Heating secondary network personalized control method based on large model and multi-modal image

Abstract

The application discloses a heat supply secondary network personalized regulation and control method based on a large model and multi-modal portraits, and comprises the following steps: acquiring multi-modal data of each heat user of a heat supply system secondary network, performing feature extraction and importance analysis by using a large model, and then constructing each heat user portrait of the secondary network; performing information extraction on each household operation mode of the heat supply system secondary network, combining the heat user portrait selected for the household operation mode, and generating an operation mode portrait of the household operation mode by using the large model; inputting each heat user portrait and the operation mode portrait of the secondary network into the large model for matching analysis, and outputting personalized operation mode recommendation results and recommendation rhetoric for each heat user; after recommending the selected operation mode to the heat user, training and optimizing a household regulation and control large model of the heat supply system secondary network according to the operation mode selected by the heat user, the heat user portrait and the established household regulation and control knowledge graph, combining hydraulic balance mechanism knowledge of the heat supply network, and outputting a household regulation and control strategy of the secondary network.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of secondary network regulation of heating systems, and particularly relates to a heating secondary network individualized regulation method based on a large model and multi-modal portraits. BACKGROUND

[0002] The main task of the secondary network of the heating system is to distribute the fluid in the primary pipe network to specific areas or heat users, and the operation of the secondary pipe network is directly related to the heat demand of the heat users, enterprises and commercial areas. Different types of heat users have different heat behavior and preference attributes. For example, some residential heat users have no one at home during the day and use heat after work. The residents can choose to reduce indoor temperature and reduce heat consumption during the period when there is no one at home during the day. For mall heat users, the heat consumption is related to the business hours and the flow density. Therefore, it is necessary to establish portraits of each heat user of the secondary network to represent the heat behavior and preferences of different heat users, so as to facilitate individualized regulation.

[0003] With the application of artificial intelligence technology in the field of intelligent heating, the optimization and regulation of the heating system have become more intelligent. However, with the increase in the amount of heat user data of the heating system, the expansion of the system operation scale and the complexity and diversity of the operation data, how to efficiently and accurately process and analyze these data to generate more optimal regulation strategies and meet the heat demand of different heat users has become a difficult problem. Although machine learning and deep learning technologies have been applied to the optimization and regulation of heating, these methods still have certain limitations, lack individualization, are difficult to reveal the complex correlation between multi-modal massive heterogeneous data, cannot output individualized regulation strategies for each heat user, and cannot fully utilize the rich structured knowledge contained in the regulation process of the secondary network of the heating system.

[0004] Based on the above technical problems, a new heating secondary network individualized regulation method based on a large model and multi-modal portraits needs to be designed. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a heating secondary network individualized regulation method based on a large model and multi-modal portraits. The method can use a large model and multi-modal data to establish portraits and operation mode portraits of each heat user of the secondary network, perform matching analysis, output individualized operation mode recommendation results and recommendation dialogues for heat users, clearly explain the recommendation reasons and mode advantages to the users, improve the heat user experience and the acceptance of the operation mode, establish a household regulation knowledge graph, fully utilize the structured knowledge of the regulation, establish a household regulation large model, mine the complex correlation between data, make the model continuously adapt to the changes in the actual situation, and improve the accuracy and effectiveness of the regulation strategy.

[0006] To solve the above technical problems, the technical scheme of the present application is:

[0007] The application provides a heating secondary network personalized regulation method based on a large model and multi-modal portraits, which comprises the following steps:

[0008] S1, multi-modal data of each heat user of a heating system secondary network is acquired, and after feature extraction and importance analysis are performed by using a large model, a portrait of each heat user of the secondary network is constructed;

[0009] S2, information extraction is performed on each household operation mode of the heating system secondary network, the portrait of the heat user selecting the household operation mode is combined, and a large model is used to generate an operation mode portrait of the household operation mode;

[0010] S3, the portrait of each heat user of the secondary network and the operation mode portrait are input into the large model for matching analysis, and a personalized operation mode recommendation result and a recommendation speech are output for each heat user;

[0011] S4, a household regulation knowledge graph is established according to household heat metering data, regulation data and regulation knowledge corresponding to the operation mode of each heat user;

[0012] S5, according to the operation mode recommendation result and the recommendation speech of each heat user, the heat user is recommended to select an operation mode, and according to the operation mode selected by the heat user, the portrait of the heat user and the household regulation knowledge graph, combined with the knowledge of the hydraulic balance mechanism of the heat network, a household regulation large model of the heating system secondary network is trained and optimized, and a household regulation strategy of the secondary network is output.

