Intelligent building method and system of building load and storage medium

Abstract

The invention discloses an intelligent building method and system of a building load and a storage medium. The intelligent building method of the building load comprises the following steps: S1, generating structured retrieval information containing building parameters according to demand information of a user; s2, based on the structured retrieval information, retrieving a semantic vector knowledge base to obtain load configuration parameters matched with the building parameters; s3, inputting the load configuration parameters and an IDF file template into an intelligent agent constructed by a large language model, and automatically constructing an IDF file; s4, inputting the IDF file and the meteorological file corresponding to the IDF file into a load calculation engine to obtain hourly load data; and S5, evaluating the hourly load data, if the hourly load data is reasonable, outputting the hourly load data, otherwise, returning to the step S3 to reconstruct the IDF file. According to the method, the calculation difficulty of the accurate hourly load data is simplified, and the use threshold of a user is reduced.

Description

Technical Field

[0001] This invention relates to the technical fields of artificial intelligence and building load construction, and in particular to an intelligent building load construction method based on a large language model (LLM). Background Technology

[0002] Accurate prediction of building load is fundamental for air conditioning system selection, building energy consumption simulation, and operational strategy optimization. Traditional calculation methods based on cooling load indices, also known as empirical estimation methods, use empirical parameters for cooling load per unit area, typically in W / m². The total cooling load is calculated by multiplying the building's total air-conditioned area by this cooling load index. The formula is: Total Cooling Load (kW) = Building Area (m²) × Cooling Load Index (W / m²) / 1000. For example, for a 10,000 square meter office building, if the empirically selected cooling load index is 100 W / m², then its estimated total cooling load is 10,000 × 100 / 1000 = 1000 kW. While the cooling load index-based calculation method is simple and quick, it simplifies the complex and dynamic load process into a static, fixed value, ignoring fluctuations in building load over time, weather, and internal usage. Therefore, its accuracy is limited and it cannot meet the needs of sophisticated application scenarios.

[0003] Besides calculation methods based on cooling indices, high-precision load calculations can also be achieved using professional simulation tools such as EnergyPlus and DeST. However, their modeling processes are complex, requiring users to manually write input files conforming to specific syntax rules and relying heavily on HVAC expertise, resulting in a high barrier to entry and hindering widespread adoption by ordinary users. Taking EnergyPlus modeling as an example, it requires defining the building's geometry and spatial division, as well as defining the detailed multi-layered material composition (such as cement mortar, insulation layer, brick wall, and interior finish) for each wall, roof, floor, and window. Users must also input thermal parameters such as the thickness, thermal conductivity, density, and specific heat capacity of each material layer, along with the changes in personnel, lighting, and equipment over different times. Furthermore, EnergyPlus's IDF file (Input Data File) is a text file that requires strict adherence to specific syntax rules. Users often need to manually write or modify thousands of lines of code; any minor syntax error (such as a missing semicolon or mismatched parentheses) or parameter logic error can lead to simulation failure or incorrect results, making the debugging process extremely time-consuming.

[0004] In recent years, large language models have demonstrated powerful capabilities in code generation and semantic understanding, but they have not yet been applied in the field of building load calculation. This is because large language models are only very good at semantic understanding, but lack HVAC expertise. Therefore, even if ordinary users express their needs through large language models such as Deepseek and ChatGpt, the large language models cannot obtain accurate building loads.

[0005] Therefore, how to provide an intelligent building load-bearing method that can be applied to ordinary users is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] In order to solve the technical problem that ordinary users cannot build building loads in the prior art, the present invention proposes an intelligent building load building method and system, and a storage medium.

[0007] The intelligent construction method for building load proposed in this invention includes:

[0008] Step S1: Based on the user's natural language input requirements and combined with the semantic vector knowledge base corresponding to the pre-built professional standard documents in the HVAC field, the requirements information is processed in a structured manner to generate structured retrieval information containing building parameters.

[0009] Step S2: Based on the structured retrieval information, retrieve the semantic vector knowledge base and obtain load configuration parameters that match the building parameters from the professional standard documents in the HVAC field;

[0010] Step S3: Input the load configuration parameters and IDF file template into the agent constructed by the large language model to automatically construct the IDF file;

[0011] Step S4: Input the IDF file and its corresponding meteorological file into the load calculation engine to obtain hourly load data;

[0012] Step S5: Evaluate the hourly load data. If it is reasonable, output it; otherwise, return to step S3 to rebuild the IDF file.

[0013] Furthermore, the semantic vector knowledge base is formed using the following steps:

[0014] Collect professional standard documents in the HVAC field;

[0015] Extract the table of contents and headings from the content of the professional standard document to form the logical structure of the professional standard document;

[0016] When dividing the content of the professional standard document into text blocks, the logical structure related to the paragraph content is attached to the corresponding paragraph as a sub-block text.

[0017] All sub-block text is input into the embedding model and transformed into high-dimensional vector data;

[0018] All high-dimensional vector data and the mapping relationship between high-dimensional vector data and the corresponding parts of the professional standard documents in the HVAC field are stored in the semantic vector knowledge base.

[0019] Furthermore, when retrieving the semantic vector knowledge base, the structured retrieval information is first converted into a query vector. By calculating the cosine similarity between the query vector and the vectors in the semantic vector knowledge base, the m high-dimensional vector data with the highest cosine similarity are obtained from the vector knowledge base.

[0020] The load configuration parameters are formed by using the content of the professional standard document corresponding to m high-dimensional vector data.

[0021] Furthermore, the load configuration information includes at least one of the following: building foundation information, building envelope parameters, building internal disturbance information, and operating parameters.

