database queries is extracted from the thought chain, and Text2SQL technology is used to retrieve the required data from the multi-source database. Finally, the thought chain and the required data are combined into a prompt and input into the LLM for fault location.

[0033] Therefore, this solution is not limited to the type of fault or the type of data in the multi-source database. This solution can generate a corresponding mind chain based on the fault through the LLM reference mind chain knowledge base, and then obtain the required data from the multi-source database based on the generated mind chain. Then, the fault location is performed by using the obtained data and mind chain. Therefore, even if there are new fault types, this solution can effectively deal with them, so as to effectively locate faults and improve the efficiency of fault location. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of a power distribution network fault accurate location system based on multi-source data fusion. Detailed Implementation

[0035] A precise fault location system for distribution networks based on multi-source data fusion includes:

[0036] Agent unit, large model LLM unit, mind chain knowledge base, multi-source database;

[0037] The Agent unit includes: a thinking chain module, an SQL retrieval module, and a fault location module; the thinking chain module is bidirectionally connected to the large model LLM unit and the thinking chain knowledge base; the SQL retrieval module is bidirectionally connected to the multi-source database and the large model LLM unit; the fault location module is bidirectionally connected to the large model LLM unit; the multi-source database is unidirectionally connected to the thinking chain module, with the thinking chain module acting as the data receiving end;

[0038] The thought chain module is used to receive multi-source data related to abnormal events in the distribution network, and based on the received multi-source data, retrieve relevant thought chain instances in the thought chain knowledge base, construct a prompt, send the prompt to the LLM, and request the LLM to return thought chain data; the thought chain data is used to locate faults in the distribution network based on the multi-source data related to abnormal events.

[0039] The SQL retrieval module is used to extract the text information of the steps that need to be retrieved from the thinking chain returned by LLM. That is, the text information indicates what data still needs to be obtained. This data is used for fault location. Based on the extracted text information corresponding to each thinking chain step, the module generates SQL instructions to query the corresponding data in the multi-source database and performs the query in the multi-source database, thereby returning the multi-source data corresponding to each thinking chain step.

[0040] The fault location module is used to receive the thought chain generated by LLM from the thought chain module and the multi-source data corresponding to each step of the thought chain from the SQL retrieval module, generate a prompt, and send it to LLM, thereby determining the fault location of the distribution network through LLM analysis.

[0041] A real-time fault location method for distribution networks based on multi-source data fusion, implemented on the aforementioned positioning system, includes the following process:

[0042] S100: Constructing a Mind Chain Knowledge Base; Each piece of data in the Mind Chain Knowledge Base consists of two parts:

[0043] Multi-source data related to distribution network faults can be encoded using encoders based on models such as BERT to obtain word vector features or sentence vector features, which can be used as index information for retrieval and matching.

[0044] The thought process for fault location in the distribution network based on reasoning using this multi-source data.

[0045] Example 1

[0046] Multi-source data related to distribution network faults: FTU overcurrent alarm and switch trip signal of a 10kV line, SOE recording time is 14:05:32;

[0047] Mind chain:

[0048] Based on the overcurrent alarm and trip information, it was determined that a short circuit fault may have occurred on the line.

[0049] Considering that lightning strikes are the main cause of thunderstorms, it is necessary to determine the meteorological data of the relevant areas of the route and first obtain the location information of the areas through which the route passes.

[0050] Obtain meteorological information corresponding to each location;

[0051] Identify areas where thunderstorms occur as potential locations for malfunctions.

[0052] Example 2

[0053] Multi-source data related to power distribution network faults: Received a fault repair work order from 95598, describing "power outage" and "wire sparks";

[0054] Mind chain:

[0055] Obtain the user's electricity meter data to verify the authenticity of the repair request;

[0056] Obtain repair request data from users located in the same residential area as the user;

[0057] If multiple users in the same community report a problem within the same time period, the location of the fault is determined to be in or near that community.

[0058] Obtain the electricity meter data for this community to confirm whether there are any sudden changes in electricity consumption;

[0059] If there is no sudden change in electricity consumption, then external environmental forces are the main cause, and information on drone inspections of the community should be obtained.

[0060] The location of abnormal information in the drone inspection data, such as the location of tree obstacles, is used as the suspected location of a malfunction.

