CN-122019773-A - Power grid business index abnormal attribution method, device, electronic equipment and storage medium

CN122019773ACN 122019773 ACN122019773 ACN 122019773ACN-122019773-A

Abstract

The invention discloses a power grid business index abnormal attribution method, a device, electronic equipment and a storage medium, belonging to the field of power grid business data analysis, wherein the method comprises the steps of obtaining a power grid index abnormal attribution inquiry text, wherein the inquiry text comprises index names, abnormal directions and corresponding time and organization information; the method comprises the steps of carrying out intention analysis on a query text to obtain a standardized index object and determining a transaction type, calling matched index quantized data from a time sequence index database based on the standardized index object, locating corresponding index nodes in a power grid operation knowledge graph, carrying out backward tracing along a causal relationship to form an upstream causal relationship chain, and carrying out comprehensive analysis on the index quantized data and the causal relationship chain to generate an index transaction attribution analysis report. By implementing the method and the device, the problems that in the prior art, the grid index abnormal analysis only depends on numerical monitoring to cause data and business logic to be split and historical experience is difficult to combine to automatically position the abnormal root cause can be solved.

Inventors

ZHAO SHUANG
XIE HANYANG
FENG WEIXIA
JIANG JIANG
HUANG XIAOQIANG
PENG ZEWU
FENG XINYAO
LENG YUQI
Zang Haoqi
LIN JIAXIN
YANG YONGJIAO
PANG PENG
SU HUAQUAN
WEN YOU
PAN HUI
DENG CHURAN
LUO XUAN
Ma Guanxiong
Wan chan
YAN YUPING
SHAO YANNING

Assignees

广东电网有限责任公司

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (10)