[0022] Furthermore, step S3 includes:

[0023] Step S31: Determine whether the building basic information in the load configuration parameters contains detailed room-level geometric information;

[0024] If included, the IDF file is generated based on the load configuration parameters and the IDF file template;

[0025] If not included, based on the building foundation information of the load configuration parameters, standard building templates are selected from the predefined building model library, the size information of the standard building templates is adapted and scaled, and then combined with the load configuration parameters to generate the IDF file;

[0026] Step S32: Verify the integrity of the generated IDF file;

[0027] In step S33, if the verification fails, the reason for the failure is corrected and the file is regenerated, and the process returns to step S32 until the number of corrections reaches the preset number; if the verification succeeds, the construction of the IDF file is completed.

[0028] Furthermore, in step S5, when evaluating the hourly load data, the hourly load data is converted into structured load characteristic parameters, and the rationality of the load characteristic parameters is evaluated in conjunction with relevant data retrieved from the semantic vector knowledge base.

[0029] Furthermore, at least one of steps S1, S2, S4, and S5 is processed by the corresponding intelligent agent generated by the large language model.

[0030] Furthermore, when steps S1 to S5 are all processed by the corresponding intelligent agents, steps S1 to S5 are processed through a workflow intelligent agent, which executes each step sequentially and calls the corresponding intelligent agent.

[0031] The intelligent building load erection system of the present invention includes:

[0032] The input processing module acquires the user's demand information based on natural language input, and combines it with a pre-built semantic vector knowledge base corresponding to professional standard documents in the HVAC field to perform structured processing on the demand information and generate structured retrieval information containing building parameters.

[0033] The knowledge base retrieval module is used to retrieve the semantic vector knowledge base based on the structured retrieval information, and to obtain load configuration parameters that match the building parameters from the corresponding professional standard documents based on the retrieved high-dimensional vector data.

[0034] The intelligent agent engine module generates different intelligent agents based on a large language model for structured processing, retrieves the semantic vector knowledge base, constructs IDF files, and evaluates hourly load data.

[0035] The simulation calculation module uses a load calculation engine to perform calculations based on the IDF file and its corresponding meteorological file to obtain hourly load data.

[0036] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent construction method for building load as described in any of the above technical solutions.

[0037] This invention builds a semantic vector knowledge base based on professional standard documents in the HVAC field, and automatically generates detailed load configuration information based on user needs. Then, based on the load configuration information, it automatically generates the corresponding IDF file through the corresponding intelligent agent generated by the large language model. After that, the IDF file and its corresponding meteorological file are sent to the load calculation engine to obtain hourly load data. Users do not need to have professional HVAC knowledge, and it can also provide an accurate data foundation for air conditioning selection.

[0038] Furthermore, this invention can also deeply integrate the general capabilities of large language models with the professional knowledge of the HVAC field by constructing a framework in which multiple professional intelligent agents work together, thereby realizing an end-to-end automated process from natural language to high-precision load calculation results, which significantly improves modeling efficiency and result reliability. Attached Figure Description

[0039] The present invention will now be described in detail with reference to the embodiments and accompanying drawings, wherein:

[0040] Figure 1 This is a flowchart of a load intelligent setup according to an embodiment of the present invention.

[0041] Figure 2 This is a flowchart illustrating the generation process of a semantic vector knowledge base according to an embodiment of the present invention.

[0042] Figure 3 This is a flowchart of the load configuration parameter generation process according to an embodiment of the present invention.

[0043] Figure 4 This is a flowchart illustrating the construction process of an IDF file according to an embodiment of the present invention.

[0044] Figure 5 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0045] Figure 6 This is a flowchart of the construction of a semantic vector knowledge base according to an embodiment of the present invention.

[0046] Figure 7 This is a schematic diagram illustrating the specific implementation process of an application embodiment of the present invention. Detailed Implementation

[0047] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0048] Therefore, a feature pointed out in this specification is used to describe one feature of one embodiment of the invention, and does not imply that every embodiment of the invention must have the described feature. Furthermore, it should be noted that this specification describes many features. Although certain features may be combined to illustrate possible system designs, these features may also be used in other combinations not explicitly stated. Therefore, unless otherwise stated, the described combinations are not intended to be limiting.

[0049] Large language models are currently a very popular intelligent tool. Common examples include OpenAI's ChatGPT, Google's Gemini, Anthropic's Claude, Baidu's Wenxin Yiyan, Deepseek, and Alibaba's Tongyi Qianwen. These large language models can not only create and edit content, but also generate corresponding code snippets based on users' natural language descriptions, and retrieve relevant information from the internet to respond to users' knowledge-based questions. For this invention to be applicable to the construction load calculation of ordinary users, a large language model is needed to at least help ordinary users generate professional IDF files that strictly adhere to specific grammatical rules, so that ordinary users can also obtain accurate construction loads.

[0050] like Figure 1 As shown, in a basic embodiment, the intelligent construction method for building load of the present invention includes the following steps.

[0051] Step S1: Based on the user's natural language input requirements and combined with the semantic vector knowledge base corresponding to the pre-built professional standard documents in the HVAC field, the requirements information is structured to generate structured retrieval information containing building parameters; building parameters refer to some or all of the basic building information, and the building parameters that can be provided vary from user to user.

[0052] This step is used to process and structure user input. This step can be automated by generating corresponding intelligent agents using large language models, or it can be obtained through interactive interfaces and processed by specialized software code.

