CN-121998682-A - LLM-based supply and demand factor anomaly identification tracing method, device and storage medium

CN121998682ACN 121998682 ACN121998682 ACN 121998682ACN-121998682-A

Abstract

The invention relates to a supply and demand factor anomaly identification tracing method, a device and a storage medium based on LLM, which are applied to the technical field of electric power operation monitoring and comprise the steps of realizing deep semantic analysis and dynamic association of unstructured data through a field fine adjustment model, constructing an unstructured text-structured factor association knowledge base, thoroughly solving the problems of information island and semantic gap, providing panoramic and high-quality characteristic input for anomaly analysis, embedding an innovative mechanism into a detection framework of data-driven fusion, remarkably improving the detection reliability under the complex scene by utilizing the complex mode in the data and being limited by physical constraint and market mechanism of an electric power system, avoiding false alarm and false alarm, constructing a LLM-driven three-layer progressive tracing mechanism, automatically generating a clear and credible causal chain map without depending on expert manual analysis, and greatly improving the interpretability and consistency of a tracing conclusion.

Inventors

DAI CHANGCHUN
LI YONGBO
Qian Hanhan
QI HUI
ZHANG WEISHI
JI CHAO
HAO YUXING
XIE DAOQING
FU JINGYU
ZHOU TAO
JIANG HAILONG
ZHAO XUETING
ZHANG WEI
Cheng Honggu
LIN ZHEMIN

Assignees

安徽电力交易中心有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (9)

1. LLM-based supply and demand factor anomaly identification traceability method is characterized by comprising the following steps: Obtaining structured data of supply and demand factors of multiple channels and unstructured data of supply and demand factors influenced by multiple dimensions; Carrying out semantic analysis on the unstructured data by adopting a pre-trained model, extracting structured association data and constructing an association knowledge base from unstructured text to structured factors; carrying out data standardization processing on the supply and demand factor structured data and the structured association data to generate a standardized supply and demand factor data set; an anomaly detection algorithm based on a decision tree embedded into the physical constraint of the power system is adopted, a scene self-adaptive dynamic prediction interval generated by a time sequence prediction model is combined, an anomaly confidence score of each factor in the standardized supply and demand factor data set is calculated, and an anomaly state and a normal state are distinguished according to the anomaly confidence score of each factor; The method comprises the steps of measuring nonlinear association strength among factors by a mutual information method, judging causal direction on time sequence by combining with a Grangel causal test, and constructing a factor coupling network with direction; based on an professional knowledge base in the electric power field, an inference template is constructed through prompt word engineering, a standardized supply and demand factor data set and an exception report are fused, three layers of progressive tracing logic are adopted to complete cause analysis, and an interpretable exception tracing result is output after verification by an expert rule base.
2. The method of claim 1, wherein the step of determining the position of the substrate comprises, The data normalization process includes: Filling the missing values in the supply and demand factor structured data and the structured association data by adopting an interpolation method; removing abnormal values in the supply and demand factor structured data and the structured association data through a3 sigma principle; Unifying the dimensions of different dimension factor data in the supply and demand factor structured data and the structured association data based on Z-score standardization; and verifying the logical consistency of the text information and the numerical data through the pre-trained LLM model, and triggering the manual review after the inconsistent data are marked.
3. The method of claim 2, wherein the step of determining the position of the substrate comprises, The semantic parsing of the unstructured data by using a pre-trained model comprises the following steps: preprocessing unstructured data by adopting a Qwen-Max large model subjected to fine adjustment of data in the electric power field, wherein the preprocessing comprises word segmentation, dead word filtering and field term normalization processing; And extracting structural association data comprising areas, time, factor types, numerical values and influencing objects from the preprocessed text by adopting an electric power field information extraction template.
4. A method according to claim 3, further comprising: and storing the standardized supply and demand factor data set, the abnormal report and the abnormal tracing result into a historical case library for model parameter iteration and associated knowledge base updating.
5. The method of claim 4, wherein the step of determining the position of the first electrode is performed, The anomaly detection algorithm based on the decision tree is an improved isolated forest algorithm, and the time sequence prediction model is an LSTM model; The physical constraint of the power system comprises a unit climbing rate limit, a new energy output non-abrupt constraint and a power balance boundary; The abnormal confidence score is 0-100, the score is greater than or equal to the first confidence threshold value and is high confidence abnormality, the low confidence abnormality is arranged between the first confidence threshold value and the second confidence threshold value, and the score is smaller than the second confidence threshold value and is in a normal state.
6. The method of claim 5, wherein the step of determining the position of the probe is performed, The fusing of the standardized supply and demand factor data sets and the anomaly report includes: And carrying out semantic alignment and context fusion on the standardized supply and demand factor data set and the exception report through a pre-trained LLM model to obtain standard reasoning input.
7. The method of claim 6, wherein the step of providing the first layer comprises, The method for completing the cause analysis by adopting the three-layer progressive tracing logic comprises the following steps: locating abnormal direct trigger events based on real-time factor changes and event logs; The causal reasoning capacity of the pre-trained LLM model is utilized to mine the dynamic coupling relation among multiple factors, and an abnormal propagation path is obtained; And searching the historical cases with similar semantics in the professional knowledge base in the electric power field, comparing the evolution rules and generating a cause chain map.
8. LLM-based supply and demand factor anomaly identification traceability device is characterized in that the device comprises: The multi-source data acquisition module is used for acquiring multi-channel supply and demand factor structured data and unstructured data of which the multi-dimension influences the supply and demand factors; The unstructured data processing module is used for carrying out semantic analysis on the unstructured data by adopting a pre-trained model, extracting structured associated data and constructing an associated knowledge base from unstructured text to structured factors; The data standardization module is used for carrying out data standardization processing on the supply and demand factor structured data and the structured association data to generate a standardized supply and demand factor data set; The abnormality detection module is used for calculating an abnormality confidence score of each factor in the standardized supply and demand factor data set by adopting an abnormality detection algorithm based on a decision tree embedded into the physical constraint of the power system and combining a scene self-adaptive dynamic prediction interval generated by a time sequence prediction model, and distinguishing an abnormality state from a normal state according to the abnormality confidence score of each factor; the system comprises an anomaly report generation module, a factor coupling network with direction, an anomaly report generation module, a priority ranking module and a priority ranking module, wherein the anomaly report generation module is used for measuring nonlinear association strength among factors through a mutual information method and judging a causal direction on a time sequence by combining with a Grangel causal test; The tracing module is used for constructing an inference template through prompt word engineering based on the professional knowledge base in the electric power field, fusing a standardized supply and demand factor data set and an exception report, completing cause analysis by adopting three layers of progressive tracing logic, and outputting an interpretable exception tracing result after verification by an expert rule base.
9. A storage medium storing a computer program which, when executed by a master, implements the steps of the LLM-based supply and demand factor anomaly identification tracing method of any one of claims 1-7.