[0013] Further, the S1 specifically comprises:

[0014] The multi-modal data of each heat user of the heating system secondary network comprises text data, voice data and image data; the text data comprises basic information of a heating user type, a user age, a house orientation, a house area, a floor on which the user is located, a user work and rest time, a heat consumption in each period, a flexible heat load response information, a complaint information, a maintenance work order information and a preferred operation mode; the voice data comprises voice evaluation information of a heating effect between a heat user and a heating enterprise, fault description voice information and special heat demand voice information; the image data comprises indoor heating equipment state images, outdoor heating pipeline images, house appearance images, house layout images and house flow personnel images; the heating user type comprises domestic heating and industrial and commercial heating;

[0015] The obtained multi-modal data is pre-processed and multi-modal fused to form a structured data set, and is input into a pre-trained large model. The structured data set and a prompt are used as inputs by combining the large model and the prompt. Then, the attention mechanism is used to extract features of the data, including features of text, voice and image. The basic features of the hot user and the user's hot behavior features are obtained. After the feature importance analysis is performed by using the feature importance evaluation method, the secondary network of each hot user portrait including multi- portrait dimensions is output.

[0016] Further, the prompt for constructing the secondary network of each hot user portrait includes:

[0017] Role setting: It is clear that the large model is an expert in the heating field during the dialogue process, and it guides the large model to output the behavior mode and response mode of the hot user portrait.

[0018] Demand description: The task demand completed by the large model is described, including feature extraction and feature importance analysis.

[0019] Multi-portrait dimension definition: It includes user basic feature definition and user hot behavior definition, which guides the large model to correctly understand the specific meaning. The user basic feature definition is the static information of the hot user, and the user hot behavior definition is the dynamic information of the hot user, including hot user work and rest behavior information, hot user hot quantity preference information in each period, flexible response information, running mode preference information and hot feedback information.

[0020] Further, the secondary network of each hot user portrait is defined as:

[0021] P i =LLMs(D i ,R);

[0022] D i is the structured data set of the i-th hot user; R is the prompt for constructing the hot user portrait; P i is the portrait of the i-th hot user; LLMs() is the large model used to construct the hot user portrait.

[0023] Further, in the S2, the running mode of each household in the secondary network of the heating system specifically includes:

[0024] Normal running mode: In any period of the heating season, the indoor temperature of the user reaches the preset temperature.

[0025] Energy-saving running mode: It has a time period division mechanism and a customized temperature selection mechanism, which can select a smaller indoor temperature than the conventional preset temperature for energy saving at a fixed time period according to the user's hot behavior, and give economic compensation at the same time.

[0026] Intelligent adaptive mode: Real-time monitoring of indoor temperature, room activity time, and number of people moving around, intelligently and automatically adjusting heating supply;

[0027] Economic optimization mode: Based on energy price fluctuation information, it provides the optimal operating plan while meeting basic comfort requirements, thereby reducing heating costs.

[0028] Furthermore, S2 specifically includes:

[0029] Obtain the definition of each household's operating mode in the secondary network of the heating system, the heat user information of each household's operating mode selected in history, and the operating parameters of the operating mode;

[0030] Set prompts to guide the generation of operational mode profiles from large models, and clarify their function in describing the generation of operational mode profiles;

[0031] The name of the operating mode, the description of its characteristics, the correlation information between each operating mode and the hot user profile, and the operating parameters of the operating mode are used as operating mode data. These are then combined with prompt words input into the large model to generate an operating mode profile, represented as follows:

[0032] P j =LLMs′(L j ,M);

[0033] L j The runtime mode data is for the j-th runtime mode; M is the prompt word used when generating the runtime mode profile; P j The image is for the j-th operating mode; LLMs′ is the large model used when generating the operating mode image.

[0034] Furthermore, the large model generates a profile of the operational mode, including:

[0035] The input sequence data is divided into words using a word segmenter, and each word is mapped to a low-dimensional vector space to obtain word embeddings. At the same time, positional encoding is added to allow the large model to learn the positional information of the words in the sequence.

