Substation site selection question and answer method and device based on large language model and storage medium
By constructing a knowledge graph model of substations and using retrieval enhancement technology, the problems of insufficient knowledge coverage and strong subjectivity in substation site selection by large language models are solved, enabling more accurate and scientific site selection decisions and adapting to dynamic environmental changes.
Patent Information
- Application Number
- CN202511502479.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-21
- Publication Date
- 2026-01-20
AI Technical Summary
Existing large language models suffer from insufficient knowledge coverage, strong subjectivity, inability to update in real time, and inability to adapt to dynamic environmental factors in substation site selection, resulting in inaccurate site selection and insufficient scientific basis for decision-making.
By constructing a knowledge graph model of substations, using the BERT-BiLSTM-CRF model for question classification and intent recognition, and combining retrieval enhancement techniques to generate retrieval enhancement question vectors, and using a space-time-source composite index for retrieval matching, more accurate site selection answers are generated.
It improves the accuracy and scientific nature of substation site selection, adapts to the temporal evolution characteristics of environmental factors and load demand, reduces knowledge blind spots and subjective biases, and supports dynamic knowledge updates.
Smart Images
Figure CN121365844A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of substation site selection question and answer, and in particular relates to a substation site selection question and answer method and device based on a large language model and a storage medium. BACKGROUND
[0002] Substation site selection is a core link of power grid planning, and needs to be balanced among geological stability, environmental sensitivity, load center distribution, policy and regulation, economic cost, public acceptance and other multi-dimensional constraints. The traditional site selection method relies on manual collection of geological exploration reports, environmental impact assessment, load prediction curves, land space planning maps and other large amounts of heterogeneous data, and then through expert consultation, site reconnaissance, scheme comparison and government approval, the cycle is long, the cost is high, and once the ecological red line is adjusted, the load hotspot is migrated or the policy index is updated, the original scheme often cannot respond in time, causing the planning to lag behind or even being forced to re-argue.
[0003] In recent years, large language models (LLM) have been tried to be used for intelligent question and answer in the power field with their strong natural language understanding and generation capabilities, in order to reduce the professional threshold and shorten the decision-making cycle through dialogue interaction. However, in the highly specialized vertical field of substations, general large language models have obvious limitations.
[0004] Firstly, the model training corpus is mainly derived from public network text, and there is insufficient coverage of special knowledge such as survey drilling data and load data of power grid enterprises, resulting in knowledge blind spots and even "hallucination" errors when answering detailed questions such as "whether there is an active fault zone in a certain area" and "whether a certain plot avoids a first-level water source protection area".
[0005] Secondly, environmental factors and load demand have obvious time evolution characteristics: the scope of ecological red line and wetland protection area may be adjusted according to annual remote sensing monitoring results, and the operation of emerging industrial parks or the large-scale grid connection of distributed photovoltaics will also quickly change the spatio-temporal distribution of load. The general large language model is frozen after training, and cannot absorb the latest site selection knowledge, remote sensing interpretation results or load prediction updates in real time, so the output results are out of date and cannot support dynamic site selection decisions.
[0006] More seriously, substation site selection often involves multiple subjects, and different subjects have different or even opposite perceptions of concepts such as "safety distance" and "load coordination". The existing large language model tends to generate answers that "users want to hear" rather than neutral conclusions based on objective data and specifications in an open dialogue scenario, which amplifies subjective bias and weakens the scientific nature and public credibility of site selection decisions.
[0007] Therefore, there is an urgent need for a large language model question and answer method that can break through knowledge boundaries, support dynamic knowledge updates, and suppress cognitive bias to meet the comprehensive requirements of real-time, accuracy, and objectivity for substation site selection. SUMMARY
[0008] Based on the problems raised in the above background art, the purpose of the present application is to provide a substation site selection question and answer method based on a large language model, a device and a storage medium, which solves the problems of subjective question and answer and insufficient knowledge coverage of the existing large language model for substation site selection.
