Vertical domain distribution network planning method and system based on large language model

Abstract

The invention relates to the technical field of power grids, in particular to a vertical domain distribution network planning method and system based on a large language model. The method comprises the following steps: collecting and arranging a distribution network structure and operation data, and constructing a'graph 'management system; extracting historical planning schemes to form a typical case library; fusing external open source data and source side data to establish a big database and an access interface; constructing a bilateral prediction model and a planning scheme generation model based on machine learning; carrying out secondary pre-training and practical packaging on the open source large language model; planning parameters are obtained in a guiding mode in a question-answer mode, and a scheme is generated. According to the method, the problems of high labor cost, long planning time, large error, unstable planning quality, poor interactivity, high learning threshold and the like of traditional distribution network planning are solved, the planning efficiency and precision are improved through an intelligent technology, the professional threshold is reduced, a brand new solution is provided for coping with the new situation of high permeability of new energy and electric energy replacement acceleration, and the method is worthy of popularization and application. And the practical value is remarkable.

Description

Technical Field

[0001] The present invention relates to the field of power grid technology, and in particular to a vertical field distribution network planning method and system based on a large language model. Background Art

[0002] As my country's energy transition deepens, the power system, particularly its distribution network, faces unprecedented challenges. By the end of 2024, installed capacity of new energy sources will continue to grow, and the number of electric vehicles will rapidly increase. This dual uncertainty on both the energy and load sides poses new challenges to traditional distribution network planning.

[0003] Traditional distribution network planning relies primarily on expert experience and simple tools. This approach has proven inadequate in addressing emerging trends such as the high penetration of renewable energy and the acceleration of electricity substitution. Conventional distribution network planning, typically based on deterministic load forecasts and fixed power output models, struggles to address bidirectional uncertainties on both the source and load sides, resulting in insufficient adaptability in planning solutions.

[0004] Currently, distribution network planning typically relies on manual consultation, with expert teams conducting analysis and making decisions using tools like Excel. This approach has significant drawbacks: first, high labor costs and lengthy planning times; second, the limited number of parameters relied upon for analysis can easily lead to significant errors; third, the varying expertise of different experts leads to inconsistent planning quality; and finally, traditional planning tools lack user interactivity and have a high learning curve, hindering widespread adoption.

[0005] At the same time, industrial enterprises, as a major source of carbon emissions, are accelerating their transition to electric energy substitution, further increasing the complexity of distribution network planning. The rapid changes in industrial energy consumption structure require higher precision and flexibility in distribution network planning.

[0006] Therefore, in order to adapt to the new situation of energy transformation, distribution network planning urgently needs to introduce new technologies, especially digital means such as artificial intelligence and big data, to achieve smarter, more accurate and more cost-effective planning schemes, and effectively support the process of new energy consumption and electricity substitution. Summary of the Invention

[0007] In view of the problems existing in the prior art, the present invention is proposed.

[0008] Therefore, the problem to be solved by this invention is how to address the common practice of manual consultation for distribution network planning, where teams of experts conduct analysis and make decisions using tools such as Excel. This approach has significant drawbacks: first, high labor costs and long planning times; second, the limited number of parameters relied upon for analysis, which can easily lead to large errors; third, the varying expertise of different experts leads to unstable planning quality; and finally, traditional planning tools have poor user interactivity and a high learning curve, making them difficult to widely promote and apply.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] In a first aspect, embodiments of the present invention provide a vertical distribution network planning method based on a large language model. The method includes collecting and organizing distribution network structure and operation data, as well as historical distribution network planning scheme data, structuring them, and constructing indexes based on scenario, time, region, scale, feeder, and substation, forming a "one-picture" management system for the distribution network.

[0011] Extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library, and strengthen cases of electricity substitution and new energy consumption;

[0012] Integrate external open source data and source-side data, further structure them, form a large database for distribution network planning, and build a large data access interface for distribution network planning to support large language models to access data;

[0013] Based on machine learning, a dual-side prediction model for the source and load sides of the distribution network and a distribution network planning scheme generation model are constructed, forming an access interface for the distribution network planning machine learning model;

[0014] Based on the open source large language model, secondary pre-training is performed to generate a large language model for a vertical domain. The large language model is then practically packaged to enhance user interaction capabilities, access capabilities for the distribution network planning large database, the ability to call the distribution network planning solution generation model, and the ability to generate the final solution.

[0015] Through the form of questions and answers, the guided acquisition of planning scenarios, goals and technical indicators is completed to obtain the target microgrid planning scheme.

[0016] As a preferred solution of the vertical field distribution network planning method based on the large language model of the present invention, the collection and collation of distribution network structure and operation data, and the historical distribution network planning scheme data include:

[0017] Pay attention to and organize new energy power generation data, new energy vehicle data, energy storage system data, and virtual power plants to adapt to and respond to the two-sided uncertainties of the current distribution network system.

[0018] As a preferred solution of the vertical field distribution network planning method based on the large language model of the present invention, wherein: the extraction of planning scenarios, goals and technical indicators from historical planning schemes to form a typical distribution network planning case library includes:

[0019] Extract the scenarios, goals, and technical parameters of distribution network planning, analyze, and generate standardized data items required for distribution network planning under different scenarios;

[0020] Abstract the normalized data items as "input" and integrate the corresponding actual data of the current operation of the distribution network as "output";

[0021] Strengthen the organization of thematic scenarios such as new energy consumption, industrial power substitution, two-way interaction of electric vehicles, and virtual power plants, and construct additional parameter items to describe the scenarios more accurately.

[0022] As a preferred solution of the vertical field distribution network planning method based on the large language model of the present invention, the fusion of external open source data and source side data includes:

[0023] Supplement open source data such as meteorological data, holiday data, geographic information data, city and regional characteristics data, and economic data;

[0024] Integrate power generation data, especially data from renewable energy sites such as photovoltaic and wind power, and perform structured processing;

[0025] The open source data is associated and integrated with the distribution network planning case data to form a complete distribution network planning database.