[0036] The multi-layer transformer encoder structure based on a large model utilizes the multi-head self-attention mechanism of each encoder to enable the large model to learn the relationship between lexical units in different representation subspaces and capture semantic information. Then, a feedforward neural network is used to perform information transformation and feature extraction on the output of the multi-head self-attention. Different abstract representations of the input data are learned at different levels. By continuously stacking, feature representations containing rich information of the input data are obtained.

[0037] In the output layer, a linear transformation is used to map the feature representation to a dimension equal to the size of the vocabulary, obtaining the predicted score of each token; using a decoding strategy, the token with the highest probability is selected as the next generated token according to the predicted score, and the text content of the operation mode image is generated step by step.

[0038] Further, the S3 specifically comprises:

[0039] Integrating the secondary network of each hot user portrait and the operation mode portrait into an input data set;

[0040] Setting prompt words for guiding the large model to output personalized operation mode recommendation results and recommendation dialogues, and clearly defining the role of the large model, task demand description and thought chain construction when recommending operation mode and dialogue;

[0041] Inputting the input data set and the prompt words into the large model, and the large model uses the internal attention mechanism and neural network layer to calculate the similarity between the hot user portrait features and the operation mode portrait features;

[0042] The large model selects the operation mode with the highest similarity as the personalized recommendation result output of the hot user according to the similarity calculation result;

[0043] The large model outputs differentiated recommendation dialogues for different hot users according to the output personalized operation mode recommendation result and in combination with the pre-established dialogue corpus.

[0044] Further, the S4 specifically comprises:

[0045] Obtaining the heat metering data of each hot user, including heat consumption, flow and temperature in different time periods;

[0046] Obtaining the sub-household regulation data of each hot user, including operation mode switching time and adjustment parameters of regulation equipment;

[0047] Obtaining the regulation knowledge corresponding to the operation mode, including the regulation mechanism and target temperature range of the regulation equipment under different operation modes;

[0048] According to the sub-household heat metering data, regulation data and regulation knowledge corresponding to the operation mode of each hot user, deep learning method is used for entity, relationship and attribute extraction; the relationship extraction includes the relationship between the hot user and the regulation equipment, the relationship between the hot user and the operation mode, and the relationship between the regulation equipment and the operation mode;

[0049] According to the extracted entity, relationship and attribute, a household regulation knowledge graph is established, including: determining the structure of the secondary network household regulation knowledge graph, adopting a directed graph form, taking nodes to represent entities and edges to represent the relationship between entities; adding entities as nodes to the knowledge graph, and creating corresponding edges according to the relationship, and adding corresponding attribute information to each node and edge.

[0050] Further, the S5 specifically includes:

[0051] The running mode data selected by the hot user is associated and integrated with the hot user portrait data;

[0052] Knowledge information related to the hot user running mode and the regulation equipment is extracted from the household regulation knowledge graph, and is converted into a vector form suitable for model input, and the processed knowledge graph vector is fused with the hot user data;

[0053] The hot network hydraulic balance mechanism knowledge is converted into a mathematical model and a constraint condition, and the parameters related to the hot network hydraulic balance are extracted;

[0054] The training target is defined: taking generating the optimal secondary network household regulation strategy as the target, setting the training objective function according to the running mode requirement of the hot user, minimizing the energy consumption of the hot network and ensuring the hydraulic balance of the hot network;

[0055] The associated data of the running mode and the hot user portrait, the knowledge graph data and the hot network hydraulic balance data are integrated as training samples and input into the large model for training, to generate a secondary network household regulation large model of the heating system, and the parameters of the large model are continuously adjusted, so that the large model can learn the relationship between the hot user running mode, the hot user portrait, the household regulation knowledge graph and the hot network hydraulic balance;

[0056] After training and optimization, the large model inputs new hot user running mode, hot user portrait and real-time running data of the secondary network, and the large model generates a secondary network household regulation strategy according to the learned knowledge and relationship, including the adjustment parameter of the pre-household regulation equipment and the switching time of the running mode.

[0057] The beneficial effects of the present application are:

[0058] (1) The present application constructs a hot user portrait by acquiring multi-modal data, using a large model for feature extraction and importance analysis, can use multi-modal data to comprehensively depict the characteristics of hot users and the characteristics of hot users, and can accurately extract key features from massive multi-modal data by using the powerful feature extraction capability of the large model, filter out redundant information, and determine the influence degree of each feature on the user's hot behavior through importance analysis, and the constructed hot user portrait provides user basic information for subsequent personalized running mode recommendation and regulation;

[0059] (2) By extracting information on individual household operation modes and combining it with hot user profiles, this invention generates operation mode profiles using a large model. It can sort out the key information and characteristics of each individual household operation mode in detail, and combine the operation mode with the hot user profile of the selected mode. This makes the operation mode profile more in line with actual user needs and applicable scenarios, and facilitates understanding of the user group characteristics applicable to each mode. It also helps to recommend suitable operation modes to hot users more accurately.