[0053] When processing with intelligent agents, one agent (such as an integrity verification agent) first obtains the user's requirement description in natural language form, input in text, voice, or file format. For example, the user inputs "a standard office building in Nanjing, with an area of ​​10,000 square meters." This agent checks whether the input information meets the basic requirements for load calculation. If key parameters (such as running time or set temperature) are missing, the user is prompted to supplement them. Next, another agent (such as a structured information generation agent) transforms the unstructured natural language into standard structured retrieval information. During this process, if the user does not explicitly provide certain necessary parameters, the structured information generation agent will proactively search for reasonable default values ​​from a semantic vector knowledge base to fill them in, and specifically declare these assumptions in the structured output.

[0054] Step S2: Based on structured retrieval information, retrieve the semantic vector knowledge base and obtain load configuration parameters that match the building parameters from professional standard documents in the HVAC field;

[0055] This step primarily aims to obtain load configuration parameters. Since this invention targets ordinary users, the user-inputted requirements, when converted into structured search information, will inevitably lack specialized HVAC knowledge. Different users can input different information, making it difficult to obtain complete load configuration parameters solely from the user's input perspective. Therefore, it's necessary to convert the user's input requirements into structured search information that is similar in form to load configuration parameters. This information is then further supplemented using professional standard documents within the HVAC field to obtain detailed load configuration parameters that match the user's needs. Thus, while the structured search information shares the same content categories as the load configuration parameters, it may contain fewer content categories, and the specific content of each category may not be complete. In this way, load configuration parameters, which previously required specialized technical personnel, can now be obtained even by ordinary users. This step can also be accomplished using at least one corresponding intelligent agent (such as a retrieval configuration agent) formed by a large language model.

[0056] Step S3: Input the load configuration parameters and IDF file template into the agent built by the large language model to automatically build the IDF file;

[0057] This step is accomplished by an agent built using a large language model (such as an IDF-building agent). As mentioned in the background section, IDF files are text files that need to strictly adhere to specific syntax rules. Even if ordinary users know the load configuration parameters, it is difficult for them to write the corresponding IDF files. However, generating code snippets is one of the tasks that agents excel at. This invention allows the corresponding agent to learn by uploading an IDF file template. Then, based on the detailed load configuration parameters already obtained, the agent can easily complete the writing of the IDF file.

[0058] Step S4: Input the IDF file and its corresponding meteorological file into the load calculation engine to obtain hourly load data;

[0059] This step requires obtaining meteorological files. When the user inputs their requirements, the system can accurately obtain or guide the user to state the location corresponding to their requirements. Based on the location, the system can obtain the historical weather conditions for that location to obtain the corresponding meteorological files. After obtaining the IDF file and its corresponding meteorological files, the system can use a load calculation engine (such as EnergyPlus) to calculate hourly load data.

[0060] Step S5: Evaluate the hourly load data. If it is reasonable, output it; otherwise, return to step S3 to rebuild the IDF file.

[0061] Since the IDF file is generated by the corresponding agent, although the agent can handle the syntax rules well, the rationality of some of the logic settings in the IDF will be reflected in the hourly load data. Therefore, the corresponding hourly load data can be evaluated. If it is reasonable, it can be output. If it is unreasonable, it can be corrected according to the unreasonable parts to rebuild the IDF file.

[0062] Based on the above technical solution, this invention forms a semantic vector library from professional standard documents in the HVAC field. This allows ordinary users to supplement their basic needs based on these documents, and then convert the information into structured search information. However, this structured search information is merely an index. To obtain detailed load configuration parameters, the semantic vector knowledge base needs to be further searched to retrieve more detailed load configuration information from the professional standard documents in the HVAC field. After obtaining the load configuration information, an IDF file can be automatically constructed using an intelligent agent built from a large language model. This method saves professional users the work of manually correcting the syntax line by line and also allows ordinary users to create IDF files. The IDF file and its corresponding meteorological data are then input into the load calculation engine to obtain hourly load data, providing an accurate data foundation for users to select air conditioning models and optimize operating strategies.

[0063] like Figure 2 As shown, in one embodiment, the semantic vector knowledge base is formed using the following steps.

[0064] First, collect professional standard documents in the HVAC field, such as design specifications, design standards, or design manuals. Alternatively, you can collect all relevant professional standard documents in the HVAC field so that the semantic vector knowledge base can contain as much professional HVAC knowledge as possible.

[0065] Next, extract the table of contents and headings at all levels from the content of each professional standard document to form the logical structure corresponding to each professional standard document;

[0066] When dividing the content of a professional standard document into text blocks, the logical structure related to the paragraph content is attached to the corresponding paragraph as a sub-block of text;

[0067] All sub-block text is input into the embedding model and transformed into high-dimensional vector data;

[0068] All high-dimensional vector data and the mapping relationships between high-dimensional vector data and corresponding parts of professional standard documents in the HVAC field are stored in the semantic vector knowledge base.

[0069] Existing technologies also include methods that divide articles or books into sub-text blocks before inputting them into an embedding model to convert them into corresponding vector data. However, the difference between this invention and existing technologies lies in the fact that this invention extracts the logical structure of the document and then attaches the corresponding logical structure to the corresponding sub-text blocks when dividing the text into blocks. This makes the content of a sub-text block no longer isolated but can be combined with the context. As a result, when the embedding model is converted into vectors, it can obtain accurate high-dimensional vector data based on more accurate information. Furthermore, when retrieving information based on user needs, it can obtain accurate load configuration parameters. In contrast, the existing technology of simply dividing text into blocks, if applied to the HVAC field, will lead to the loss of interrelated information and make it difficult to obtain accurate load configuration parameters.