[0026] As a preferred solution of the vertical field distribution network planning method based on the large language model of the present invention, the construction of a two-sided prediction model on the source side and the load side of the distribution network based on machine learning includes:

[0027] By using factors such as meteorological forecast data, historical power output data on the source side, and geographical data, a basic model for power output prediction on the source side is constructed;

[0028] By analyzing the load side characteristics, a basic load forecasting model is constructed;

[0029] According to the topology of the target distribution network, the flow constraints are analyzed and the two basic models are reasonably combined to form a two-sided prediction capability.

[0030] As a preferred solution of the vertical field distribution network planning method based on the large language model of the present invention, the distribution network planning solution generation model includes:

[0031] Using the output of the two-sided prediction model and the distribution network big data as input, the planning scheme of the target distribution network is output;

[0032] The planning scheme includes important parameters such as maximum load, load density, load growth rate, network topology, line cross-sectional flow, investment economic analysis, and reliability analysis.

[0033] As a preferred solution of the vertical field distribution network planning method based on the large language model of the present invention, the secondary pre-training includes:

[0034] Organize the data, logic, and specialized vocabulary used in the system construction process into a training library for a large language model;

[0035] Through incremental pre-training, instruction supervision fine-tuning, and reward model training, the model is restricted to the vertical field of service distribution network planning, and the context is pre-processed.

[0036] Secondly, embodiments of the present invention provide a vertical distribution network planning system based on a large language model. The system includes a data collection and organization module for collecting and organizing distribution network structure and operation data, as well as historical distribution network planning scheme data, structuring them, and constructing an index to form a "one-picture" management system for the distribution network.

[0037] A case library construction module is used to extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library;

[0038] The big data fusion module is used to integrate external open source data and source-side data, and further structure them to form a distribution network planning big data database. It also builds a distribution network planning big data access interface to support large language models to access data.

[0039] The model access module is used to build a two-sided prediction model for the source and load sides of the distribution network and a distribution network planning solution generation model based on machine learning, forming an access interface for the distribution network planning machine learning model;

[0040] The large language model pre-training module is used to perform secondary pre-training based on the open source large language model to generate a large language model for a specific vertical domain. The module also provides practical packaging for the large language model, improving user interaction capabilities, access to the large database for distribution network planning, the ability to call the distribution network planning solution generation model, and the ability to generate the final solution.

[0041] The user interaction module is used to complete the guided acquisition of planning scenarios, goals and technical indicators through question-and-answer format to obtain the target microgrid planning scheme.

[0042] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the vertical field distribution network planning method based on a large language model as described in the first aspect of the present invention are implemented.

[0043] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored, wherein: when the computer program instructions are executed by a processor, the steps of the vertical field distribution network planning method based on a large language model as described in the first aspect of the present invention are implemented.

[0044] The beneficial effects of the present invention are as follows: the present invention solves the problems of data dispersion and experience dependence in traditional distribution network planning by constructing a "one-picture" management system and a typical case library for distribution networks, thereby improving data access efficiency and planning standardization. The large database established by integrating external open source data and source-side data expands the information dimension of distribution network planning and overcomes the limitation of large errors caused by insufficient parameters of traditional methods. The two-sided prediction model based on machine learning achieves accurate prediction of new energy output and load changes, and improves the robustness of planning. The practical encapsulation of large language models in vertical fields breaks through the limitations of traditional tools with single interaction and high learning thresholds, and realizes guided acquisition of planning parameters through natural language question-and-answer interaction, which reduces the complexity of user operations, partially replaces manual labor, reduces planning costs, and improves planning quality, providing strong support for the new situation of high penetration of new energy and accelerated substitution of electric energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A flowchart of a vertical distribution network planning method based on a large language model;

[0047] Figure 2 Computer equipment diagram for vertical field distribution network planning method based on large language model;

[0048] Figure 3 A flowchart is constructed for the system of vertical field distribution network planning method based on large language model. DETAILED DESCRIPTION

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it constitute a separate or selective embodiment that is mutually exclusive with other embodiments.

[0052] Example 1

[0053] Reference Figures 1 and 2 , which is the first embodiment of the present invention, provides a vertical field distribution network planning method based on a large language model, including:

[0054] S100: Collects and organizes distribution network structure and operation data, historical distribution network planning scheme data, structures them, and builds indexes by scenario, time, region, scale, feeder, and substation to form a "one-picture" management system for the distribution network;

[0055] In this embodiment, collecting and organizing distribution network structure and operational data is a fundamental step in system construction. First, for distribution network structure data, static information such as grid topology, distribution transformer capacity, line type and cross-section, and protection device type and parameters must be collected. Operational data primarily includes dynamic information such as voltage at each distribution network node, power flow distribution, load curves, and fault records. Historical distribution network planning data includes past planning documents, implementation plans, and acceptance reports.

[0056] In specific implementation, data collection methods can be divided into three approaches: first, direct acquisition through the distribution network automation system interface, such as the SCADA system, distribution management system (DMS), distribution network automation master station system, etc.; second, indirect acquisition through the enterprise asset management system (EAM), engineering management system, etc.; third, obtaining data that is difficult to collect automatically through manual methods such as on-site surveys and expert interviews.

[0057] Data structuring is a key step in ensuring data quality. For unstructured document data, such as PDF planning reports, the system uses OCR technology combined with natural language processing to extract information. For semi-structured data, such as Excel spreadsheets, the system automatically identifies the data format and extracts information using pre-set parsing rules. For structured data, such as records in a database, the system uses ETL tools for data cleansing and conversion.

[0058] In an optional embodiment, the system utilizes a knowledge graph-based data integration solution, treating each element of the distribution network as a node in the graph and the relationships between elements as edges. For example, substation A is connected to distribution substation C via 10kV line B, which can be represented as a relationship of three nodes and two edges. This representation intuitively reflects the topology of the distribution network and the relationships between its elements, facilitating subsequent planning and analysis.

[0059] Index construction is a crucial step in achieving efficient data access. This system has built a multi-dimensional indexing system, including: scenario indexing (for urban core areas, industrial parks, rural power grids, etc.), time indexing (data organized by year, quarter, and month), regional indexing (province, city, county, distribution station service area, and other levels), scale indexing (classified by substation capacity, power supply radius, etc.), feeder indexing (classified by voltage level, line length, load density, etc.), and substation indexing (classified by capacity, number of main transformers, number of outgoing line loops, etc.). The indexes of each dimension are interconnected, forming a complete index network.