1. The utility model provides a power grid business index abnormal attribution method which is characterized by comprising the following steps: Acquiring a power grid index abnormal attribution inquiry text, wherein the power grid index abnormal attribution inquiry text comprises index name keywords representing target power grid service indexes, abnormal direction keywords representing the target power grid service indexes in abnormal states, and time description information and organization description information aiming at the target power grid service indexes; inputting the query text of the power grid index abnormal attribution into an intention analysis model, so that the intention analysis model maps to obtain a standardized index object according to the index name keyword, and determines the abnormal type according to the abnormal direction keyword; retrieving index quantization data matched with the standardized index object, the time description information and the organization description information from a preset time sequence index database; map nodes corresponding to the standardized index objects are matched from a preset power grid operation knowledge map, and the map nodes are used as starting points to carry out reverse traversal along causal relation edges, so that an upstream causal relation chain pointing to potential trigger factors is generated; data aggregation is carried out on the index quantized data and the upstream causal relationship chain to form a transaction attribution evidence set; Invoking a reference case template matched with the standardized index object and the abnormal action type from a preset historical case library; and inputting the abnormal attribution evidence set and the reference case template into a preset attribution generation model to generate an index abnormal analysis report.
2. The grid business index anomaly attribution method of claim 1, wherein training the intent resolution model comprises: The method comprises the steps of acquiring a power grid intention recognition training data set, wherein the power grid intention recognition training data set comprises a plurality of power grid index query sample texts extracted from a history log, and each power grid index query sample text is marked with a real entity tag sequence of a keyword representing an index name and a real category tag representing a transaction type; dividing the power grid intention recognition training data set into a plurality of batches of training samples according to a preset batch size; inputting training samples of each batch into an intention analysis model to be trained for iterative training until the number of preset training rounds is met, wherein in each iteration, the intention analysis model to be trained outputs a predicted index entity label sequence and a predicted transaction type confidence coefficient according to the training samples of the current batch; Calculating an entity extraction loss value according to the predicted index entity tag sequence and the corresponding real index entity tag sequence; And carrying out weighted summation on the entity extraction loss value and the intention classification loss value to obtain a comprehensive loss value, and updating network parameters of an intention analysis model to be trained by using a preset optimizer according to the comprehensive loss value.
3. The method of power grid business index transaction attribution according to claim 2, wherein retrieving index quantization data matching the standardized index object, the time description information and the organization description information from a preset time sequence index database comprises: analyzing the time description information, and converting the time description information into a standard time window containing a start time stamp and an end time stamp; Matching the tissue description information with a preset tissue structure topology tree, and determining a unique tissue structure identification code corresponding to the tissue description information; Generating a time sequence database query statement according to the standardized index object, the standard time window and the unique organization identification code; And extracting time sequence values in the standard time window from a preset time sequence index database according to the time sequence database query statement to serve as the index quantized data.
4. The power grid business index transaction attribution method according to claim 3, wherein the matching of map nodes corresponding to the standardized index objects from a preset power grid operation knowledge map, and the reverse traversal is performed along a causal relation side with the map nodes as a starting point, and an upstream causal relation chain pointing to a potential trigger factor is generated, and the method comprises the following steps: Searching a map node consistent with the standardized index object name in a preset power grid operation knowledge map, and defining the map node as an initial node; constructing a node sequence only comprising the initial node, and defining the node sequence as an initial path; Constructing a path queue to be traversed, which only contains the initial path; constructing an accessed node set only comprising the initial node; repeatedly executing path reproduction operation until the path queue to be traversed is empty, and determining all paths in the result set of the attribution chains as upstream causality chains; The path propagation operation includes: extracting a path of the head of the queue from the path queue to be traversed as a current processing path, and taking a node positioned at the tail end of the current processing path as a current analysis node; retrieving an incoming degree causal relationship edge pointing to the current analysis node according to the current analysis node; Positioning a plurality of upstream adjacent nodes at the starting end of the incoming degree causal relationship edge according to the incoming degree causal relationship edge; Comparing the plurality of upstream adjacent nodes with nodes in the accessed node set, taking the upstream adjacent nodes which do not exist in the accessed node set as effective upstream nodes, and adding the effective upstream nodes into the accessed node set; Taking the effective upstream node containing the preset root attribute tag as a first node, and taking the effective upstream node not containing the preset root attribute tag as a second node; Splicing each first node to the current processing path to generate corresponding complete attribution chains, and storing each generated complete attribution chain into a preset attribution chain result set, wherein the initial attribution chain result set is empty; And respectively splicing each second node to the current processing path to generate a corresponding intermediate extension path, and sequentially adding each generated intermediate extension path to the tail of the path queue to be traversed so as to finish iterative updating of the path queue to be traversed.
5. The grid business index anomaly attribution method according to claim 4, wherein the data aggregating the index quantized data with the upstream causal relationship chain to form an anomaly attribution evidence set comprises: for each upstream causal link in the attribution link result set, resolving an entity node positioned at the tail end of the link from the current upstream causal link, and determining the entity node as a root cause node; Generating a database query instruction aiming at the root node according to the root node and the standard time window, extracting a target time sequence value from the time sequence index database according to the database query instruction, and determining the target time sequence value as root node quantized data; According to a preset time sequence similarity calculation function, comparing the numerical sequence of the index quantized data with that of the root cause node quantized data, and calculating to obtain a trend correlation coefficient; according to the current upstream causal relationship chain, the root cause node quantized data and the trend correlation coefficient, packaging and generating a standard format abnormal attribution evidence list; And adding the abnormal attribution evidence list into a preset abnormal attribution evidence set, wherein the initial abnormal attribution evidence set is empty.
6. The grid business index anomaly attribution method according to claim 5, wherein the retrieving a reference case template matching the standardized index object and the anomaly type from a preset historical case base comprises: Performing feature coding on the standardized index object and the abnormal type to generate a combined feature vector, and determining the combined feature vector as a target retrieval vector; Traversing a preset historical case library to obtain index feature vectors which are pre-constructed for each historical case; According to a preset vector similarity calculation function, similarity values between the target retrieval vector and each index feature vector are calculated respectively; sorting the historical cases according to the sequence of the similarity values from high to low to obtain a case sorting list; and intercepting the first historical case in the case sequencing list, and determining the first historical case as a reference case template.
7. The grid business index anomaly attribution method of claim 6, wherein training the attribution generation model comprises: Obtaining an attribution generation training data set, wherein the attribution generation training data set comprises a plurality of attribution generation samples extracted from a historical attribution file, and each attribution generation sample comprises a historical transaction attribution evidence set, a historical reference case template and a marked standard attribution analysis report; Dividing the attribution generated training data set into a plurality of batches of training samples according to a preset batch size; Inputting training samples of each batch into an attribution generating model to be trained for iterative training until the number of training rounds is met, wherein in each iteration, the attribution generating model to be trained takes the historical abnormal attribution evidence set and the historical reference case template contained in the training samples of the current batch as input data, and outputs a predictive attribution analysis text; calculating text to generate a loss value according to the predicted attribution analysis text and the corresponding standard attribution analysis report; And updating the network parameters of the attribution generating model to be trained according to the text generating loss value by using a preset optimizer.
8. The power grid business index abnormal attribution device is characterized by comprising an intention analysis module, a data retrieval module, a chain generation module, an evidence aggregation module and a report generation module; The system comprises an intention analysis module, a calculation module and a calculation module, wherein the intention analysis module is used for acquiring an intention analysis module, the intention analysis module comprises an index name keyword for representing a target power grid business index, an abnormal direction keyword for representing the target power grid business index in an abnormal state, time description information and organization description information for the target power grid business index, and the intention analysis module is used for inputting the intention analysis module into the intention analysis module so that the intention analysis module can map to obtain a standardized index object according to the index name keyword and determine the abnormal type according to the abnormal direction keyword; The data retrieving module is configured to retrieve, from a preset time sequence index database, index quantized data that matches the standardized index object, the time description information, and the tissue description information; The chain generation module is used for matching map nodes corresponding to the standardized index objects from a preset power grid operation knowledge map, and performing reverse traversal along a causal relation side by taking the map nodes as a starting point to generate an upstream causal relation chain pointing to potential trigger factors; the evidence aggregation module is used for carrying out data aggregation on the index quantized data and the upstream causal relationship chain to form a transaction attribution evidence set; The report generation module is used for calling a reference case template matched with the standardized index object and the abnormal movement type from a preset historical case library, inputting the abnormal movement attribution evidence set and the reference case template into a preset attribution generation model, and generating an index abnormal movement analysis report.
9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, the processor implementing the grid business index anomaly attribution method of any of claims 1 to 7 when the computer program is executed.
10. A storage medium comprising a stored computer program, wherein the computer program, when run, controls a device in which the storage medium is located to perform the grid business index profile attribution method according to any one of claims 1 to 7.