[0070] like Figure 3 As shown, in a further embodiment, when retrieving the semantic vector knowledge base, the structured retrieval information is first converted into a query vector. By calculating the cosine similarity between the query vector and the vectors in the semantic vector knowledge base, the m high-dimensional vector data with the highest cosine similarity are obtained from the semantic vector knowledge base. The load configuration parameters are formed by using the content of the professional standard documents corresponding to the m high-dimensional vector data.

[0071] The higher the cosine similarity between two vectors, the closer the two vectors are. Based on the previous embodiment, high-quality vectors can be obtained. That is, the vectors in the semantic vector knowledge base of this invention are of high quality. By querying the vectors, the m high-quality high-dimensional vector data with the highest similarity can be found. Then, based on these vector data, data in professional standard documents can be found, thereby forming accurate and detailed load configuration parameters that meet the user's needs.

[0072] The load configuration information of the present invention includes at least one of building foundation information, building envelope parameters, building internal disturbance information, and operating parameters.

[0073] Basic building information may include at least one of the following: the building's location, whether it is a residential or office building, its specific purpose, its area, and its floor height.

[0074] Building envelope parameters may include at least one of the following: the shape, area, material, or thickness of the building's windows, walls, roof, doors, and other structural elements.

[0075] Building disturbance information may include at least one of the following: the number of people in the building, their activity time, equipment load, etc.

[0076] Operating parameters may include at least one of the following: air conditioner's on / off time, operating temperature, etc.

[0077] Load configuration parameters vary depending on the location and the building. For a building, the more specific and precise the load configuration information, the more accurate the hourly load data will be. User demand information, or structured retrieval information derived from demand information, includes one or more of the above. For example, if a user's demand information is for a standard office building in Nanjing, then the user's demand information clearly only includes basic building information.

[0078] like Figure 4 As shown, in one embodiment, step S3 of the present invention includes the following.

[0079] Step S31: Determine whether the building basic information in the load configuration parameters contains detailed room-level geometric information;

[0080] If included, an IDF file will be generated based on the load configuration parameters and the IDF file template;

[0081] If not included, standard building templates are selected from the predefined building model library based on the building foundation information of the load configuration parameters. The size information of the standard building templates is adapted and scaled, and then combined with the load configuration parameters to generate an IDF file.

[0082] Step S32: Verify the integrity of the generated IDF file;

[0083] In step S33, if the verification fails, the reason for the failure is corrected and the file is regenerated, and the process returns to step S32 until the number of corrections reaches the preset number; if the verification succeeds, the construction of the IDF file is completed.

[0084] This step describes in detail a specific embodiment of the present invention for constructing an IDF file. In this embodiment, it is determined whether the load configuration parameters include room-level geometric information of the building. Room-level geometric information includes the dimensions of the room, the dimensions of the room's doors and windows, etc. If not, the corresponding room-level geometric information is generated based on the standard building template in order to further improve the content of the load configuration parameters, thereby obtaining a more accurate IDF file.

[0085] Since the IDF file is formed based on the agent in this step, verifying the IDF file can effectively improve the accuracy.

[0086] In a further embodiment, when evaluating hourly load data in S5, the hourly load data is transformed into structured load characteristic parameters, and the rationality of the load characteristic parameters is evaluated by combining relevant data retrieved from the semantic vector knowledge base.

[0087] This part can also be accomplished by a corresponding intelligent agent built from a large language model. Since ordinary users lack professional HVAC knowledge, when evaluating hourly load data, the intelligent agent can retrieve the corresponding data from professional standard documents in the HVAC field to evaluate the rationality of the load characteristic parameters. This can further verify the hourly load data and improve its accuracy.

[0088] In one embodiment, at least one of steps S1, S2, and S5 of the present invention is processed by a corresponding intelligent agent generated by a large language model. Generating a corresponding intelligent agent through a large language model allows ordinary users to easily obtain accurate hourly load data simply by providing their needs.

[0089] In a further embodiment, when steps S1 to S5 are all processed by the corresponding intelligent agents, steps S1 to S5 are processed through a workflow intelligent agent, which executes each step sequentially and calls the corresponding intelligent agent.

[0090] This workflow agent can not only automatically execute steps S1 to S5, but also call the corresponding agent, thus automating the entire process and simplifying user operations.

[0091] like Figure 5 As shown, the present invention also protects a building load intelligent construction system, which includes an input processing module, a knowledge base retrieval module, an intelligent agent engine module, and a simulation calculation module.

[0092] The input processing module obtains the user's demand information based on natural language input, and combines it with a pre-built semantic vector knowledge base corresponding to professional standard documents in the HVAC field to perform structured processing on the user's demand information and generate structured retrieval information containing building parameters.

[0093] The knowledge base retrieval module is used to retrieve semantic vector knowledge bases based on structured retrieval information, and based on the high-dimensional vector data obtained from the retrieval, and based on the mapping relationship between the high-dimensional vector data and the corresponding parts of professional standard documents in the HVAC field, to obtain load configuration parameters that match the building parameters from the corresponding professional standard documents.

[0094] The agent engine module generates different agents based on a large language model, performs structured processing, retrieves semantic vector knowledge bases, constructs IDF files, and evaluates hourly load data. The large language model typically generates agents by describing their settings or configurations and providing corresponding prompts to enable them to perform specific tasks.

[0095] For example, the prompts for the integrity verification agent could be designed as follows: "You are a senior HVAC engineer. Please check whether the information provided by the user (including any one or more of the user's input text, voice, or text) meets the building load calculation requirements based on the necessary content for load calculation in the load configuration parameters and the input parameters for load calculation. If the information is missing, please prompt the user for the specific content that needs to be supplemented."

[0096] Based on the above prompts, the agent generated by the large language model can effectively guide ordinary users to express their load requirements.