[0009] The present application is realized by the following technical solutions:
[0010] The present application provides a substation site selection question and answer method based on a large language model in the first aspect, comprising the following steps:
[0011] Obtain a substation site selection question, classify the substation site selection question, and obtain a question entity and a question intent;
[0012] Based on the question intent, the question entity is retrieved and enhanced to generate a retrieval enhanced question vector;
[0013] A substation knowledge graph model is constructed, the retrieval enhanced question vector is input into the substation knowledge graph model for retrieval matching, and a substation site selection answer is generated according to the retrieval matching result.
[0014] In the above technical solution, the substation site selection question is a text information input by the user, after obtaining the substation site selection question, the text is processed, and the question classification is realized through the text processing to obtain the question entity and the question intent; wherein the question entity refers to the question text itself, and the question intent refers to the substation site selection related requirements obtained after semantic recognition based on the text content, for example: the substation needs to be established in a certain area, the substation site selection needs to avoid the first water source protection area and other substation site selection intents.
[0015] The site selection intent is used as the question intent of the substation site selection question to retrieve and enhance the question entity, and the retrieval enhanced question vector is used for subsequent retrieval matching, and the retrieval enhanced content is matched with the knowledge to avoid the knowledge blind area of the professional knowledge, so as to avoid the inaccuracy and unprofessional of the substation site selection.
[0016] The substation knowledge graph model is constructed with space-time-source as a composite index, and the substation knowledge graph model and the retrieval enhanced question are matched in space, time and source three levels, so as to adapt to the time evolution characteristics of environmental factors and load demand, and make the substation site selection more in line with the present situation.
[0017] In an alternative embodiment, the substation siting problem is classified, including the following steps:
[0018] Identifying candidate words in the substation siting problem, performing entity recognition on the candidate words to obtain problem entities;
[0019] Performing intent recognition on the problem entities using a BERT-BiLSTM-CRF model to obtain problem intents.
[0020] In an alternative embodiment, the BERT-BiLSTM-CRF model is used to perform intent recognition on the problem entities, including:
[0021] Using a word segmentation method to segment the problem entities to obtain a problem sequence;
[0022] Inputting the problem sequence into an embedding layer for embedding to obtain word embeddings, position embeddings, and sentence embeddings;
[0023] Using a BERT module to process the word embeddings, position embeddings, and sentence embeddings to obtain semantic sensitive feature vectors;
[0024] Using a BiLSTM module to process the semantic sensitive feature vectors in both directions to obtain time sequence semantic labels;
[0025] Using a CRF module to perform path analysis on the time sequence semantic labels, and generating problem intents based on the path analysis results.
[0026] In an alternative embodiment, the problem entities are enhanced based on the problem intents, including the following steps:
[0027] Using a word segmentation method to segment the problem intents to obtain several intent image segments;
[0028] Using an embedding model to map each intent image segment to a vector space to generate several embedding vectors;
[0029] Obtaining substation siting retrieval information, performing vector aggregation on the several embedding vectors and the substation siting retrieval information to generate several enhanced intent vectors;
[0030] Embedding the several enhanced intent vectors into the problem entities to obtain an enhanced problem.
[0031] In an alternative embodiment, the several enhanced intent vectors are embedded into the problem entities, including the following steps:
[0032] Using a word segmentation method to segment the problem entities to obtain a problem sequence;
[0033] Perform semantic recognition on the question sequence, and match the question sequence with a plurality of enhanced intention vectors according to a result of the semantic recognition;
[0034] Embed the plurality of enhanced intention vectors into the question sequence according to a matching result, to obtain an enhanced question sequence;
[0035] Integrate the enhanced question sequence to form a retrieval enhanced question.
[0036] In an optional embodiment, a substation knowledge graph model is constructed, comprising:
[0037] Obtain substation data and map data, and pre-process the substation data and the map data to obtain standardized data;
[0038] Extract domain ontology, entities and relationships from the standardized data based on domain knowledge, and form triples;
[0039] Perform synonym ablation on the triples, obtain real-time data, and dynamically update the triples after synonym ablation using the real-time data, to obtain dynamic triples;
[0040] Establish a graph database with space-time-source as a composite index, load the dynamic triples into the graph database, and obtain a substation knowledge graph model.