[0060] Specifically, for new energy access scenarios, the system includes additional indexes for power generation type (photovoltaic, wind power, biomass, etc.), installed capacity, and grid connection points. For electric vehicle charging infrastructure, the system also includes indexes for charging station location, number of charging piles, and charging power. These specialized indexes provide targeted support for distribution network planning in specific scenarios.

[0061] It should be noted that the distribution network "one map" management system integrates the aforementioned data and indexes. The system uses a GIS (Geographic Information System) as a base map, overlaying information layers such as distribution network topology, load distribution, and power distribution to enable visual management of distribution network assets and operational data. Users can flexibly view distribution network information for the desired area and content through multi-scale zooming, layer switching, and conditional filtering. The system also supports historical review and trend forecasting over time, enabling users to fully understand the distribution network's history, current status, and future development plans.

[0062] At the system implementation level, the distribution network "One Map" utilizes a distributed architecture design, with data storage utilizing a sharded cluster model to ensure the efficient storage and access of massive amounts of data. The system employs an incremental data update mechanism, automatically pulling the latest data from various source systems on a regular basis. Data consistency and accuracy are maintained through data comparison and verification. For frequently changing operational data, the system employs a multi-level caching strategy to balance data real-time performance with system performance.

[0063] In terms of data security, the system implements strict permission management and data encryption mechanisms. Data access permissions are differentiated based on user roles, sensitive data is encrypted for storage, and all data operations are logged to ensure traceability. Furthermore, the system regularly conducts data backup and recovery drills and establishes a comprehensive disaster recovery mechanism to ensure data security and reliability.

[0064] For example, this example collects five years of historical data for the distribution network planning of an industrial park. This includes detailed information on 10 110kV substations, 35 35kV distribution stations, 86 10kV feeders, and 245 distribution substations, totaling approximately 30GB of structured data and 15GB of unstructured document data. By building a "single-image" management system, the system integrates data scattered across multiple systems, enabling planners to intuitively view the development history and current status of the park's distribution network, providing comprehensive data support for subsequent planning.

[0065] S101: Collect and organize distribution network structure and operation data. Historical distribution network planning data includes:

[0066] Pay attention to and organize new energy power generation data, new energy vehicle data, energy storage system data, and virtual power plants to adapt to and respond to the two-sided uncertainties of the current distribution network system.

[0067] S200: Extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library, and strengthen cases for electricity substitution and new energy consumption;

[0068] In this example, extracting key information from historical planning proposals is fundamental to building a representative case library. The system uses natural language processing technology to perform semantic analysis and information extraction on planning documents, primarily extracting three core pieces of information: planning scenarios, planning objectives, and technical indicators.

[0069] Planning scenario information extraction primarily focuses on the planning context, applicable area type, and load characteristics. Examples include industrial park expansion scenarios, urban residential area renovation scenarios, and centralized new energy grid connection scenarios. The system uses methods such as keyword matching and topic modeling to identify paragraphs describing scenarios in documents and extract scenario feature parameters. For complex scenario descriptions, the system uses a BERT-based sequence annotation model to identify scenario elements in the text and construct a scenario knowledge graph.

[0070] Planning objective information extraction primarily focuses on the expected effects and optimization directions of the plan. Examples include improving power supply reliability, meeting load growth demands, and enhancing renewable energy absorption capacity. The system uses target vocabulary matching and semantic similarity analysis to identify target descriptions in documents. It then uses dependency parsing to extract the target's subject, attributes, and metrics. For implicitly expressed targets, the system uses semantic reasoning techniques to identify them.

[0071] Technical indicator information extraction primarily focuses on specific technical parameters and constraints. Examples include voltage qualification rate, line load rate, and renewable energy penetration rate. The system uses regular expressions and numerical unit recognition to extract parameter values ​​and their units from documents. For technical indicators presented in tabular form, the system uses a table structure recognition algorithm to extract the table content and convert it into structured data. For indicators presented graphically, the system uses image recognition technology to extract numerical information.

[0072] In an optional embodiment, the system utilizes a knowledge-enhanced deep learning model that combines a distribution network domain knowledge base with a pretrained language model to improve information extraction accuracy. This model, building on a standard pretrained model, incorporates a specialized vocabulary and rule base for distribution network applications. Through transfer learning, it is fine-tuned to the specific characteristics of distribution network planning documents. Experiments have shown that compared to general information extraction models, this knowledge-enhanced model improves the accuracy of distribution network terminology and technical specifications by 15%.

[0073] Data structuring and standardization are crucial components of building a case library. Based on the extracted information, the system designs a unified data model to transform unstructured raw information into structured case data. The data model primarily comprises a scenario description module, a goal definition module, a technical indicator module, an implementation plan module, and an effectiveness evaluation module. Relationships exist between these modules, such as the selection of goals for a specific scenario and the technical indicators corresponding to their achievement.

[0074] Specifically, the system has established a data dictionary and standardization system for distribution network planning, unifying terminology and metrics. For example, the "voltage qualification rate" metric is clearly defined as "the percentage of time the voltage is within ±7% of the rated value," with specific calculation cycles and measurement point selection methods. This standardization ensures that cases from different sources can be compared and analyzed within a unified framework.

[0075] Based on standardized data, the system has built a multi-dimensional case search mechanism. Users can perform combined searches based on scenario categories, technical indicator ranges, planning objectives, and other criteria to quickly locate similar historical cases. Furthermore, based on a content similarity algorithm, the system proactively recommends historical cases relevant to the current planning task for reference.

[0076] The system has taken special strengthening measures for the two key scenarios of electric energy substitution and new energy consumption. First, the collection scope of relevant cases has been expanded to include not only the planning cases of the unit, but also typical cases in the industry and international advanced experience. Secondly, a more in-depth parameter analysis of such cases has been conducted, and more detailed technical indicators and implementation measures have been extracted. For example, for electric energy substitution cases, additional parameters such as energy structure, substitution rate, and carbon emission reduction before and after substitution are recorded; for new energy consumption cases, additional information such as new energy output characteristics, fluctuation patterns, and control strategies are recorded. In addition, the system also specially marks and classifies such cases to facilitate users' targeted retrieval and reference.