[0097] The simulation calculation module uses the load calculation engine to perform calculations based on the IDF file and its corresponding meteorological file to obtain hourly load data.

[0098] The intelligent building load calculation system of this invention, through various intelligent agents, semantic vector knowledge base, professional standard documents, load calculation engine, etc., can obtain relatively accurate hourly load data based on the needs of ordinary users, providing a strong data foundation for the selection, construction and intelligent control of air conditioning systems.

[0099] The intelligent agent of the building load intelligent construction system of the present invention also includes a workflow intelligent agent, which can connect various modules and call the corresponding intelligent agents to automatically complete the calculation of load data.

[0100] The following detailed description of a preferred embodiment of the present invention will provide a more detailed explanation of the technical concept of the invention.

[0101] This embodiment is a building load intelligent construction system based on Large Language Model (LLM). It utilizes the structured output capability of LLM to convert the user's natural language input (text, voice, file) into standard structured data, establishes and retrieves a semantic vector knowledge base in the HVAC field, and completes the automatic construction of building load.

[0102] A HVAC knowledge base needs to be built in advance, that is, a semantic vector knowledge base corresponding to professional standard documents. Figure 6 Taking the "GB50189 Standard for Energy-Saving Design of Public Buildings" as an example, the steps for building a semantic vector knowledge base are described.

[0103] First, divide the standard document into blocks according to the content of the clauses, and when dividing the blocks, attach the corresponding logical structure to the corresponding sub-block text;

[0104] Then, the embedding model is used to convert each sub-block text block into high-dimensional vector data;

[0105] Finally, the high-dimensional vector data and the mapping relationship between the high-dimensional vector data and the corresponding parts of the professional standard documents in the HVAC field are stored in the semantic vector knowledge base.

[0106] Conventional "fixed-length blocks" (such as blocks of 512 characters each) or simple "segment blocks" crudely sever the terms in the HVAC field from their context, resulting in the complete semantics being broken down, thus failing to obtain high-quality vectors. However, this invention can obtain higher-quality vectors.

[0107] A predefined building model library is also required. This library contains standard building models, which include detailed geometric information of each room in a building. In other words, this invention predefines standardized "building prototype IDF templates" for different types of buildings (such as offices, hotels, and retail spaces) and different building sizes.

[0108] For example, a "medium-sized office building prototype" is predefined as follows: 10 floors, each divided into 5 zones – 4 outer zones and 1 inner zone. This prototype template has already defined the relative relationships and names of the geometry (walls, windows, floors, roof) of all rooms, but its specific parameters (such as the building envelope heat transfer system, room area, occupants, fresh air supply, and lighting) are pending.

[0109] After the above preparations are completed, the system of this invention can be used. (Refer to...) Figure 7 .

[0110] The first step is for the user to input their requirements.

[0111] The information input by users typically includes basic building information, and the input methods include at least one of text input, voice input, and document input.

[0112] If the input is voice, the corresponding intelligent system built by the large language model will convert the voice into text based on the voice recognition function;

[0113] If the input is a file (xlsx, docx, pdf, etc.), the corresponding AI built by the large language model will convert the file into text based on the file recognition function;

[0114] If the input is text, proceed to the next step.

[0115] The second step is to perform integrity verification on the user input.

[0116] The intelligent agent constructed by the large language model verifies whether the information required for building load calculation in the demand information is complete. If necessary information is missing, it prompts the user to supplement it; if the information is complete, it generates structured retrieval information based on the user's input.

[0117] The first and second steps can be accomplished by different agents. For example, an integrity verification agent can guide user input and perform integrity verification, while a structured information generation agent can generate structured retrieval information.

[0118] Structured search information is used for subsequent searches. Structured search information may include building location, building area, user habits (set temperature, running time), building purpose, etc.

[0119] In one specific embodiment, an example of structured output is shown below.

[0120] {

[0121] "location": "Nanjing", / / Location of the building, Nanjing

[0122] "area": ​​"10000", / / Building area, 10000 m² 2

[0123] “usage”: “office”, / / Building use, office

[0124] “operating_date”: “06 / 01-10 / 15”, / / Operating dates, from June 1st to October 15th

[0125] ...

[0126] }

[0127] While large language models have the ability to transform unstructured natural language into structured retrieval data, they present some challenges in the HVAC (Heating, Ventilation, and Air Conditioning) field, which demands high accuracy and has low error tolerance.

[0128] Users may not be able to provide all the information for load calculation directly, i.e., they may not be able to provide detailed information on load configuration parameters, and some of the information provided by the user may be implicit. For example, if a user only says "a standard office building in Nanjing", then the user has assumed the operating time (e.g., weekdays 8:00-18:00), personnel density (e.g., 0.1 people / m²), and set temperature (e.g., 26℃).

[0129] This embodiment guides users through an agent formed by a big oracle model. When the user does not explicitly provide necessary information (e.g., setting a temperature), the agent is instructed not to leave it blank, but to actively search for the corresponding high-dimensional vector data from the semantic vector knowledge base. Based on the high-dimensional vector data, it finds a reasonable default value from professional standard documents (e.g., office building standard 26℃), fills in this value, and declares it in the "assumptions" field. By marking the "assumptions" field, when the specific load configuration parameters are obtained, the agent can provide prompts to the user based on this mark, avoiding discrepancies between the found default value and the user's actual needs. This allows the user to make modifications that better suit their actual requirements based on the default value prompts.

[0130] For example, the structured retrieval data output by the agent can be seen in the following example.

[0131] {

[0132] "location": "Nanjing",

[0133] "area": ​​"10000",

[0134] "usage": "office",

[0135] "T_set_cooling": 26, ...