[0041] In an optional embodiment, the retrieval enhanced question vector is input into the substation knowledge graph model for retrieval matching, comprising the following steps:
[0042] Perform time matching on the retrieval enhanced question, and filter a first node set from the substation knowledge graph model according to a result of the time matching;
[0043] Perform space matching on the retrieval enhanced question, and filter a second node set from the first node set according to a result of the space matching;
[0044] Perform source matching on the retrieval enhanced question, and filter a third node set from the second node set according to a result of the source matching;
[0045] Extract an enhanced intention vector from the retrieval enhanced question, perform retrieval matching on the enhanced intention vector and the third node set, and generate a substation site selection answer according to a result of the retrieval matching.
[0046] The second aspect of the application provides a substation site selection question and answer device based on a large language model, comprising:
[0047] The problem classification module is configured to obtain a substation site selection problem, perform problem classification on the substation site selection problem, and obtain a problem entity and a problem intention.
[0048] The retrieval enhancement module is configured to perform retrieval enhancement on the problem entity based on the problem intention, and generate a retrieval enhanced problem vector.
[0049] The retrieval matching module is configured to construct a substation knowledge graph model, input the retrieval enhanced problem vector into the substation knowledge graph model for retrieval matching, and generate a substation site selection answer based on a result of the retrieval matching.
[0050] In an optional embodiment, the retrieval enhancement module comprises:
[0051] The word segmentation unit is configured to perform word segmentation on the problem intention using a word segmentation method, and obtain a plurality of intention image segments.
[0052] The mapping unit is configured to map each intention image segment into a vector space using an embedding model, and generate a plurality of embedding vectors.
[0053] The aggregation unit is configured to obtain substation site selection information, perform vector aggregation on the plurality of embedding vectors and the substation site selection information, and generate a plurality of enhanced intention vectors.
[0054] The embedding unit is configured to embed the plurality of enhanced intention vectors into the problem entity, and obtain a retrieval enhanced problem.
[0055] The third aspect of the present application provides a computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the substation site selection question and answer method based on the large language model.
[0056] Compared with the prior art, the present application has the following advantages and beneficial effects:
[0057] 1. The retrieval enhanced problem vector is used for subsequent retrieval matching, and the content of retrieval enhancement is matched with knowledge to avoid the inaccuracy and unprofessionalism of substation site selection caused by the knowledge blind area of special knowledge.
[0058] 2. The substation knowledge graph model is constructed by taking space-time-source as a composite index, and the substation knowledge graph model and the retrieval enhanced problem are matched in three aspects of space, time and source, so as to adapt to the time sequence evolution characteristics of environmental factors and load demand, and make the substation site selection more in line with the current situation. BRIEF DESCRIPTION OF DRAWINGS
[0059] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those of ordinary skill in the art, other related drawings can also be obtained without creative labor. In the drawings:
[0060] Figure 1 The flowchart of the substation site selection question and answer method based on a large language model provided for Embodiment 1 of the present application;
[0061] Figure 2 The structural diagram of the substation site selection question and answer device based on a large language model provided for Embodiment 2 of the present application. DETAILED DESCRIPTION
[0062] In order to make the purpose, technical solutions and advantages of the present application more clear and obvious, the following will further describe the present application in combination with embodiments and drawings. The exemplary embodiments of the present application and their descriptions are only used to explain the present application, and should not be regarded as a limitation on the present application.
[0063] Embodiment 1 of the present application provides a substation site selection question and answer method based on a large language model, as shown in Figure 1 The substation site selection question and answer method based on a large language model includes the following steps:
[0064] Obtaining a substation site selection question, classifying the substation site selection question, obtaining a question entity and a question intention;
[0065] Based on the question intention, the question entity is retrieved and enhanced to generate a retrieval enhanced question vector;
[0066] A substation knowledge graph model is constructed, the retrieval enhanced question vector is input into the substation knowledge graph model for retrieval matching, and a substation site selection answer is generated according to the retrieval matching result.
[0067] It should be noted that the substation site selection question is the text information input by the user. After obtaining the substation site selection question, the text processing is performed on the substation site selection question, and the question classification is realized through the text processing to obtain the question entity and the question intention. The question entity refers to the question text itself, and the question intention refers to the substation site selection related requirements obtained after the semantic recognition based on the text content, for example: the substation needs to be established in a certain area, the substation site selection needs to avoid the first water source protection area and other substation site selection intentions.