[0077] Regarding case evaluation, the system has established a case quality assessment mechanism, scoring cases based on data integrity, technological advancement, and implementation effectiveness. The scoring results serve as a weighting factor in case recommendations, ensuring that high-quality cases are prioritized. The system also supports user commentary and tagging of cases, forming a collaborative case library maintenance mechanism.

[0078] It should be noted that the case library is not a static resource but a dynamically updated knowledge base. The system incorporates a regular update mechanism that automatically extracts information from completed planning projects to supplement the case library. Furthermore, as new technologies and standards change in the industry, the system proactively adjusts the standardized processing of case parameters to ensure the case library remains current. Furthermore, the system implements version management for the case library, recording the change history of each version and supporting historical version query and comparative analysis.

[0079] In terms of knowledge accumulation, the system not only stores basic information about planning cases but also extracts planning experience and best practices. For example, by analyzing the selection criteria and effectiveness evaluation of planning solutions in different scenarios, it summarizes planning patterns and considerations for specific scenarios. This knowledge is stored in the system as rules, templates, or case annotations, providing empirical reference for new planning tasks.

[0080] For example, this example examines 300 distribution network planning projects completed by a provincial power grid company over the past five years. Over 1,500 planning scenarios, 3,000 planning objectives, and 15,000 technical indicators were extracted, and a planning case library containing 500 typical cases was constructed. Of these, 120 cases focused on electric energy substitution and 150 on renewable energy consumption, representing over 50% of the total, demonstrating a strong focus on key scenarios. By leveraging this case library, new planning projects can quickly reference historical plans for similar scenarios, improving planning efficiency and quality.

[0081] S201: Extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library including:

[0082] Extract the scenarios, goals, and technical parameters of distribution network planning, analyze, and generate standardized data items required for distribution network planning under different scenarios;

[0083] Abstract the normalized data items as "input" and integrate the corresponding actual data of the current operation of the distribution network as "output";

[0084] Strengthen the organization of thematic scenarios such as new energy consumption, industrial power substitution, two-way interaction of electric vehicles, and virtual power plants, and construct additional parameter items to describe the scenarios more accurately.

[0085] S300: Integrates external open-source data and source-side data, further structures it, and forms a large database for distribution network planning. It also builds a large data access interface for distribution network planning to support data access using large language models.

[0086] In this example, integrating external open-source data with source-side data is crucial for enriching the information dimension of distribution network planning. Distribution network planning involves not only the grid's own data but also external factors such as meteorology, economy, and population. Therefore, it is necessary to integrate heterogeneous data from multiple sources to support comprehensive analysis.

[0087] The collection of external open-source data primarily falls into five categories: meteorological data, holiday data, geographic information data, city and regional characteristic data, and economic data. Meteorological data includes historical weather records, temperature changes, sunshine intensity, wind speed, and other factors, which directly affect renewable energy power generation output and residential electricity demand. Holiday data includes statutory holidays, adjusted workdays, and major events, which are often accompanied by abnormal fluctuations in electricity load. Geographic information data includes satellite imagery, topography, administrative divisions, and transportation networks, which help understand the spatial distribution characteristics of the distribution network. City and regional characteristic data includes population density, building types, and land use planning, which influence load distribution and growth trends. Economic data includes GDP growth rate, industrial structure, and fixed asset investment, which help predict long-term changes in regional electricity demand.

[0088] Data acquisition channels are diverse, including the Meteorological Administration's public data platform, the National Bureau of Statistics database, the government's open data portal, and industry data sharing platforms. Some non-public but valuable data is purchased through commercial data service providers or acquired through data sharing agreements with relevant institutions. The system has established a data source catalog management mechanism, recording information such as the source, update frequency, data format, and access method of various data types to facilitate automated collection and management.

[0089] Source-side data primarily refers to data from power sources, including conventional and renewable energy generation data. For conventional power sources like thermal and hydropower, information is collected on installed capacity, output characteristics, and regulation capabilities. For renewable energy sources like photovoltaic and wind power, data on output curves, fluctuation characteristics, and geographic distribution is focused. Due to the high uncertainty inherent in renewable energy generation, the system has established a historical renewable energy output database, recording output variations under different weather conditions and seasons to provide a data foundation for output forecasts.

[0090] Data association is a core step in fusion. The system establishes a multidimensional data association system, linking data from different sources and types across time, space, and business dimensions. For example, distribution network load data for a region can be associated with temperature data and holiday information for the same period to analyze the correlation between load and external factors. Output data for a photovoltaic power station can also be associated with meteorological data to establish a mapping between sunlight intensity and power generation. This multidimensional association transforms data from isolated points of information into a semantically rich knowledge network.

[0091] The distribution network planning big data access interface bridges the gap between the database and the large language model. Designed in a RESTful style, the interface provides a standardized HTTP API, enabling the large language model to retrieve required data through simple requests. The interface features data query, statistical analysis, and trend forecasting. Input parameters support natural language descriptions, and output is in a structured JSON format, making it easier for the large language model to process and understand.

[0092] Specifically, the access interface implements a semantic understanding layer that converts natural language queries from large language models into precise data manipulation instructions. For example, when the model asks "What is the growth trend of power load in a certain region over the past three years?", the semantic understanding layer analyzes the time range (the past three years), geographic scope (a certain region), data type (power load), and analysis requirements (growth trend). It then converts the query into a corresponding database query statement, executes the query, and returns the results. This natural language interaction method significantly reduces the technical barriers to data access for large language models.

[0093] In terms of security controls, the interface implements mechanisms such as identity authentication, permission verification, and operation auditing to ensure secure and controllable data access. Furthermore, to prevent excessive system load from excessive access to large language models, the interface implements access frequency limits and query complexity controls, balancing the needs of data openness with system protection.

[0094] It should be noted that the distribution network planning database is not only a collection of data but also a carrier of knowledge. By processing and correlating raw data, the system extracts domain knowledge and patterns relevant to distribution network planning. For example, by analyzing electricity consumption data from different regions and seasons, typical load curve characteristics can be summarized; by correlating renewable energy output with meteorological data, power generation forecasting models for different weather conditions can be established. This knowledge, stored in the system in the form of data models, parameter relationships, and prediction algorithms, becomes a vital resource for planning and analysis using the Big Language Model.