[0137] "assumptions": [

[0138] "T_set_cooling: 26°C (User did not specify, using 'GB 50189'standard for office.)" ]

[0140] }

[0141] The structured search information currently provided only displays a few fields of the email address. In reality, the structured search information should at least include the core parameters sufficient to start a load calculation engine (such as EnergyPlus), including but not limited to the following:

[0142] Basic building information: location (where the building is located), area (building area), height (building floor height), usage (purpose, such as office, hotel, hospital, etc.);

[0143] Building envelope parameters (if missing, they can be generated from professional standard documents by searching the semantic vector knowledge base): orientation (building orientation), window_to_wall_ratio (window-to-wall ratio), u_value_wall (external wall heat transfer coefficient), u_value_roof (roof heat transfer coefficient), shgc_window (window SHGC value), etc.

[0144] Building disturbance information (if missing, it can be generated from professional standard documents by searching the semantic vector knowledge base): personnel_density, lighting_load_density, equipment_load_density.

[0145] Operating parameters (if missing, they can be generated from professional standard documents by searching the semantic vector knowledge base): operating_schedule (operating schedule, such as "weekdays 9-18 o'clock"), T_set_cooling (cooling set temperature), T_set_heating (heating set temperature), fresh_air_rate (fresh air volume), etc.

[0146] When generating an intelligent agent using structured information from a large language model, the corresponding prompt could be designed as: "You are a building load simulation expert. Please output structured retrieval information based on <user input text> and <data style of structured retrieval information>." When the user does not explicitly provide necessary information (e.g., setting a temperature), a reasonable default value (e.g., office building standard 26℃) should be actively searched from the semantic vector knowledge base, and then this value should be filled in and declared in the "assumptions" field.

[0147] Structured retrieval information is transformed into high-dimensional vector data through an embedding model. Then, a semantic vector knowledge base is queried to obtain the m high-dimensional vector data with the highest cosine similarity. Based on these high-dimensional vector data, m matching document data are then obtained from professional standard documents.

[0148] For example, the following document data can be retrieved.

[0149] I. For general offices, the personnel density is 0.1 people / m². 2 Selected personnel load: 14W / m 2 Lighting load 50 W / m 2 Fresh air volume 25 m³ 3 / (person·h), fresh air load 27 W / m 2 —《Air Conditioning Design for Civil Buildings》;

[0150] II. Reference index for office building cooling load: 70~120 W / m 2 —《Air Conditioning Design for Civil Buildings》;

[0151] Ⅲ. The air conditioning system of office buildings shall be operated from 7:00 to 18:00 on weekdays and shut down on holidays – “Air Conditioning Design for Civil Buildings”;

[0152] IV. The indoor temperature setting for air conditioning in office buildings and teaching buildings shall be 26℃ during operation – GB 50189-2015 Standard for Energy Conservation Design of Public Buildings;

[0153] V. In hot-summer and cold-winter regions, the heat transfer coefficient of the exterior wall cladding structure is ≤1.0 W / (m²). 2 ·K)——《GB 50189-2015 Standard for Energy-Saving Design of Public Buildings》.

[0154] The third step is to obtain the load configuration parameters.

[0155] Based on the structured retrieval information corresponding to the user's input requirements, the semantic vector knowledge base is searched, and load configuration parameters matching the building parameters are obtained from professional standard documents in the HVAC field.

[0156] For example, if the user does not provide the air conditioner's operating time and set temperature for a certain office in the information they enter, then according to the above search results III and IV, the air conditioner's operating time will be selected as 7:00-18:00 on weekdays and off on holidays, with a temperature set to 26℃; if the user's information includes the operating time of room xx as 9:00-17:00 and the temperature as 24℃, then the settings will be based on the user's input values.

[0157] This part of the work can be accomplished by generating intelligent agents based on building loads using large language models.

[0158] The fourth step is to build the IDF file.

[0159] The IDF file is constructed by inputting the load configuration parameters and the IDF file template into the agent built by the large language model. The corresponding agent maintains the building envelope (shape, area, material, and thickness of windows, walls, roofs, doors, etc.), building disturbance information (personnel load, equipment load, etc.), and operating parameters (start-up and shutdown time, operating temperature, etc.) of each room in turn, and automatically constructs the IDF file.

[0160] There are two scenarios when building an IDF file.

[0161] First, the user input in the first step provides complete room geometry information of the building. The agent of the large language model can directly parse this information and then maintain the enclosure structure (shape, area, material, thickness of windows, walls, roof, doors, etc.), internal disturbance information (personnel, load, equipment load, etc.), and operating parameters (start-up and shutdown time, operating temperature, etc.) of each room based on the user input or the retrieval results of the semantic vector knowledge base, thus completing the creation of the IDF file.

[0162] Secondly, users only provide overall building information, not specific room details. This invention will utilize a pre-established building model library to create the IDF file.

[0163] Assuming the user input only specifies a medium-sized office building in Nanjing, without providing other room geometry information, this invention will select an IDF template of a medium-sized office building from the building model library. The corresponding AI will perform a geometric scaling operation. For example, when it reads the IDF template of the medium-sized office building, it finds that the total area of ​​the building in the template is 5000m² (hypothetically), while the user input is 10000m². The corresponding AI will calculate a scaling factor of 2.0 and apply this scaling factor to all zones in the prototype, thereby automatically completing the allocation of the 10000m² area.