[0060] (3) This invention inputs hot user profiles and operation mode profiles into a large model for matching analysis, and outputs personalized recommendation results and scripts. It can utilize the powerful data analysis and matching capabilities of the large model to accurately find the most suitable operation mode for each hot user based on hot user profiles and operation mode profiles, thereby improving the accuracy of recommendations. In addition, according to the characteristics and needs of different hot users, personalized recommendation scripts are output to clearly explain the reasons for the recommendation and the advantages of the mode to the user, thereby improving the user experience and acceptance of the operation mode.

[0061] (4) This invention establishes a household control knowledge graph based on household heat metering data, control data, and control knowledge. This integrates various types of data and control knowledge to form a structured knowledge graph, which facilitates unified management and query of household control-related information. In addition, it provides rich knowledge support for household control in the secondary network of the heating system, which helps the large model to make more scientific and reasonable decisions by comprehensively considering various factors when generating control strategies. Furthermore, through the correlation analysis of the knowledge graph, the potential relationship between household heat metering data, control data, and control knowledge can be discovered, providing a basis for optimizing control strategies.

[0062] (5) This invention guides users to select a mode based on recommendation results and wording, trains and optimizes the large-scale control model by combining multiple information sources, and outputs control strategies. It can train and optimize the large-scale control model based on multiple information sources such as the user's selected operating mode, heat user profile, and household control knowledge graph, combined with the knowledge of the hydraulic balance mechanism of the heating network, so that the model can continuously adapt to changes in actual conditions and improve the accuracy and effectiveness of the control strategies. In addition, considering the knowledge of the hydraulic balance mechanism of the heating network, it ensures that the output secondary network household control strategies can maintain the hydraulic balance of the heating network while meeting the personalized needs of users, and ensure the stable operation of the entire heating system.

[0063] Other features and advantages will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings needed in the specific embodiment or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0066] Figure 1 A flow chart of the method for personalized regulation and control of the secondary network of the heating system based on a large model and multi-modal portraits according to the present application;

[0067] Figure 2 A principle block diagram of the personalized regulation and control of the secondary network of the heating system based on a large model and multi-modal portraits according to the present application;

[0068] Figure 3 A flow chart of the method for outputting personalized operation mode recommendation results and recommendation dialogues according to the present application;

[0069] Figure 4 A flow chart of the method for establishing a household regulation and control knowledge graph according to the present application;

[0070] Figure 5 A flow chart of the method for training and optimizing a large model for household regulation and control of the secondary network of the heating system according to the present application. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely in combination with the drawings. Obviously, the described embodiments are some embodiments of the present application, but not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0072] As shown in Figure 1 and Figure 2 , the present embodiment 1 provides a method for personalized regulation and control of the secondary network of the heating system based on a large model and multi-modal portraits, which comprises:

[0073] S1, multi-modal data of each heat user of the secondary network of the heating system is obtained, and after feature extraction and importance analysis by using a large model, a portrait of each heat user of the secondary network is constructed;

[0074] S2, information extraction is performed on each household operation mode of the secondary network of the heating system, the heat user portrait of the household operation mode is selected, and an operation mode portrait of the household operation mode is generated by using a large model;

[0075] S3, inputting the secondary network of each thermal user portrait and operation mode portrait into the large model for matching analysis, outputting personalized operation mode recommendation results and recommendation language for each thermal user;

[0076] S4, establishing a household regulation knowledge graph according to the household heat metering data, regulation data and operation mode corresponding regulation knowledge of each thermal user;

[0077] S5, according to the operation mode recommendation results and recommendation language of each thermal user, recommending the selection of operation mode to the thermal user, and according to the operation mode selected by the thermal user, the thermal user portrait and the household regulation knowledge graph, combining the hydraulic balance mechanism knowledge of the heat supply system, training and optimizing the secondary network household regulation big model of the heat supply system, and outputting the secondary network household regulation strategy.