[0068] The site selection intention is retrieved as the problem intention of the substation site selection problem, and the retrieval enhanced problem vector after retrieval enhancement is used for subsequent retrieval matching. The retrieval enhanced content is matched with knowledge to avoid the knowledge blind spot of professional knowledge, so as to avoid inaccurate and unprofessional substation site selection.
[0069] The substation knowledge graph model is constructed by taking space-time-source as a composite index, and the substation knowledge graph model is matched with the retrieval enhanced problem in space, time and source three aspects, so as to adapt to the time evolution characteristics of environmental factors and load demand, and make the substation site selection more in line with the current situation.
[0070] In an optional embodiment, the substation site selection problem is classified, including the following steps:
[0071] The candidate words in the substation site selection problem are identified, and the problem entities are obtained by performing entity recognition on the candidate words.
[0072] The problem entities are identified by using a BERT-BiLSTM-CRF model to obtain the problem intention.
[0073] BERT is Bidirectional Encoder Representations from Transformers, a bidirectional encoder representation based on Transformer. The semantic information on the left and right sides of the word is captured by the Transformer encoder at the same time, solving the problem that the traditional one-way model cannot fully understand the context.
[0074] BiLSTM is Bidirectional Long Short-Term Memory, a bidirectional long short-term memory network. In the forward LSTM, the input sequence is processed in time order to capture historical information. In the reverse LSTM, the sequence is processed in reverse order to capture future context. Finally, the bidirectional hidden state is combined by splicing or weighting operation to form a feature representation containing complete context information, so as to learn the time evolution characteristics of environmental factors and load demand in the text.
[0075] CRF is Conditional Random Field, a conditional random field used to recombine the substation site selection label related to environmental factors and load demand obtained by BiLSTM according to the time evolution characteristics.
[0076] In an optional embodiment, the problem entities are identified by using a BERT-BiLSTM-CRF model, including:
[0077] perform word segmentation on the question entity using a word segmentation method to obtain a question sequence;
[0078] input the question sequence into an embedding layer for embedding to obtain word embedding, position embedding and sentence embedding;
[0079] process the word embedding, the position embedding and the sentence embedding using a BERT module to obtain a semantic sensitive feature vector;
[0080] perform bidirectional processing on the semantic sensitive feature vector using a BiLSTM module to obtain a time sequence semantic label;
[0081] perform path analysis on the time sequence semantic label using a CRF module, and generate a question intention according to the path analysis result.
[0082] In an optional embodiment, the question entity is enhanced based on the question intention, including the following steps:
[0083] perform word segmentation on the question intention using a word segmentation method to obtain a plurality of intention image segments;
[0084] map each intention image segment into a vector space using an embedding model to generate a plurality of embedding vectors;
[0085] obtain substation site selection retrieval information, aggregate the plurality of embedding vectors with the substation site selection retrieval information to generate a plurality of enhanced intention vectors;
[0086] embed the plurality of enhanced intention vectors into the question entity to obtain an enhanced retrieval question.
[0087] It should be noted that the retrieval enhancement is the core idea of the method, wherein the question intention is first segmented into a plurality of intention image segments using a word segmentation method, each intention image segment includes a word or a complete word, such as "water source protection" and "high load", and then each intention image segment is mapped into a vector space using an embedding model. Specifically, in this embodiment, a Transformer model is used for dynamic embedding.
[0088] Then, the substation site selection retrieval information is obtained, wherein the substation site selection retrieval information is information obtained by retrieval on historical successful substation site selection, including its location, requirements, advantages and disadvantages, and then the plurality of embedding vectors are aggregated with the substation site selection retrieval information to obtain more representative site selection requirements of the substation site selection.
[0089] Finally, the enhanced intention vector is embedded into the question entity to obtain an enhanced retrieval question.
[0090] In an alternative embodiment, embedding a plurality of enhanced intent vectors into the question entity comprises the following steps:
[0091] segmenting the question entity using a segmentation method to obtain a question sequence;
[0092] performing semantic recognition on the question sequence, and matching the question sequence with a plurality of enhanced intent vectors according to the result of semantic recognition;
[0093] embedding a plurality of enhanced intent vectors into the question sequence according to the matching result to obtain an enhanced question sequence;
[0094] integrating the enhanced question sequence to form a retrieval enhanced question.