[0164] When the agent traverses the IDF file template, it iterates through the objects one by one. When it encounters an envelope object, it uses the semantic vector knowledge base and patent standard documents to find that the external wall heat transfer coefficient is ≤1.0, and then fills this 1.0 into the corresponding field of the object. When it encounters a Lights object, it uses the semantic vector knowledge base and patent standard documents to find that the lighting load is 50 W / m², and fills this value in. When it encounters a People object, it uses the semantic vector knowledge base and patent standard documents to find that the personnel load is 14 W / m², and fills it in. When it encounters a Schedule:Compact object, it checks whether there is any user input (such as "9:00~17:00"). If there is, it uses the user input; otherwise, it uses the semantic vector knowledge base and patent standard documents to retrieve "weekdays 7:00~18:00" and "26℃" to dynamically generate Schedule data.

[0165] The prompts for the agent used to automatically build IDF files could be: "You are an IDF generation expert. Please generate an IDF file that meets the EnergyPlus computing requirements based on the user's input information." "Please confirm whether the user has provided detailed room information or overall building information." "If detailed room information is provided, please build the IDF file based on the user-provided information and the search results from the semantic vector knowledge base. If only overall building information is available, please populate the IDF template based on the building prototype."

[0166] After generating the IDF file, it is necessary to verify the IDF file.

[0167] The integrity of the IDF file generated by the agent corresponding to the large language model can be verified by comparing it with the standard template of the IDF file. If the verification fails, an error message is returned to the large language model so that the agent can regenerate the IDF file. If the number of failures exceeds x, for example, x=5, the current workload generation is terminated.

[0168] IDF file verification involves two technical methods.

[0169] The first is conventional syntax checking, and the second is logic checking based on a large language model.

[0170] For routine syntax validation, EnergyPlus can automatically check whether the generated IDF file conforms to the syntax rules. For example, is the Zone object missing a Name field? Is the Area field filled with text instead of numbers? Once EnergyPlus detects errors in the IDF file, it returns the error information to the large language model, making it easier for the corresponding agent to regenerate the IDF file.

[0171] LLM-based logical validation. An IDF file might pass syntax validation but be logically flawed. For example, a Window object might be defined but not properly "mounted" to any Wall, rendering it useless in the simulation; or a Schedule might be defined but never invoked by any Lights or People objects.

[0172] This part of the logic verification can also be done by the intelligent agent corresponding to the large language model.

[0173] This agent can be generated based on the following prompt: You are an EnergyPlus simulation expert. Please review the IDF file for logical inconsistencies. For example, an object may be defined but not used, or two conflicting temperature control methods may be set for a single area. If logical problems are found, please return to the relevant section.

[0174] The fifth step is to calculate the hourly load data.

[0175] Based on the user's input requirements, the system retrieves the building's location and then obtains the corresponding meteorological file, such as the meteorological file for Nanjing. The meteorological file, also called an EPW (EnergyPlus Weather) file, is the standard meteorological data format for EnergyPlus, containing 8760 hours of meteorological data for the entire year (dry and wet bulb temperatures, solar radiation, wind speed, wind direction, etc.).

[0176] The IDF file obtained in step four, along with the meteorological file mentioned above, are input into the load calculation engine to calculate hourly load data.

[0177] Step 6: Result evaluation and verification.

[0178] The agent generated by the large language model can generate structured feature parameters from the obtained hourly load data. Examples of feature parameters are shown below.

[0179] {

[0180] "location": "Nanjing", / / Location of the building, Nanjing

[0181] "area": ​​10000, / / Building area, 10000 m² 2

[0182] "max_cooling_load": 120, / / Maximum cooling load per unit area, W / m2

[0183] "max_heatingling_load": 60, / / Maximum heating load per unit area, kW / m² 2

[0184] "max_human_load": 80.0, / / Maximum human load per unit area, W / m2

[0185] ...

[0186] }

[0187] Then, the obtained load feature parameters and the retrieval data most relevant to these load feature parameters are input into the large language model. The retrieval data here is also obtained by searching the semantic vector knowledge base.

[0188] The large language model corresponds to an intelligent agent that automatically judges whether the load result is reasonable. For example, the maximum cooling load per unit area in the calculation result is 120W / m. 2 The large language model, combined with the retrieved "office building cooling load reference index 70~120 W / m", 2—Analysis of the "Air Conditioning Design for Civil Buildings" shows that this indicator is reasonable. The maximum occupancy load per unit area is 80W / m². 2 This is consistent with the recommended value of 14 W / m. 2 If the deviation is too large and the verification fails, the evaluation result and the reason for failure will be returned to the building load generation agent, and step three will be executed again. If the number of failures is greater than y, for example, y=5, then the current load generation will end.

[0189] Determining the reasonableness of the obtained load results can also be accomplished through a corresponding intelligent agent. The corresponding intelligent agent generated by the large language model generates a "question" for retrieval based on load feature parameters (such as "max_cooling_load": 120) and building information (such as "usage": "office"), for example: "What is the reference index for cooling load per unit area of ​​an office building?" This question is then transformed into a question vector. By comparing the cosine similarity between vectors, the IDs or indices of the m most relevant high-dimensional vectors to the question vector are returned from the semantic vector knowledge base. These IDs point to corresponding content in at least one professional standard document. For example, it is possible to retrieve "Reference index for cooling load of office buildings: 70~120 W / m²". 2 —《Air Conditioning Design for Civil Buildings》.

[0190] This part of the intelligent agent can be generated based on the following prompt: You are an HVAC expert. Please evaluate the rationality of the load calculation results by combining the retrieval from the semantic vector knowledge base. Building Information: Nanjing City Office Building, Area 10000m² 2 ...>; Simulation results: Maximum cooling load per unit area: 120W / m 2 The maximum personnel load per unit area is 80W / m². 2 If this is unreasonable, please provide a reason and re-invoke the <Building Load Generation Agent> to generate the building load.