[0078] In the embodiment, the S1 specifically comprises:

[0079] Obtaining multi-modal data of each thermal user of the secondary network of the heat supply system, including text data, voice data and image data; the text data includes basic information of the heat user type, user age, house orientation, house area, floor, and user work and rest time, heat consumption in each period, flexible heat load response information, complaint information, maintenance work order information and preferred operation mode; the voice data includes heat supply effect voice evaluation information, fault description voice information and special heat demand voice information between the heat user and the heat supply enterprise; the image data includes indoor heat supply equipment state image, outdoor heat supply pipeline image, house appearance image, house layout image and house flow personnel image; the heat supply user type includes domestic heat supply and industrial and commercial heat supply;

[0080] The obtained multi-modal data is preprocessed and multi-modal fused to form a structured data set, and is input into a pre-trained large model, using a method of combining a large model and a prompt word, taking the structured data set and the prompt word as input, and then using an attention mechanism to extract features of the data, including text, voice and image features, obtaining thermal user basic features and user heat consumption behavior features, and then using a feature importance evaluation method to analyze the feature importance, and outputting a secondary network of each thermal user portrait including multiple portrait dimensions.

[0081] In practical applications, for civil heating, different residential users have different preferences for heating behavior. For example, young people have irregular work and rest, are more active at night, and are away from home during the day. They can try energy-saving heating mode, and when there is no one at home during the day, they can adjust the indoor temperature to flexibly adjust the heat load accordingly, and obtain part of the economic compensation. When they come home from work, they switch to the heating operation mode to ensure the indoor heating demand. The elderly and children have relatively regular time at home, and their physical condition is relatively weak. They have a low tolerance for cold and a high requirement for room temperature, and they generally do not frequently adjust the temperature. For different types of industrial and commercial users, the preferences for heating behavior are also different. For example, for production-type industries, different production processes have specific requirements for temperature, humidity, and other thermal environment parameters. The heating demand is usually stable and large, and the interruption of heating may cause production stagnation, product scrap, and other serious losses. For food processing enterprises, heating has periodicity, and the heating demand increases in the production season and decreases in the off-season. For shopping malls and shopping centers, the business hours are fixed and relatively long, and the indoor temperature needs to be kept comfortable during business hours to attract customers and improve the shopping experience. The demand for heating increases on weekends and holidays, and the temperature needs to be controlled in different zones according to the function and personnel density of different floors and areas, and the heating amount needs to be adjusted adaptively. For hospital heating objects, different departments have different requirements for temperature and humidity. For schools, the heating demand has obvious seasonality and time periodicity. During the winter semester, classrooms and offices need to be heated to ensure the comfort of teachers and students. During the winter and summer vacations, the heating demand is greatly reduced, and the school will adjust the operation of the heating system according to the course arrangement and work schedule to save energy. Therefore, different secondary network heating objects need to establish portraits of secondary network heating users, set different heating operation modes for different secondary network heating user portraits, and provide personalized time-periodic, zoned, and household control strategies.

[0082] In this embodiment, the prompt words when constructing the portrait of each heating user of the secondary network include:

[0083] Role setting: clearly define that the large model is an expert in the heating field during the dialogue process, and guide the large model to output the behavior mode and response mode of the heating user portrait;

[0084] Demand description: describe the task demand completed by the large model, including feature extraction and feature importance analysis;

[0085] Multi-portrait dimension definition: including user basic feature definition and user heating behavior definition, guiding the large model to correctly understand the specific meaning; wherein the user basic feature definition is the static information of the heating user; the user heating behavior definition is the dynamic information of the heating user, including heating user work and rest behavior information, heating user time-periodic heating amount preference information, flexible response information, operation mode preference information, and heating feedback information.

[0086] It should be noted that the hot user daily living behavior information: refers to the hot user's daily living, activity time regularity and other related information, such as the white-collar workers who go out early and come back late, and spend less time at home during the day; the elderly are used to going to bed early and getting up early, and are active at home in the morning and at night. These information can help the heat supplier to understand the user's time at home, so as to adjust the heat supply strategy, such as preheating before the user comes home and reducing the heat supply intensity when the user leaves home;

[0087] Hot user heat consumption preference information: is the demand tendency of heat users for heat at different time periods. For example, some families like slightly lower temperature when resting at night, and hope to quickly raise the indoor temperature in the morning after getting up; commercial places such as shopping malls have different heat consumption preferences in different time periods such as before business, during business and after business due to different number of personnel and activity conditions. Mastering these information helps the heat supply system to more accurately allocate heat and avoid energy waste;