[0095] In an alternative embodiment, a substation knowledge graph model is constructed, comprising:
[0096] obtaining substation data and map data, and pre-processing the substation data and the map data to obtain standardized data;
[0097] extracting domain ontology, entities and relationships from the standardized data based on domain knowledge, and forming triples;
[0098] performing synonym ablation on the triples, obtaining real-time data, and dynamically updating the triples after synonym ablation using the real-time data to obtain dynamic triples;
[0099] establishing a graph database with space-time-source as a composite index, loading the dynamic triples into the graph database to obtain a substation knowledge graph model.
[0100] In an alternative embodiment, the retrieval enhanced question vector is input into the substation knowledge graph model for retrieval matching, comprising the following steps:
[0101] performing time matching on the retrieval enhanced question, and screening a first node set from the substation knowledge graph model according to the result of time matching;
[0102] performing space matching on the retrieval enhanced question, and screening a second node set from the first node set according to the result of space matching;
[0103] performing source matching on the retrieval enhanced question, and screening a third node set from the second node set according to the result of source matching;
[0104] extracting an enhanced intent vector from the retrieval enhanced question, performing retrieval matching on the enhanced intent vector and the third node set, and generating a substation site selection answer according to the result of retrieval matching.
[0105] Embodiment 2 of the present application provides a substation site selection question and answer device based on a large language model, as shown in Figure 2 The substation site selection question and answer device based on a large language model comprises:
[0106] A question classification module is configured to obtain a substation site selection question, perform question classification on the substation site selection question, and obtain a question entity and a question intention.
[0107] A retrieval enhancement module is configured to perform retrieval enhancement on the question entity based on the question intention, and generate a retrieval enhanced question vector.
[0108] A retrieval matching module is configured to construct a substation knowledge graph model, input the retrieval enhanced question vector into the substation knowledge graph model for retrieval matching, and generate a substation site selection answer according to the retrieval matching result.
[0109] In an optional embodiment, the retrieval enhancement module comprises:
[0110] A word segmentation unit is configured to perform word segmentation on the question intention using a word segmentation method to obtain a plurality of intention image segments.
[0111] A mapping unit is configured to map each intention image segment into a vector space using an embedding model to generate a plurality of embedding vectors.
[0112] An aggregation unit is configured to obtain substation site selection information, perform vector aggregation on the plurality of embedding vectors and the substation site selection information to generate a plurality of enhanced intention vectors.
[0113] An embedding unit is configured to embed the plurality of enhanced intention vectors into the question entity to obtain a retrieval enhanced question.
[0114] Embodiment 3 of the present application further provides a computer readable storage medium, and the computer executable instructions are used to implement the substation site selection question and answer method based on a large language model provided in Embodiment 1 when executed by a computer processor.
[0115] The storage medium provided in the embodiment of the present application comprises computer executable instructions, which are not limited to the method operations provided in Embodiment 1, and can also perform related operations in the substation site selection question and answer method based on a large language model provided in any embodiment of the present application.
[0116] The above detailed description of the specific embodiments of the present application has been given to understand the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A substation site selection question and answer method based on a large language model, characterized in that, a substation site selection question is obtained, the substation site selection question is classified, question entities and question intentions are obtained; the question entities are retrieved and enhanced based on the question intentions, and a retrieval and enhancement question vector is generated; a substation knowledge graph model is constructed, the retrieval and enhancement question vector is input into the substation knowledge graph model for retrieval matching, and a substation site selection answer is generated according to the retrieval matching result.
2. The substation siting question answering method based on a large language model according to claim 1, characterized in that, The substation site selection question is classified, including the following steps: candidate words in the substation site selection question are identified, and the candidate words are subjected to entity recognition to obtain question entities; a BERT-BiLSTM-CRF model is used to identify the intentions of the question entities to obtain question intentions.