[0095] Data updating and maintenance are key to ensuring the continued effectiveness of the database. The system establishes a tiered data update mechanism: high-frequency operational data (such as load and power generation) is updated in real time or near real time; medium-frequency business data (such as equipment inventory and customer information) is updated in batches on a regular basis; and low-frequency basic data (such as geographic information and historical statistics) is updated on demand. The system also implements data lifecycle management, setting different retention periods and archiving strategies for different data types to balance data value and storage costs.

[0096] For example, the distribution network planning big data database constructed in this embodiment integrates nearly 10 years of distribution network operation data for a province, 5 years of meteorological data, 3 years of economic and social development data, as well as province-wide geographic information and new energy station data, with a total storage capacity of 50TB. Through the big data access interface, the big language model can easily query complex information such as "the ratio of maximum load to installed photovoltaic capacity in a city's industrial park in 2020," obtaining accurate data-backed answers and providing a solid foundation for distribution network planning analysis.

[0097] S301: Fusion of external open source data and source-side data includes:

[0098] Supplement open source data such as meteorological data, holiday data, geographic information data, city and regional characteristics data, and economic data;

[0099] Integrate power generation data, especially data from renewable energy sites such as photovoltaic and wind power, and perform structured processing;

[0100] The open source data is associated and integrated with the distribution network planning case data to form a complete distribution network planning database.

[0101] S400: Based on machine learning, a dual-side prediction model for the distribution network source and load sides and a distribution network planning solution generation model are constructed, forming an access interface for the distribution network planning machine learning model;

[0102] In this embodiment, building a two-sided prediction model and a planning solution generation model is the core step in achieving intelligent planning. Due to the dual uncertainty of renewable energy and electricity load, traditional deterministic planning methods are no longer suitable for the development needs of modern distribution networks. Therefore, machine learning technology is introduced for prediction and optimization.

[0103] The distribution network source-side prediction model primarily addresses the problem of forecasting renewable energy power generation output. The model uses meteorological forecast data, historical output data, and geographic data as input to predict renewable energy output for future periods. The model architecture utilizes an ensemble learning approach, combining the strengths of multiple algorithms. For photovoltaic power generation, a modified XGBoost algorithm is used to capture the nonlinear relationship between solar intensity and power generation. For wind power, a time series LSTM network is used to model the temporal characteristics of wind speed variations. For small, distributed renewable energy sources, a random forest algorithm is used to account for output characteristics influenced by multiple variables. The outputs of these multiple models are weighted and fused to produce the final forecast, with weights dynamically adjusted based on historical forecast accuracy.

[0104] To improve forecast accuracy, the model incorporates a multi-granularity forecasting strategy. Short-term forecasts (within an hour) use a 5-minute granularity and are primarily used for real-time scheduling; medium-term forecasts (from one day to one week) use an hourly granularity and are used for day-ahead planning; and long-term forecasts (from one month to one year) use a daily or monthly granularity for capacity planning. Predictions at different time scales utilize different feature combinations and model parameters, balancing computational complexity and forecast accuracy.

[0105] To train the source-side prediction model, we employed a combination of feature engineering and deep learning. During feature engineering, we constructed a rich feature set, including primitive features (such as temperature, wind speed, and sunshine duration), temporal features (such as time of day, season, and year), statistical features (such as recent averages, maximums, and minimums), and interactive features (such as the product of temperature and wind speed). These features were screened through feature importance analysis, retaining the most predictive subset. During deep learning, we utilized autoencoders for feature dimensionality reduction and noise filtering, extracting essential patterns from the data. Predictions were then made using a supervised learning model.

[0106] In an optional embodiment, the source-side prediction model also integrates the advantages of physical and data-driven models. For photovoltaic power generation, a physical photovoltaic conversion model based on panel parameters and installation angle is introduced, which is used in parallel with the machine learning model for prediction, and then fused through confidence weighting. For wind power, the power curve characteristics of wind turbines are combined with adaptive corrections from machine learning to improve prediction accuracy under extreme weather conditions. This hybrid physical-data prediction approach demonstrates significant advantages in prediction accuracy and robustness.

[0107] The load-side forecasting model for the distribution network primarily addresses the problem of predicting electricity load. This model uses historical load data and incorporates external factors (such as temperature, holidays, and economic indicators) to predict future load changes. Considering the diversity of loads, the system employs a categorized modeling strategy for different load types: for industrial loads, factors such as production plans and capacity utilization are considered; for commercial loads, factors such as business hours and customer flow are considered; and for residential loads, factors such as temperature and daily routines are considered. Forecast results for each load type are aggregated in a bottom-up manner to form a regional total load forecast.

[0108] Load forecasting utilizes a hierarchical modeling strategy. The bottom layer provides micro-forecasts for individual users or substations, the middle layer provides regional forecasts for feeder lines or substation service areas, and the top layer provides macro-forecasts for the entire planning unit. Forecasts at different levels are cross-checked and revised to ensure consistency and reliability. The system also establishes a library of typical daily load curves, including standard curves for different seasons, weather conditions, and different date types (weekdays, weekends, and holidays), providing a reference for load forecasting in areas without historical data.

[0109] Specifically, for new loads like electric vehicles, the system employs a forecasting approach based on behavioral analysis. By collecting data on charging station usage, vehicle trajectories, and charging habits, it builds an electric vehicle charging demand model and predicts the spatiotemporal distribution of charging loads. For industrial electricity replacement loads, scenario analysis is used to predict the progress and scale of future electricity replacement, taking into account factors such as industrial policy, technological maturity, and economic benefits.

[0110] The integration of two-sided forecasting models is key to achieving comprehensive forecasting capabilities. Based on the target distribution network topology, the system analyzes the power flow distribution constraints at each node in the network and balances the forecast results for the source and load sides. Through the power flow analysis model, the network's operational status under the forecast scenario is verified, identifying potential overloads, undervoltages, and other issues, providing targeted improvement directions for planning. Furthermore, the system analyzes the source-load interaction, such as the matching of renewable energy output with the load curve and the degree of peak shifting, to assess the distribution network's self-balancing capabilities.

[0111] The distribution network planning scheme generation model generates optimal planning schemes based on two-sided predictions using a hybrid integer optimization algorithm. The model uses the output of the two-sided prediction model and distribution network big data as input, and outputs planning schemes that include key parameters such as maximum load, load density, load growth rate, network topology, line cross-sectional flow, investment economics analysis, and reliability analysis.