[0088] Flexible response information: refers to the degree of cooperation and response capability of the heat user to the adjustment of the heat supply system. For example, when the heat supplier needs to adjust the heat consumption behavior of the user due to special circumstances (such as equipment maintenance, energy allocation), whether the user can respond in time and adjust the heat consumption equipment according to the requirements. At the same time, it also includes the willingness and ability of the heat user to participate in demand side response activities, such as reducing heat supply load during peak period;

[0089] Running mode preference information: that is, the preference and selection tendency of the heat user for different heat supply running modes. Common running modes include energy-saving mode, regular mode, adaptive mode, etc. Some users pay more attention to energy saving and will tend to choose energy-saving mode; some users pursue comfortable indoor environment and prefer regular heat consumption mode. Understanding the running mode preference of the user, the heat supply enterprise can provide heat supply service that meets the needs of the user and improve the user satisfaction;

[0090] Heat consumption feedback information: is the evaluation, opinion and suggestion of the heat user on the heat supply service, etc. For example, the user feedbacks that the indoor temperature is too high or too low, the heat supply is unstable, the equipment has problems, etc.; it can also be the evaluation of the heat supply service quality, such as customer service response speed, service attitude of maintenance personnel, etc. After collecting these feedback information, the heat supplier can improve the heat supply system and service in time, and improve the heat supply quality and management level.

[0091] In the embodiment, the portrait of each heat user of the secondary network is defined as:

[0092] P i = LLMs(D i , R);

[0093] D i is the structured data set of the i-th heat user; R is the prompt word when the heat user portrait is constructed; P iprofile of the ith hot user; LLMs() is a large model used when building a hot user profile.

[0094] In the present embodiment, the S2, the individual household operation mode of the secondary network of the heating system specifically includes:

[0095] Normal operation mode: in any period of the heating season, the indoor temperature of the user reaches the preset temperature;

[0096] Energy-saving operation mode: with time period division mechanism and customized temperature selection mechanism, it can select a lower indoor temperature than the conventional preset temperature in a fixed time period according to the user's heating behavior to save energy, while giving economic compensation;

[0097] Intelligent adaptive mode: real-time monitoring of indoor temperature, room activity time, and number of people, intelligent and automatic adjustment of heating capacity;

[0098] Economic optimization mode: according to energy price fluctuation information, provide the optimal operation scheme under the premise of meeting the basic comfort, reduce the heating cost.

[0099] In actual application, the characteristics and applicable objects of the individual household operation mode of the secondary network of the heating system are introduced as follows:

[0100] 1) Normal operation mode

[0101] Features: During the entire heating season, the system will maintain the indoor temperature of the user at the preset temperature value at any time, ensuring a warm and comfortable indoor environment. This mode does not consider changes in user behavior, energy prices, and other factors, providing consistent and stable heating for users who require high stability and do not need to participate in temperature adjustment;

[0102] Heating object: suitable for users who require high stability in indoor temperature and are not concerned about changes in heating cost, such as hospitals, nursing homes, and other environmentally sensitive locations, as well as some residents who require high comfort and have good economic conditions;

[0103] 2) Energy-saving operation mode

[0104] Features: Through time period division mechanism, the day or the entire heating season is divided into different time periods, such as daytime and nighttime, weekdays and weekends, etc. At the same time, users can choose a lower customized temperature than the conventional preset temperature in a fixed time period according to their heating behavior, thereby reducing heat consumption and achieving energy-saving purposes. To encourage users to adopt this mode, economic compensation such as fee reduction, point rewards, etc. will be given;

[0105] Hot object: Suitable for users with certain energy-saving awareness and regular work and rest, who can adjust their temperature needs according to time periods. For example, office worker families can set lower temperatures during the day when there is no one or less activity at home; business and industrial users who focus on cost control, such as warehouses and other places with relatively low temperature requirements;

[0106] 3) Intelligent adaptive mode

[0107] Features: With the help of sensors, real-time monitoring of indoor temperature, time of room activity, and number of people flowing, etc. Based on these data, the system can intelligently and automatically adjust the heating capacity. For example, when the room is empty for a long time, the heating capacity is automatically reduced; when the number of people increases or the activity is frequent, the heating capacity is appropriately increased, which saves energy to the maximum extent while ensuring comfort.