3. The substation siting question answering method based on a large language model according to claim 2, characterized in that, The BERT-BiLSTM-CRF model is used to identify the intentions of the question entities, including: the question entities are subjected to word segmentation using a word segmentation method to obtain a question sequence; the question sequence is input into an embedding layer for embedding to obtain word embeddings, position embeddings, and sentence embeddings; the word embeddings, the position embeddings, and the sentence embeddings are processed using a BERT module to obtain semantic sensitive feature vectors; the semantic sensitive feature vectors are processed bidirectionally using a BiLSTM module to obtain time sequence semantic labels; the time sequence semantic labels are subjected to path analysis using a CRF module, and question intentions are generated according to the path analysis result.
4. The substation siting question answering method based on a large language model according to claim 1, characterized in that, The question entities are retrieved and enhanced based on the question intentions, including the following steps: the question intentions are subjected to word segmentation using a word segmentation method to obtain several intention pictures; each intention picture is mapped into a vector space using an embedding model to generate several embedding vectors; substation site selection retrieval information is obtained, the several embedding vectors are aggregated with the substation site selection retrieval information to generate several enhanced intention vectors; the several enhanced intention vectors are embedded into the question entities to obtain a retrieval and enhancement question.
5. The substation siting question answering method based on a large language model according to claim 4, characterized in that, The several enhanced intention vectors are embedded into the question entities, including the following steps: the question entities are subjected to word segmentation using a word segmentation method to obtain a question sequence; the question sequence is subjected to semantic recognition, and the question sequence is matched with the several enhanced intention vectors according to the semantic recognition result; the several enhanced intention vectors are respectively embedded into the question sequence according to the matching result to obtain an enhanced question sequence; the enhanced question sequence is integrated to form a retrieval and enhancement question.
6. The substation siting question answering method based on a large language model according to claim 1, characterized in that, The substation knowledge graph model is constructed, including: substation data and map data are obtained, and the substation data and the map data are preprocessed to obtain standardized data; domain knowledge is used to extract domain ontology, entities, and relationships from the standardized data, and form triples; synonym ablation is performed on the triples, real-time data is obtained, the real-time data is used to dynamically update the triples after synonym ablation, and dynamic triples are obtained; A graph database is established by taking space-time-source as a composite index, and the dynamic triplets are loaded into the graph database to obtain a substation knowledge graph model.
7. The substation siting question answering method based on a large language model according to claim 6, characterized in that, Time matching is performed on the retrieval enhanced question, and a first node set is screened from the substation knowledge graph model according to a result of the time matching; Space matching is performed on the retrieval enhanced question, and a second node set is screened from the first node set according to a result of the space matching; Source matching is performed on the retrieval enhanced question, and a third node set is screened from the second node set according to a result of the source matching; An enhanced intention vector is extracted from the retrieval enhanced question, the enhanced intention vector is matched with the third node set, and a substation site selection answer is generated according to a result of the matching.
8. A substation siting question answering device based on a large language model, characterized by, The substation site selection question and answer device for implementing the substation site selection question and answer method based on a large language model according to any one of claims 1 to 7 comprises: a question classification module configured to obtain a substation site selection question, perform question classification on the substation site selection question, and obtain question entities and question intentions; a retrieval enhancement module configured to perform retrieval enhancement on the question entities based on the question intentions, and generate a retrieval enhanced question vector; a retrieval matching module configured to construct a substation knowledge graph model, input the retrieval enhanced question vector into the substation knowledge graph model for retrieval matching, and generate a substation site selection answer according to a result of the retrieval matching.
9. The substation siting question answering apparatus based on a large language model according to claim 8, characterized by, The retrieval enhancement module comprises: a word segmentation unit configured to perform word segmentation on the question intentions using a word segmentation method to obtain a plurality of intention image segments; a mapping unit configured to map each of the intention image segments to a vector space using an embedding model to generate a plurality of embedding vectors; an aggregation unit configured to obtain substation site selection information, perform vector aggregation on the plurality of embedding vectors and the substation site selection information to generate a plurality of enhanced intention vectors; and an embedding unit configured to embed the plurality of enhanced intention vectors into the question entities to obtain a retrieval enhanced question.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the substation site selection question and answer method based on a large language model according to any one of claims 1 to 7.