[0112] The model is based on a multi-objective optimization framework, simultaneously considering multiple objectives, including economic efficiency, reliability, and flexibility. Economic objectives include minimizing investment, operating, and loss costs; reliability objectives include maximizing power supply reliability indicators (such as SAIDI and SAIFI); and flexibility objectives include maximizing renewable energy integration and load growth adaptability. These objectives are weighted to form a comprehensive objective function, with weights adjustable based on planning preferences.

[0113] S401: Based on machine learning, a two-sided prediction model for the source and load sides of the distribution network is constructed, including:

[0114] By using factors such as meteorological forecast data, historical power output data on the source side, and geographical data, a basic model for power output prediction on the source side is constructed;

[0115] By analyzing the load side characteristics, a basic load forecasting model is constructed;

[0116] According to the topology of the target distribution network, the flow constraints are analyzed and the two basic models are reasonably combined to form a two-sided prediction capability.

[0117] S401: The distribution network planning scheme generation model includes:

[0118] Using the output of the two-sided prediction model and the distribution network big data as input, the planning scheme of the target distribution network is output;

[0119] The planning scheme includes important parameters such as maximum load, load density, load growth rate, network topology, line cross-sectional flow, investment economic analysis, and reliability analysis.

[0120] S500: Based on the open-source large language model, it performs secondary pre-training to generate a large language model for a specific domain. It then provides practical packaging for the large language model, enhancing user interaction capabilities, access to a large distribution network planning database, the ability to call distribution network planning solution generation models, and the ability to generate final solutions.

[0121] S501: Performing secondary pre-training includes:

[0122] Organize the data, logic, and specialized vocabulary used in the system construction process into a training library for a large language model;

[0123] Through incremental pre-training, instruction supervision fine-tuning, and reward model training, the model is restricted to the vertical field of service distribution network planning, and the context is pre-processed.

[0124] S600: Complete the guided acquisition of planning scenarios, goals, and technical indicators through question-and-answer format to obtain the target microgrid planning solution.

[0125] Furthermore, this embodiment also provides a vertical field distribution network planning system based on a large language model, including:

[0126] The data collection and collation module is used to collect and collate distribution network structure and operation data, historical distribution network planning scheme data, structure them, and build indexes to form a "one-picture" management system for the distribution network;

[0127] A case library construction module is used to extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library;

[0128] The big data fusion module is used to integrate external open source data and source-side data, and further structure them to form a distribution network planning big data database. It also builds a distribution network planning big data access interface to support large language models to access data.

[0129] The model access module is used to build a two-sided prediction model for the source and load sides of the distribution network and a distribution network planning solution generation model based on machine learning, forming an access interface for the distribution network planning machine learning model;

[0130] The large language model pre-training module is used to perform secondary pre-training based on the open source large language model to generate a large language model for a specific vertical domain. The module also provides practical packaging for the large language model, improving user interaction capabilities, access to the large database for distribution network planning, the ability to call the distribution network planning solution generation model, and the ability to generate the final solution.

[0131] The user interaction module is used to complete the guided acquisition of planning scenarios, goals and technical indicators through question-and-answer format to obtain the target microgrid planning scheme.

[0132] In summary, by building a multi-dimensional indexing system (scenario, time, region, scale, feeder, and substation) and knowledge graph technology, comprehensive visualization of distribution network asset and operational data is achieved. This integrated data management approach enables planners to intuitively and quickly access comprehensive information, avoiding the fragmented and difficult-to-find data challenges inherent in traditional approaches. This significantly improves data access efficiency and utilization, laying a solid data foundation for subsequent precise planning. In particular, the collection and organization of new data, such as those on renewable energy generation and electric vehicles, effectively addresses the challenges of bilateral uncertainty in the current distribution network. By applying natural language processing technology and knowledge-enhanced deep learning models, key information from historical planning documents is effectively extracted and structured, establishing a standardized case library. This case library intelligently aggregates planning experience and best practices, transforming implicit expert knowledge into explicit, reusable assets. This significantly reduces reliance on expert experience and addresses the inconsistent planning quality inherent in traditional approaches, which can lead to uneven expert expertise. The specific emphasis on case studies on electricity substitution and renewable energy consumption further enhances the system's ability to address the challenges of energy transition.

[0133] By integrating heterogeneous external data from multiple sources, including meteorological, holiday, geographic, urban, and economic data, the information dimension of distribution network planning has been significantly expanded, establishing a rich data association system. This multidimensional association transforms data from isolated information points into a semantically rich knowledge network, overcoming the limitations of traditional methods that often suffer from large errors due to a limited number of analysis parameters. Furthermore, a standardized RESTful API interface and a natural language-based semantic understanding layer enable seamless integration between large language models and specialized data, significantly lowering the technical barrier and improving the system's applicability.

[0134] By pre-training and packaging an open-source large language model for specialized domains, we have built a vertically intelligent assistant with specialized capabilities for distribution network planning. This specialized large language model overcomes the limitations of traditional planning tools, which often have limited interaction methods and high learning barriers. By leveraging natural language interaction, it significantly improves the system's usability, enabling non-professionals to participate in the planning process. Furthermore, the various automated capabilities (data access, model invocation, and solution generation) incorporated into the model's packaging significantly reduce planning complexity and improve efficiency.

[0135] Through natural language question-and-answer interaction, the system enables guided acquisition of key planning parameters, simulating the experience of expert consulting services. This interactive approach overcomes the poor interactivity of traditional planning tools, allowing users to express their needs in a natural, conversational manner without having to understand complex technical terminology and operational procedures. Through intelligent guidance, the system ensures comprehensive and accurate acquisition of planning parameters, reduces the risk of user error, improves planning quality, partially replaces manual labor, and reduces planning costs.

[0136] Example 2

[0137] Reference Figure 1 - Figure 3 , which is the second embodiment of the present invention.

[0138] Collect and organize existing distribution network structure and operation data, and historical distribution network planning scheme data: collect and organize relevant data, structure them, and build indexes such as scenario, time, region (planning unit), scale, feeder, and substation, basically realizing "one-picture" management of distribution network and its planning data, helping users to effectively grasp the current status and history of distribution network.