Description

Power grid business index abnormal attribution method, device, electronic equipment and storage medium Technical Field The invention relates to the technical field of power grid business data analysis, in particular to a power grid business index abnormal attribution method, a device, electronic equipment and a storage medium. Background With the comprehensive construction of the electric power Internet of things and the intelligent power grid, operation data of power grid enterprises show explosive growth, and key business indexes such as line loss rate, electricity sales quantity and the like are monitored in real time and attribution analysis is carried out, so that the method becomes a core link for guaranteeing economic operation of the power grid and improving management efficiency. The method can accurately diagnose the reasons of abnormal fluctuation of indexes, not only can help operation and maintenance personnel to rapidly check metering faults, electricity stealing behaviors or hidden equipment hazards, but also can provide important data support for lean management decisions of the power grid, and is a key place for improving the intelligent level of the power grid business. However, current power grid index anomaly analysis approaches rely mainly on statistical-based numerical monitoring or simple threshold alarms, with obvious limitations in facing complex anomaly attribution tasks. The main problem is that the prior art is often limited to the fluctuation display of the index quantized data, but lacks the deep association analysis of complex business logic and topological relation behind the index. The root cause of this problem is that a simple time series data analysis only reveals what happens (i.e. numerical anomaly), but cannot automatically correlate with a potential "upstream and downstream causal chain" (i.e. anomaly root cause) in the grid business, so that the "data layer" and the "logic layer" are mutually split. After the operation and maintenance personnel find that the index value is abnormal, the operation and maintenance personnel often need to manually query the topological relation across systems, deduce the upstream trigger factors by means of personal experience and manually write analysis reports, and the mode is low in efficiency and difficult to realize automatic attribution diagnosis by combining historical experience with multidimensional evidence. Disclosure of Invention The embodiment of the invention provides a power grid business index abnormal condition attribution method, a device, electronic equipment and a storage medium, which can solve the problems that in the prior art, data and business logic are split due to the fact that power grid index abnormal condition analysis only depends on numerical monitoring, and abnormal root causes are difficult to automatically position by combining historical experience. The embodiment of the invention provides a power grid business index abnormal attribution method, which comprises the following steps: Acquiring a power grid index abnormal attribution inquiry text, wherein the power grid index abnormal attribution inquiry text comprises index name keywords representing target power grid service indexes, abnormal direction keywords representing the target power grid service indexes in abnormal states, and time description information and organization description information aiming at the target power grid service indexes; inputting the query text of the power grid index abnormal attribution into an intention analysis model, so that the intention analysis model maps to obtain a standardized index object according to the index name keyword, and determines the abnormal type according to the abnormal direction keyword; retrieving index quantization data matched with the standardized index object, the time description information and the organization description information from a preset time sequence index database; map nodes corresponding to the standardized index objects are matched from a preset power grid operation knowledge map, and the map nodes are used as starting points to carry out reverse traversal along causal relation edges, so that an upstream causal relation chain pointing to potential trigger factors is generated; data aggregation is carried out on the index quantized data and the upstream causal relationship chain to form a transaction attribution evidence set; Invoking a reference case template matched with the standardized index object and the abnormal action type from a preset historical case library; and inputting the abnormal attribution evidence set and the reference case template into a preset attribution generation model to generate an index abnormal analysis report. Further, training the intent resolution model comprises: The method comprises the steps of acquiring a power grid intention recognition training data set, wherein the power grid intention recognition training data set comprises a plurality