[0191] This invention does not rely on a specific large language model (such as GPT, Qwen, Deepseek, etc.), but instead constructs multiple intelligent agents based on prompt word engineering to complete workflows such as user input information verification, generation of structured retrieval information, retrieval of semantic vector knowledge base, generation and verification of IDF file, evaluation of load calculation results, and output of final results.

[0192] The entire workflow can also be completed by an intelligent agent, for example, using the following prompts: You are an HVAC expert who needs to perform building load calculations based on user input. You need to complete the following tasks in sequence: 1. Validate user input, <call intelligent agent…>; 2. Generate structured retrieval information, <call intelligent agent…>; 3. Retrieve semantic vector knowledge base and generate IDF file, <call intelligent agent…>; 4. Validate IDF file content and call EnergyPlus for calculation, <call intelligent agent…>; 5. Verify load calculation results, <call intelligent agent…>.

[0193] This prompt allows an agent to call upon different agents at different stages of the work to complete the corresponding tasks, thereby automating the entire process. Even ordinary users without any professional knowledge can accurately obtain the hourly load data of the building.

[0194] Finally, the corresponding intelligent agent can be used to output hourly load data in the form required by the user, such as data tables, line graphs, and load distribution frequency graphs.

[0195] This invention also protects a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent construction method for building loads of any of the above-described technical solutions.

[0196] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligently constructing building loads, characterized in that, include: Step S1: Based on the user's natural language input requirements and combined with the pre-built semantic vector knowledge base corresponding to professional standard documents in the HVAC field, the requirements information is processed in a structured manner to generate structured retrieval information containing building parameters. Step S2: Based on the structured retrieval information, retrieve the semantic vector knowledge base and obtain the load configuration parameters that match the building parameters from the professional standard documents in the HVAC field; Step S3: Input the load configuration parameters and IDF file template into the agent constructed by the large language model to automatically construct the IDF file; Step S4: Input the IDF file and its corresponding meteorological file into the load calculation engine to obtain hourly load data; Step S5: Evaluate the hourly load data. If it is reasonable, output it; otherwise, return to step S3 to rebuild the IDF file.

2. The intelligent construction method for building load according to claim 1, characterized in that, The semantic vector knowledge base is formed using the following steps: Collect professional standard documents in the HVAC field; Extract the table of contents and headings from the content of the professional standard document to form the logical structure of the professional standard document; When dividing the content of the professional standard document into text blocks, the logical structure related to the paragraph content is attached to the corresponding paragraph as a sub-block text. All sub-block text is input into the embedding model and transformed into high-dimensional vector data; All high-dimensional vector data and the mapping relationship between high-dimensional vector data and the corresponding parts of the professional standard documents in the HVAC field are stored in the semantic vector knowledge base.

3. The intelligent construction method for building load according to claim 1, characterized in that, When retrieving the semantic vector knowledge base, the structured retrieval information is first converted into a query vector. By calculating the cosine similarity between the query vector and the vectors in the semantic vector knowledge base, the m high-dimensional vector data with the highest cosine similarity are obtained from the vector knowledge base. The load configuration parameters are formed by using the content of the professional standard document corresponding to m high-dimensional vector data.

4. The intelligent construction method for building load according to claim 1, characterized in that, The load configuration information includes at least one of the following: building foundation information, building envelope parameters, building internal disturbance information, and operating parameters.

5. The intelligent construction method for building load according to claim 4, characterized in that, Step S3 includes: Step S31: Determine whether the building basic information in the load configuration parameters contains detailed room-level geometric information; If included, the IDF file is generated based on the load configuration parameters and the IDF file template; If not included, based on the building foundation information of the load configuration parameters, standard building templates are selected from the predefined building model library, the size information of the standard building templates is adapted and scaled, and then combined with the load configuration parameters to generate the IDF file; Step S32: Verify the integrity of the generated IDF file; In step S33, if the verification fails, the reason for the failure is corrected and the file is regenerated, and the process returns to step S32 until the number of corrections reaches the preset number; if the verification succeeds, the construction of the IDF file is completed.

6. The intelligent construction method for building load according to claim 1, characterized in that, In step S5, when evaluating the hourly load data, the hourly load data is converted into structured load characteristic parameters, and the rationality of the load characteristic parameters is evaluated in conjunction with relevant data retrieved from the semantic vector knowledge base.

7. The intelligent construction method for building load as described in claim 1, characterized in that, At least one of steps S1, S2, S4 and S5 is processed by the corresponding intelligent agent generated by the large language model.

8. The intelligent construction method for building load as described in claim 7, characterized in that, When steps S1 to S5 are all processed by the corresponding intelligent agents, steps S1 to S5 are processed through a workflow intelligent agent, which executes each step sequentially and calls the corresponding intelligent agent.

9. A building load intelligent erection system, characterized in that, The system includes: The input processing module acquires the user's demand information based on natural language input, and combines it with a pre-built semantic vector knowledge base corresponding to professional standard documents in the HVAC field to perform structured processing on the demand information and generate structured retrieval information containing building parameters. The knowledge base retrieval module is used to retrieve the semantic vector knowledge base based on the structured retrieval information, and to obtain load configuration parameters that match the building parameters from the corresponding professional standard documents based on the retrieved high-dimensional vector data. The intelligent agent engine module generates different intelligent agents based on a large language model for structured processing, retrieves the semantic vector knowledge base, constructs IDF files, and evaluates hourly load data. The simulation calculation module uses a load calculation engine to perform calculations based on the IDF file and its corresponding meteorological file to obtain hourly load data.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent construction method for building loads according to any one of claims 1 to 8.

Images