[0139] Planning scenarios, objectives, and technical indicators are extracted from historical planning schemes to form a typical distribution network planning case library, with special emphasis on cases involving electricity substitution and new energy consumption. The following extracts the scenarios, objectives, and technical parameters of distribution network planning, thereby analyzing and forming the standardized data items required for distribution network planning under different scenarios. These data items are abstracted as "inputs," and the corresponding actual operating data of the distribution network is integrated as "outputs." On this basis, paired inputs and outputs are organized into a typical distribution network planning case library to provide support for the subsequent construction of a large database. In particular, to meet the requirements of the current new situation, thematic scenarios such as new energy consumption, industrial electricity substitution, two-way interaction of electric vehicles, and virtual power plants are strengthened and organized. Additional parameter items are constructed to more accurately describe the scenarios, thereby achieving the construction of a distribution network planning case library that is currently adapted to the current era.

[0140] Integrate external open source data and source-side data, and further structure them to form a large database for distribution network planning: supplement open source data such as meteorological data, holiday data, geographic information data, city and regional characteristic data, economic data, etc., integrate power generation side data, especially data from new energy sites such as photovoltaic and wind power, and perform structured processing, and associate and integrate them with distribution network planning case data to form a complete large database for distribution network planning.

[0141] Build a big data access interface for distribution network planning: Build a big database interface that can be accessed by a large language model to support the large language model in obtaining and analyzing data.

[0142] Based on LGBT machine learning, a two-sided prediction model for the source and load sides of the distribution network is constructed. A distribution network planning scheme generation model is constructed based on a hybrid integer optimization mathematical model. Based on LGBT or other appropriate machine learning algorithms, a two-sided prediction model for the distribution network (source and load) is constructed. This model uses factors such as meteorological forecast data, historical source-side output data, and geographic data to construct a basic output prediction model for the source side. It then analyzes the characteristics of the load side, namely distribution network data, to construct a basic load prediction model. Finally, based on the topology of the target distribution network and flow constraints, the two basic models are rationally combined for application, forming a two-sided prediction capability. Furthermore, a distribution network planning scheme generation model is constructed based on the hybrid integer optimization data model. This model uses the output of the aforementioned two-sided model and distribution network big data as input to output a planning scheme for the target distribution network, including key parameters such as maximum load, load density, load growth rate, network topology, line cross-sectional flow, investment economics analysis, and reliability analysis.

[0143] Form an access interface for the distribution network planning machine learning model: Build a model interface that can be accessed by the large language model to support the large language model in calling the model, analyzing, and generating planning solutions.

[0144] Based on the open source big language model, secondary pre-training is performed to generate a vertical domain big language model: the data, logic, and specialized vocabulary used in the system construction process are organized into a training library for the big language model. Secondary training is then performed based on the open source big language model, limiting the model to the vertical domain of serving distribution network planning. The context is pre-processed to complete the specialized construction of distribution network planning.

[0145] Practical encapsulation of large language models: Based on the distribution network-specific construction described above, it supplements user interaction capabilities, distribution network planning large database access capabilities, distribution network planning solution generation model call capabilities, and final solution generation capabilities, so that the model can be used in a user-friendly manner and the required data or models can be accessed intelligently and automatically.

[0146] User interaction: Based on specialized encapsulation, it enables friendly interaction between large language models and users, and has the ability to complete the guided acquisition of planning scenarios, goals and technical indicators through question-and-answer format, thereby achieving natural language communication-based acquisition of key parameters required for distribution network planning, and simulating expert services through software systems, partially replacing human labor.

[0147] In general, this method can be implemented by pure software or a combination of software and hardware. Both BS and CS architectures can implement the method. There are no special requirements for the programming language and underlying operating system. It is recommended to use Python to implement the core model.

[0148] "Collect and organize existing distribution network structure and operation data, as well as historical distribution network planning data." For power grid companies, this can be collected and organized from relevant systems such as distribution network planning, dispatching, and assets. For consulting, planning, construction, and other non-grid companies, this can be obtained and constructed from project cases, industrial park sharing, and open source databases. Particular attention should be paid to and organized from data on renewable energy power generation, new energy vehicles, energy storage systems, and virtual power plants to adapt to and address the current uncertainties on both sides of the distribution network.

[0149] "Integrate external open-source data with source-side data, further structure it, and form a large-scale database for distribution network planning." Data can be obtained from the internet, academic databases, publications by relevant departments, and collections by domestic and international non-governmental organizations. It can also be accumulated through offline visits and research. Environmental data, such as energy prices, weather data, and carbon emission indices, can be collected simultaneously from data released by relevant departments.

[0150] Based on LGBT machine learning, a two-sided prediction model for the source and load sides of the distribution network is constructed. After selecting a machine learning model and preparing the training data, the model can be trained and tuned using PyTorch or other open source technologies, following typical machine learning model training methods. The model's final input is comprehensive distribution network big data, and its output is predicted data for the target distribution network and its associated source side.

[0151] "Building a distribution network planning scheme generation model based on a hybrid integer optimization data model": In the final stage of distribution network planning, after completing source-side output forecasts and load-side load forecasts, the main issues for distribution network planning shift to optimal equipment investment analysis and topology analysis to balance economic and safety requirements. This can generally be further abstracted into a hybrid integer optimization mathematical problem. This model can be built independently using open source libraries such as pyomo, or other mature solutions can be utilized.

[0152] Based on the open-source big language model, perform secondary pre-training to generate a vertical domain big language model: The feature parameters described above, as well as the knowledge base related to distribution network planning, are organized into the corpus required for the secondary pre-training big language model and put into training. Taking the distribution network planning of a region containing a mixed new technology and industrial park as an example, a guiding corpus can be constructed, such as "Human: I want to conduct distribution network planning for Region A. Assistant: OK, the distribution network information for Region A has been loaded from the database. Please tell me your planning goals and requirements. Region A was not found in the database. Please tell me more parameters, starting with your region. Human: Assume a new energy penetration rate of 15%, a comprehensive absorption rate greater than 95%, and an annual load growth rate of 5%. Assistant: OK, we have analyzed the technical requirements. If you need further requirements, please continue to describe them. If you want to start generating a distribution network planning solution, please tell me to start generating it."