Description

LLM-based supply and demand factor anomaly identification tracing method, device and storage medium Technical Field The invention relates to the technical field of electric power operation monitoring, in particular to a supply and demand factor anomaly identification tracing method, a supply and demand factor anomaly identification tracing device and a storage medium based on LLM. Background With the deep promotion of the improvement of the electric power market and the continuous rising of the new energy installation ratio, the running environment of an electric power system is increasingly complex, and the frequency and the amplitude of the fluctuation of the electricity price are obviously increased. The power spot market electricity discharge price is interactively influenced by multiple factors such as primary energy price fluctuation, new energy output uncertainty, meteorological condition change, power transmission and transformation equipment maintenance arrangement, policy adjustment, market main game policy and the like, abnormal electricity price can interfere reasonable allocation of market resources, the enthusiasm of market main participation in transactions is reduced, the interests of power generation enterprises can be possibly damaged, the follow-up investment and production enthusiasm are influenced, and the market running risk is increased. Anomaly monitoring and root cause analysis of power supply and demand factors (namely, thermal power bidding space, the value of which is the difference between the total power demand and the total non-thermal power output) are the core of power market risk management. The method has four key defects in the field, namely, firstly, multi-source heterogeneous data are fused on the surface, structured data (such as SCADA (supervisory control and data acquisition) and unstructured data (such as weather early warning text and maintenance information) are subjected to simple rule mapping or feature stitching, deep semantic understanding and dynamic association cannot be realized, so that 'data are rich and information is poor', secondly, prediction and anomaly detection model mechanisms are separated from data driving, a pure data driving model ignores physical constraints of an electric power system, a pure rule model is stiff, scene generalization capability is weak, false alarm rate is high, thirdly, anomaly cause tracing automation degree is low, expert manual analysis is relied, and stays at a correlation analysis level, interpretable causal reasoning capability is lacked, a clear causal chain cannot be generated, and fourthly, system solidification and rigidification do not have the capability of continuously evolving from operation feedback, and cannot adapt to dynamic evolution of an electric power market structure and rules. The prior art cannot meet the requirement of complex risk analysis in the electric power market, so that a technical scheme with depth data fusion, accurate anomaly identification, intelligent causal tracing and continuous evolution capability is needed. Disclosure of Invention In view of the above, the invention aims to provide a supply and demand factor anomaly identification tracing method, a supply and demand factor anomaly identification tracing device and a storage medium based on LLM, so as to solve the problems that in the prior art, multisource heterogeneous data are fused on the surface, a prediction and anomaly detection model mechanism is disjointed with data driving, the anomaly cause tracing automation degree is low, and expert manual analysis is relied on. According to a first aspect of an embodiment of the present invention, there is provided a supply and demand factor anomaly identification tracing method based on LLM, the method comprising: Obtaining structured data of supply and demand factors of multiple channels and unstructured data of supply and demand factors influenced by multiple dimensions; Carrying out semantic analysis on the unstructured data by adopting a pre-trained model, extracting structured association data and constructing an association knowledge base from unstructured text to structured factors; carrying out data standardization processing on the supply and demand factor structured data and the structured association data to generate a standardized supply and demand factor data set; an anomaly detection algorithm based on a decision tree embedded into the physical constraint of the power system is adopted, a scene self-adaptive dynamic prediction interval generated by a time sequence prediction model is combined, an anomaly confidence score of each factor in the standardized supply and demand factor data set is calculated, and an anomaly state and a normal state are distinguished according to the anomaly confidence score of each factor; The method comprises the steps of measuring nonlinear association strength among factors by a mutual information method, judging causal direction on time s