The fifth article in the series on large models and their applications: Application cases of large models in the power industry

Written by

Audrey Miles

Updated on:June-25th-2025

The application cases of big models in C-end users are more familiar to everyone, but the application cases in B-end are relatively rare. On the one hand, due to the complexity of the business mechanism of B-end, the application of big models is still in the exploratory stage; on the other hand, there is relatively little publicity in this regard. This article introduces several typical cases of big models in the power industry . (

Intelligent generation and verification of dispatch operation tickets

A power grid dispatching operation ticket refers to a written instruction filled out by the dispatcher in a certain format and specification during the power grid dispatching operation, based on the changes in the power grid operation mode, equipment maintenance plan, and system operation needs, used to direct on-site operators to operate electrical equipment. A general operation ticket contains the operation task, operation sequence, equipment name and number, operation time, operator, guardian, and auditor, etc. The traditional way of generating a power grid dispatching operation ticket is generally to fill it out in the system based on the experience of the dispatcher, and then manually review and confirm the operation ticket. This method is inefficient, completely dependent on human experience, and prone to errors. Based on the combination of the general large model and the vertical large model, a knowledge graph of the dispatching system is constructed, and dispatching operation tickets are intelligently generated, which greatly improves work efficiency and reduces the error rate.

This solution relies on the dispatching cloud platform and integrates the power grid dispatching command system. The general large model and vertical large model are used to establish the knowledge graph of operation tickets in the field of power dispatching. A dispatching operation business graph is constructed, and the results are derived through the information association of the knowledge graph to realize the intelligent generation and verification of operation tickets, providing intelligent support for the rapid fault handling of control and operation personnel.

Figure 1 System architecture diagram

Construction of knowledge graph . Based on model measurement, construct the physical model graph of the power grid: with specific physical equipment as the entity and topological connection as the relationship, the graph is automatically generated from the power system CIM model and real-time measurement files, including: primary equipment of 35kV and above power grid, secondary equipment of 500kV and above. Based on the operating procedures, construct the semantic graph of operating procedures: construct with the operation object as the entity and the operation requirements as the relationship, parse the documents such as adjustment regulations and operation regulations, extract the core text content and related structures, understand the documents through pre-trained models supplemented by manual annotation, and construct semantic graphs.

Figure 2: Construction of knowledge graph

Fusion of knowledge graphs. To support unified basic knowledge services, we study the model fusion technology between graphs. Through relationship mapping and graph analysis, we merge the two graph knowledge bases into an operation ticket knowledge graph, build a dispatching operation business graph, and have the ability to integrate ticket models, error prevention rules, terminology knowledge, structure recognition, and equipment status perception.

The knowledge graph enables the intelligent generation and verification of operation tickets. The operation ticket knowledge graph model is connected to the business system. The business system provides basic data such as power grid model, maintenance order, and return order information. The operation ticket application calls the knowledge graph API to automatically generate standard steps for the operation ticket through knowledge association and reasoning, and can verify the compliance and safety of the operation execution process.

Construction results. According to the pilot, the DC equipment calibration accuracy of a certain branch is 95.7%, the AC equipment calibration accuracy has reached 97.6%, the operation ticket generation accuracy rate exceeds 99%, and the operation ticket preparation time has been reduced from an average of ten minutes to about 2 minutes.

Intelligent monitoring of illegal operations at production sites

Production operations in the power industry often involve high risks, and operating procedures must be strictly followed. Any carelessness can lead to major accidents. Therefore, intelligent monitoring of the production process is necessary. A power grid company combines the basic big model with the vertical big model of the power industry to build an artificial intelligence intelligent monitoring system with three core functions: suspected violation identification, key process identification, and intelligent operation data analysis.

Use visual recognition to monitor production operations on site . Through the visual recognition system, the work behaviors collected by the on-site video are analyzed to identify illegal operations. The whole process management of risk prevention before operation, on-site supervision during operation and tracing after operation is realized.

Figure 3 Intelligent monitoring operation scenario

Construction results . Since the system was launched, it has helped a municipal bureau in the Pearl River Delta to find violations such as people adjusting the guy wires on the pole tower (Class A) and working at heights without wearing safety belts (Class A); helped a municipal bureau in eastern Guangdong and a municipal bureau in western Guangdong to find violations such as not wearing insulating gloves during electrical operations (Class B); helped a municipal bureau in eastern Guangdong and a municipal bureau in northern Guangdong to find violations such as deliberately avoiding video surveillance.

Intelligent analysis of charging pile site selection and operation

Existing power grid charging piles often have pain points such as low charging pile utilization and difficulty for users to find charging piles. A power grid company uses the basic large model plus the vertical large model of the power industry to develop a charging pile precise site selection model to support scientific and reasonable charging pile planning and construction. On the one hand, it improves the utilization rate of charging piles, and on the other hand, it brings great convenience to car owners.

Construction plan .

Firstly, the operating data of existing charging piles are analyzed: a comprehensive analysis is conducted based on the results of site monitoring, site importance indicators, and charging demand forecast analysis, and finally the optimal site selection result of nonlinear fitting is obtained.

Figure 4 Charging pile data collection

Analysis of urban area types of charging facilities based on urban POI . Collect road network structure data of the cities to be sited, and analyze the impact of factors such as urban road network structure, including road traffic flow, road network density, road type, road grade, intersection density, etc. on urban network rasterization processing, so as to generate a directed graph that conforms to the actual site selection situation.

Evaluation and analysis of the importance of charging facilities in urban areas based on machine learning . First, data processing and feature extraction are performed based on the collected urban traffic, population, charging piles, etc. Second, the importance of the divided urban areas is quantified and ranked based on improved web page evaluation algorithms such as PageRank. Third, combined with machine learning algorithms such as support vector machines and random forests, urban areas are classified or clustered, and the distribution of charging facilities in different urban areas is further analyzed to provide a more accurate basis for the site selection of charging facilities.

Analysis of charging spatiotemporal demand forecasting technology based on historical charging data drive and time series model . Establish SOC analysis based on user charging data, including charging start time, charging end time, charging duration, charging amount, vehicle model, charging station location, etc., to more realistically reflect the characteristics of electric vehicle user behavior, thereby assisting in site selection.

Construction results : Through real-time monitoring of the operation of public charging facilities and optimized layout, the number of "zombie charging piles" and inefficient charging piles in a certain city was reduced by nearly 20%.

The case materials are derived from the "White Paper on Excellent Case Studies of Artificial Intelligence Application Scenarios in Central Enterprises (2024 Edition)" (Central Enterprise Artificial Intelligence Collaborative Innovation Platform China Southern Power Grid Co., Ltd., December 2024), and have been compiled by the author.