[0108] Hot object: Suitable for places with high requirements for comfort and energy saving, and frequent changes in indoor personnel activity.

[0109] Such as office buildings, shopping malls and other commercial places;

[0110] 4) Economic optimization mode

[0111] Features: This mode obtains real-time energy price fluctuation information, calculates the optimal operation scheme based on the premise of meeting the basic comfort, for example, when the energy price is low, appropriately increase the heating intensity and store part of the heat; when the energy price is high, reduce the heating intensity and use the stored heat to maintain the indoor temperature, thereby effectively reducing the heating cost;

[0112] Hot object: Mainly suitable for business and industrial users who are sensitive to heating costs, such as factories, large office buildings, etc. These places have large heating capacity and high heating cost proportion, and usually have some independent energy supply equipment (such as air source heat pump, ground source heat pump, photovoltaic heating equipment, etc.), which can be independently controlled. Through this mode, the operation cost can be significantly reduced.

[0113] In this embodiment, the S2 specifically comprises:

[0114] Obtain the definition of each sub-house operation mode of the secondary network of the heating system, the historical selection of each sub-house operation mode by the heat user, and the operation parameters of the operation mode;

[0115] Set a prompt word for guiding the generation of the operation mode portrait by the large model, and clearly define its function in the generation and description of the operation mode portrait;

[0116] The name of the operation mode, the mode characteristic description, the association information between each operation mode and the heat user portrait, and the operation parameters of the operation mode are taken as operation mode data, combined with the prompt word, input into the large model, and the operation mode portrait is generated, represented as:

[0117] P j = LLMs'(L j ,M);

[0118] L j is the running mode data of the jth running mode; M is the prompt word when the running mode image is generated; P j is the image of the jth running mode; LLMs' is a large model used to generate the running mode image.

[0119] In the embodiment, the large model generating the running mode image comprises:

[0120] The input sequence data is segmented into word pieces using a word piece tokenizer, and each word piece is mapped into a low-dimensional vector space to obtain word piece embeddings, and position encoding is added to enable the large model to learn the position information of the word pieces in the sequence;

[0121] Based on the multi-layer transformer encoder structure of the large model, the multi-head self-attention mechanism of each encoder is used to enable the large model to learn the relationship between word pieces in different representation subspaces and capture semantic information, and a feedforward neural network is used to transform and extract features from the output of the multi-head self-attention, to learn different abstract representations of the input data at different levels, and to obtain feature representations containing rich information of the input data by continuously stacking;

[0122] In the output layer, a linear transformation is used to map the feature representation to a dimension equal to the size of the vocabulary to obtain the predicted score of each word piece; a decoding strategy is used to select the word piece with the highest probability as the next generated word piece according to the predicted score, and the text content of the running mode image is gradually generated.

[0123] In practical applications, a large language model suitable for processing structured and unstructured data fusion is selected. The model is adapted to understand the professional terms and data structures in the heating field, fine-tuned to enhance the understanding of heating-related knowledge, and the household operation mode information and associated heat user portrait data are sorted into a format suitable for large model input. The sorted data is input into the large model, and a specific prompt word is used to guide the model to generate the operation mode portrait. The prompt word can be designed as: "According to the following household operation mode and related heat user portrait information, generate a detailed portrait of this operation mode, including but not limited to the main features of this mode, applicable scenarios, user group characteristics, energy efficiency performance, etc." The model generates a portrait description about the household operation mode based on the input data and prompt word. The operation mode portrait generated by the large model is manually reviewed and optimized. Check if the generated portrait accurately reflects the characteristics of the operation mode and the user group characteristics. If there are inaccuracies or incompleteness, adjust the input data or prompt word and run the model again until a satisfactory operation mode portrait is generated. At the same time, it can be verified by comparing with actual operation data and user feedback to ensure the authenticity and reliability of the portrait.

[0124] It should be noted that tokenization: using a tokenizer to split the input sequence into individual tokens. For example, "energy-saving operation mode" may be tokenized into "energy-saving", "operation", "mode", etc. Embedding layer: map each token to a low-dimensional vector space to get token embedding. At the same time, in order to let the model learn the position information of the token in the sequence, position encoding is also added. Position encoding can be fixed sine cosine function encoding or learnable parameters. Finally, each token in the input sequence corresponds to an embedding vector containing token information and position information.