[0061] The specific process of one of the data construction methods in the aforementioned mind chain knowledge base is as follows:

[0062] S110: Obtain historical records of fault location performed by maintenance personnel;

[0063] S120: Based on historical records, several thought chain data are artificially constructed. That is, it is possible to identify historical faults and, after a fault occurs, to determine the step-by-step thought process of the maintenance personnel from the occurrence of the fault to the completion of the maintenance, the results of fault location, and the maintenance results. Based on this information, thought chain data related to this fault can be generated. The thought process is given by the maintenance personnel based on the actual work process. The results of fault location include: locating the specific fault occurrence point, that is: what equipment / line in what location failed, and the cause of the fault. Thus, the data in the thought chain knowledge base is artificially constructed.

[0064] The specific process of the second data construction method in the aforementioned mind chain knowledge base is as follows:

[0065] The following process follows S120 above:

[0066] S130: Using several data points from the knowledge base built in S120 as examples, and given a known fault, combine the fault-related data and examples into a prompt and input it into the LLM;

[0067] S140: Let the LLM generate a thought chain data; you can directly refer to the methods in steps S200 to S400, so I will not repeat them here.

[0068] S150: The generated thought chain data is used as a reference for maintenance personnel to troubleshoot this fault. If the maintenance personnel find that the generated thought chain data is correct after troubleshooting and can locate the fault according to this thought chain data, then the relevant data and thought chain data of this fault will be added to the thought chain knowledge base.

[0069] S200: The agent acquires multi-source fault data related to distribution network faults; specifically, the agent's thought chain module acquires multi-source fault data related to the distribution network, such as:

[0070] "The overcurrent protection of line L1 of 10kV line tripped at 14:30:25, and the current of FTU at the end of line L1 is 15A."

[0071] The sources of this multi-source fault data can be: due to a fault occurring, this fault is directly associated with fault-related data. Fault-related data consists of basic information directly related to this fault contained in the multi-source database. While this basic information is directly accessible, maintenance personnel may not be able to pinpoint the exact problem using it alone. Therefore, subsequent steps identify the thought chain data used for fault localization to assist maintenance personnel in fault location, providing a final fault localization result and serving as a reference for subsequent maintenance work.

[0072] S300: Retrieved multiple thought chain information related to the fault in this distribution network from the thought chain knowledge base;

[0073] Specifically, the thought chain module can use, for example, a BERT-based encoder to encode the multi-source fault data to obtain corresponding word vectors or sentence vector features. Then, these vector features are used to perform matching in the thought chain knowledge base to obtain several thought chain information links with similar distribution network faults.

[0074] S400: Using multi-source fault data and the matched multiple thought chain information, a prompt is constructed. This prompt is used to enable LLM to generate a thought chain for fault localization, resulting in the final thought chain for fault localization.

[0075] For example, the prompt could look like this:

【

[0077] Mind Chain Example 1{}

[0078] Mind Chain Example 2{} ......

[0080] Please refer to the above example of the thought process and provide a thought process for locating the fault for the following fault information:

[0081] {The 10kV line L1 overcurrent protection tripped (time: 14:30:25, and the FTU current at the end of line L1 is 15A} 】

[0083] Then, the thought chain module inputs the above prompt into the LLM, thereby obtaining the corresponding thought chain information from the LLM:

[0084] A SCADA tripping event should cause a complete power outage on line L1. However, the presence of current at the end of line L1 indicates that the relay protection device has malfunctioned, or that there is an independent power source downstream of the fault point.

[0085] Therefore, it is necessary to retrieve the self-test log information of the relay protection device of line L1, as well as the power output information of the power source downstream of line L1.

[0086] If the relay protection device malfunctions during self-test, the fault can be located to that relay protection device.

[0087] If the relay protection device performs a normal self-test, then the possibility of an operational logic error can be ruled out.

[0088] If the downstream power supply of L1 is normal, the possibility of the downstream power supply transmitting power can be ruled out.

[0089] Otherwise, the fault is located in the downstream power supply.

[0090] S500: Extract text for database queries from the thought chain, and then use Text2SQL technology to obtain data from multiple source databases; that is, the text content indicates which data needs to be obtained from multiple source databases based on the thought chain information, and the text is used to generate SQL statements;

[0091] For example, the SQL retrieval module can extract the following text from the above thought chain information:

[0092] Retrieve the self-test log information of the relay protection device on line L1;

[0093] Retrieve the power output information of the power source downstream of line L1.

[0094] Specifically, for example, the SQL retrieval module can be used to input the thought chain information back into the large language model, and the prompt words can be set to "according to this thought chain {...} for fault location, what data mentioned in it should be obtained", so that the large language model can output the text content related to the data to be retrieved.