[0153] After preparing the corpus, you can select a suitable open source library such as LLAMA and perform secondary pre-training on it. Other training methods for practicing the functions of the present invention on the open source large language model are also within the scope of protection, such as incremental pre-training, supervised fine-tuning, reward modeling, PPO training, and DPO training.

[0154] "User interaction": Build appropriate user interaction methods, generally text interaction, which can be expanded to voice interaction.

[0155] "Practical packaging of large language models": The packaging model is not limited and can be web, software, app, mini-program, etc. The server's working and deployment methods are not limited, including centralized and distributed.

[0156] The planning model is invoked, the returned results are analyzed, and then organized into user-friendly content. Based on the data returned by the model, further personalized packaging can be performed, such as supplementary line capacity, feeder connection suggestions, analysis of operating status over the next N years, analysis of power substitution, analysis of new energy consumption, and other extended secondary results of interest to users.

[0157] "Obtain the target microgrid planning scheme", the scheme form is not limited, including web chart presentation, document, report, graphic and text display, etc.

[0158] Example 3

[0159] This embodiment also provides a computer device, which is suitable for a vertical field distribution network planning method based on a large language model, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement a forced oscillation detection and positioning method for a distribution network as proposed in the above embodiment.

[0160] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0161] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A vertical domain distribution network planning method based on a large language model, characterized by: This includes collecting and organizing distribution network structure and operation data, historical distribution network planning scheme data, structuring them, and building indexes by scenario, time, region, scale, feeder, and substation to form a "one-picture" management system for the distribution network. Extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library, and strengthen cases of electricity substitution and new energy consumption; Integrate external open source data and source-side data, further structure them, form a large database for distribution network planning, and build a large data access interface for distribution network planning to support large language models to access data; Based on machine learning, a dual-side prediction model for the source and load sides of the distribution network and a distribution network planning scheme generation model are constructed, forming an access interface for the distribution network planning machine learning model; Based on the open source large language model, secondary pre-training is performed to generate a large language model for a vertical domain. The large language model is then practically packaged to enhance user interaction capabilities, access capabilities for the distribution network planning large database, the ability to call the distribution network planning solution generation model, and the ability to generate the final solution. Through the form of questions and answers, the guided acquisition of planning scenarios, goals and technical indicators is completed to obtain the target microgrid planning scheme.

2. The vertical domain distribution network planning method based on a large language model according to claim 1, characterized in that: The collection and collation of distribution network structure and operation data and historical distribution network planning data include: Pay attention to and organize new energy power generation data, new energy vehicle data, energy storage system data, and virtual power plants to adapt to and respond to the two-sided uncertainties of the current distribution network system.

3. The vertical domain distribution network planning method based on a large language model according to claim 2, characterized in that: Extracting planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library includes: Extract the scenarios, goals, and technical parameters of distribution network planning, analyze, and generate standardized data items required for distribution network planning under different scenarios; Abstract the normalized data items as "input" and integrate the corresponding actual data of the current operation of the distribution network as "output"; Strengthen the organization of thematic scenarios such as new energy consumption, industrial power substitution, two-way interaction of electric vehicles, and virtual power plants, and construct additional parameter items to describe the scenarios more accurately.

4. The vertical domain distribution network planning method based on a large language model according to claim 3 is characterized by: The integration of external open source data and source-side data includes: Supplement open source data such as meteorological data, holiday data, geographic information data, city and regional characteristics data, and economic data; Integrate power generation data, especially data from renewable energy sites such as photovoltaic and wind power, and perform structured processing; The open source data is associated and integrated with the distribution network planning case data to form a complete distribution network planning database.

5. The vertical domain distribution network planning method based on a large language model according to claim 4 is characterized in that: The machine learning-based construction of a two-sided prediction model for the source and load sides of the distribution network includes: By using factors such as meteorological forecast data, historical power output data on the source side, and geographical data, a basic model for power output prediction on the source side is constructed; By analyzing the load side characteristics, a basic load forecasting model is constructed; According to the topology of the target distribution network, the flow constraints are analyzed and the two basic models are reasonably combined to form a two-sided prediction capability.

6. The vertical domain distribution network planning method based on a large language model according to claim 5, characterized in that: The distribution network planning scheme generation model includes: Using the output of the two-sided prediction model and the distribution network big data as input, the planning scheme of the target distribution network is output; The planning scheme includes important parameters such as maximum load, load density, load growth rate, network topology, line cross-sectional flow, investment economic analysis, and reliability analysis.

7. The vertical domain distribution network planning method based on a large language model according to claim 6, characterized in that: The secondary pre-training comprises: Organize the data, logic, and specialized vocabulary used in the system construction process into a training library for a large language model; Through incremental pre-training, instruction supervision fine-tuning, and reward model training, the model is restricted to the vertical field of service distribution network planning, and the context is pre-processed.

8. A vertical field distribution network planning system based on a large language model, based on the vertical field distribution network planning method based on a large language model according to any one of claims 1 to 7, characterized in that: It also includes a data collection and collation module, which is used to collect and collate distribution network structure and operation data, historical distribution network planning scheme data, structure them, and build indexes to form a "one-picture" management system for the distribution network; A case library construction module is used to extract planning scenarios, goals, and technical indicators from historical planning schemes to form a typical distribution network planning case library; The big data fusion module is used to integrate external open source data and source-side data, and further structure them to form a distribution network planning big data database. It also builds a distribution network planning big data access interface to support large language models to access data. The model access module is used to build a two-sided prediction model for the source and load sides of the distribution network and a distribution network planning solution generation model based on machine learning, forming an access interface for the distribution network planning machine learning model; The large language model pre-training module is used to perform secondary pre-training based on the open source large language model to generate a large language model for a specific vertical domain. The module also provides practical packaging for the large language model, improving user interaction capabilities, access to the large database for distribution network planning, the ability to call the distribution network planning solution generation model, and the ability to generate the final solution. The user interaction module is used to complete the guided acquisition of planning scenarios, goals and technical indicators through question-and-answer format to obtain the target microgrid planning scheme.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the vertical field distribution network planning method based on a large language model as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the vertical field distribution network planning method based on a large language model according to any one of claims 1 to 7 are implemented.

Images