[0095] For example, named entity recognition can be used to extract text content related to the data to be retrieved. The Chinese names corresponding to each field can be extracted from technical documents in multi-source databases to obtain each entity. Then, the thought chain information can be segmented to obtain word vectors, and the entity vectors of each entity can be determined. The corresponding vectors can be generated by the encoder of BERT, and the text content related to the data to be retrieved can be extracted by similarity matching.

[0096] After obtaining the above text content, the SQL retrieval module can use Text2SQL technology to generate SQL query instructions corresponding to the above text through LLM and input them into the multi-source database, thereby retrieving the self-test log information of the relay protection device of line L1 and the power output information of the downstream power supply of line L1.

[0097] S600: Combines the thought chain and data obtained from multiple source databases into a prompt, enabling the LLM to locate faults.

[0098] For example, a prompt message could look like this:

[0099] Fault Information:

[0100] {

[0101] The overcurrent protection of line L1 on the 10kV line tripped at 14:30:25, and the FTU current at the end of line L1 was 15A.

[0102] }

[0103] Mind chain information: {

[0104] A SCADA tripping event should cause a complete power outage on line L1. However, the presence of current at the end of line L1 indicates a malfunction in the relay protection device or a separate power source downstream of the fault. Therefore, it is necessary to retrieve the self-test log information of the relay protection device on line L1, as well as the power output information of the power source downstream of line L1.

[0105] If the relay protection device fails a self-test, the fault can be located to that relay protection device; if the relay protection device passes a self-test, the possibility of an error in the operating logic can be ruled out.

[0106] If the downstream power supply of L1 is normal, the possibility of the downstream power supply transmitting power can be ruled out.

[0107] Otherwise, the fault is located between the downstream power source and the beginning of line L1.

[0108] }

[0109] Returned search data:

[0110] {

[0111] Relay protection device log;

[0112] Output data of distributed photovoltaic power downstream of line L1;

[0113] }

[0114] Based on the above fault information, thought process information, and search data, please locate the fault.

[0115] The content returned by LLM can be as follows:

[0116] "Since the protection device's self-test is normal and its operating logic is correct, the assumption that the relay protection device is abnormal is ruled out."

[0117] The L1-05 photovoltaic cluster, a branch of L1, suddenly increased from 0kW to 200kW at 14:30:20 (5 seconds before the trip)!

[0118] Therefore, through spatiotemporal correlation analysis, the fault point was determined to be in the section between the beginning of line L1 and the L1-05 branch.

[0119] In this invention, the multi-source data types are as follows:

[0120] (a) Equipment operation data

[0121] Data returned by the power distribution terminal (e.g., FTU / DTU / TTU):

[0122] Data types: three-phase voltage / current, waveform power, sudden zero-sequence current, switch trip signal (SOE).

[0123] Value: Directly captures the electrical characteristics of faults (such as phase loss ground current) and traces the faulty section by combining topological relationships.

[0124] Data related to the fault indicator:

[0125] Data types: Fault current direction, marked temperature abnormality, short circuit peak record.

[0126] Value: Quickly pinpoints faulty branches, especially useful for suburban lines without automation coverage.

[0127] Data related to online monitoring devices:

[0128] Data types: cable joint temperature, partial discharge intensity, insulation condition parameters.

[0129] Value: Anticipating equipment deterioration-related failures, such as overload causing connectors to melt.

[0130] (ii) User-side data

[0131] Smart meter AMI data:

[0132] Data type: User-side voltage dip / interruption event recovery timestamp.

[0133] Value: Accurately pinpointing the scope of a power outage, such as the loss of voltage in all meters under a certain transformer area indicating an upstream fault.

[0134] Repair order data:

[0135] Data type: Repair address, time, and description text.

[0136] (III) Environmental meteorological data

[0137] Data from the lightning location system:

[0138] Data types: lightning strike time, latitude and longitude, coordinates, lightning current intensity.

[0139] Value: To determine the faults caused by lightning.

[0140] Micro-weather station data:

[0141] Data types: wind speed and direction, rainfall, temperature.

[0142] Value: Predicting faults in environmentally sensitive areas.

[0143] Satellite / radar imagery:

[0144] Data types: Surface temperature heat map, vegetation cover density.

[0145] Value: Identifying line faults in wildfire risk areas.

[0146] (iv) Geospatial data

[0147] GIS topology database data:

[0148] Data types: line connection relationships, switch segment positions, tower coordinates.

[0149] Value: The core basis for fault tracing path calculation.

[0150] Image data from drones / inspection robots:

[0151] Data types: visible light / infrared images, laser point clouds.

[0152] Value: Visually confirming fault locations and assessing repair routes.

[0153] (v) System operation data

[0154] SCADA / EMS system data:

[0155] Data type: Network power flow distribution protection action sequence voltage over-limit record.

[0156] Value: Differentiate between short-circuit / ground fault types and verify consistency.

[0157] Historical fault database data:

[0158] Data type: Fault location, cause, processing time.

[0159] New energy power output data:

[0160] Data type: Real-time status of distributed photovoltaic / wind power inverters.

Claims

1. A precise fault location system for distribution networks based on multi-source data fusion, characterized in that, include: Agent unit, large model LLM unit, mind chain knowledge base, multi-source database; The Agent unit includes: a thinking chain module, an SQL retrieval module, and a fault location module; the thinking chain module is bidirectionally connected to the large model LLM unit and the thinking chain knowledge base; the SQL retrieval module is bidirectionally connected to the multi-source database and the large model LLM unit; the fault location module is bidirectionally connected to the large model LLM unit; the multi-source database is unidirectionally connected to the thinking chain module, with the thinking chain module acting as the data receiving end; The thought chain module is used to receive multi-source data related to abnormal events in the distribution network, and based on the received multi-source data, retrieve relevant thought chain instances in the thought chain knowledge base, construct a prompt, send the prompt to the LLM, and request the LLM to return thought chain data. The SQL retrieval module is used to extract the text information of the steps to be retrieved from the thought chain returned by LLM, generate SQL instructions for querying the corresponding data in the multi-source database based on the extracted text information corresponding to each thought chain step, and perform the query in the multi-source database to return the multi-source data corresponding to each thought chain step. The fault location module is used to receive the thought chain generated by LLM from the thought chain module and the multi-source data corresponding to each step of the thought chain from the SQL retrieval module, generate a prompt, and send it to LLM, thereby determining the fault location of the distribution network through LLM analysis.

2. A method for accurate fault location in distribution networks based on multi-source data fusion, characterized in that, The process includes the following: S100: Building a mind chain knowledge base; S200: The agent acquires multi-source fault data related to distribution network faults; S300: Retrieved multiple thought chain information related to the fault in this distribution network from the thought chain knowledge base; S400: Using multi-source fault data and the matched multiple thought chain information, a prompt is constructed. This prompt is used to enable LLM to generate a thought chain for fault localization, resulting in the final thought chain for fault localization. S500: Extracts text from the thought chain for database queries, and then uses Text2SQL technology to obtain data from multiple source databases; S600: Combines the thought chain and data obtained from multiple source databases into a prompt, enabling the LLM to locate faults.

3. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 2, characterized in that, Each piece of data in the S100 mind chain knowledge base includes two parts: Multi-source data related to distribution network faults can be encoded by a model-based encoder to obtain word vector features or sentence vector features, which can be used as index information for retrieval and matching. The thought process for fault location in the distribution network based on reasoning using this multi-source data.

4. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 3, characterized in that, The specific process of constructing the data in the aforementioned mind chain knowledge base is as follows: S110: Obtain historical records of fault location performed by maintenance personnel; S120: Based on historical records, identify historical faults, and determine the step-by-step thought process of maintenance personnel from the occurrence of the fault to the completion of maintenance, the results of fault location, and the maintenance results. Based on this information, thought chain data related to this fault can be generated. The thought process is provided by the maintenance personnel based on the actual work process. The results of fault location include: locating the specific fault location, that is: what equipment / line in what location has failed, and the cause of the fault.

5. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 4, characterized in that, The following process follows S120: S130: Using the data in the knowledge base built in S120 as examples, when a fault is known, combine the fault-related data and examples into a prompt and input it into the LLM; S140: Enable LLM to generate a thought chain data; S150: The generated thought chain data is used as a reference for maintenance personnel to troubleshoot this fault. If the maintenance personnel find that the generated thought chain data is correct after troubleshooting and can locate the fault according to this thought chain data, then the relevant data and thought chain data of this fault will be added to the thought chain knowledge base.

6. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 2, characterized in that, The multi-source data in S100 is encoded by an encoder based on the BERT model.

7. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 2, characterized in that, In the S300, the thinking chain module uses a BERT-based encoder to encode the multi-source fault data to obtain corresponding word vectors or sentence vector features; then, it uses these vector features to match in the thinking chain knowledge base to obtain several thinking chain information similar to the distribution network faults.

Images