[0125] In the multi-head self-attention mechanism, attention calculation: in each encoder layer, first perform multi-head self-attention calculation. The multi-head self-attention mechanism maps the input embedding vectors through multiple linear transformations to obtain query (Query), key (Key) and value (Value) matrices. Then, calculate the similarity score between the query and the key, usually using dot product operation, and then convert the score to probability distribution through softmax function, and finally weight the sum of the value according to the probability distribution to get the attention output; multi-head combination: concatenate the attention outputs of multiple heads, and then reduce the dimension through a linear transformation to get the final multi-head self-attention output. Multi-head self-attention mechanism can make the model learn the relationship between tokens in different representation subspaces and capture more rich semantic information.

[0126] The output of multi-head self-attention is fed into a feedforward neural network. This feedforward network typically consists of two fully connected layers, with a non-linear transformation using activation functions such as ReLU in between. The feedforward network further transforms and extracts features from the multi-head self-attention output, enhancing the model's expressive power. Furthermore, residual connections are used on both the multi-head self-attention and feedforward network outputs, meaning the input is directly added to the output. Residual connections alleviate the vanishing gradient problem, making the model easier to train.

[0127] The multi-head self-attention, feedforward neural network, and residual connection described above are combined into an encoder layer. Then, multiple encoder layers are stacked together to form a multi-layer Transformer encoder. Each encoder layer learns a higher-level abstract representation based on the previous layer. By continuously stacking these layers, the model can gradually capture the complex semantic and structural information of the input data.

[0128] like Figure 3 As shown, in this embodiment, S3 specifically includes:

[0129] The profiles of hot users and operating modes of each secondary network are integrated into the input dataset;

[0130] Set up prompts to guide the large model to output personalized operating mode recommendation results and recommended dialogue, and clarify the role, task requirements and thought process of the large model when recommending operating modes and dialogue.

[0131] The input dataset and prompt words are fed into the large model, which uses its internal attention mechanism and neural network layers to calculate the similarity between the hot user profile features and the profile features of each operating mode.

[0132] Based on the similarity calculation results, the large model selects the operating mode with the highest similarity as the personalized recommendation result output for popular users;

[0133] The large model recommends results based on the personalized operating mode output, and combines them with a pre-built corpus of terms to output differentiated recommended terms for different popular users.

[0134] It's important to clarify that when recommending operating modes and communication strategies, the role of the large-scale model should be defined: explain the responsibilities of this role to the large-scale model, for example, "As an intelligent heating consultant, you need to provide the most suitable heating operating mode recommendation based on the detailed information of heat users, and give the reasons for the recommendation and usage suggestions in easy-to-understand and user-friendly language." This ensures the large-scale model clearly understands its positioning and the goals it needs to achieve.

[0135] Task requirement description, including input data description: detailedly inform the data content input by the large model, such as "you will receive the hot user's work and rest behavior information, the heat preference information in each period, the flexible response information, the running mode preference information and the heat feedback information", so that the large model knows the data range that can be relied on. Recommendation result requirement: clearly present the form of the running mode recommendation result, such as "the recommendation result should contain at least two most suitable running modes, and be sorted in descending order of matching degree". At the same time, the style and content points of the recommendation language are specified, "the recommendation language should be concise and clear, highlight how each running mode meets the specific needs of the user, and the length of the language is controlled between 100-150 words".

[0136] Thought chain construction: guide the large model to analyze according to certain logical steps, for example "firstly, analyze the hot user's work and rest behavior information, judge the user's time distribution law at home, and preliminarily screen out the running modes that may be suitable. Then, combined with the heat quantity preference information in each period, further determine the energy saving or comfort advantage of each mode in different periods. Then, according to the flexible response information, evaluate the user's acceptance degree to different adjustment methods, and adjust the recommended mode"; finally, according to the recommended running mode, output the recommendation language.

[0137] As shown in Figure 4 , in the embodiment, the S4 specifically includes:

[0138] Obtain the heat metering data of each hot user, including heat consumption, flow and temperature in different time periods;

[0139] Obtain the sub-household regulation and control data of each hot user, including running mode switching time and adjustment parameters of regulation and control equipment;

[0140] Obtain the regulation and control knowledge corresponding to the running mode, including the regulation and control mechanism of the regulation and control equipment under different running modes and the target temperature range;

[0141] According to the sub-household heat metering data, regulation and control data and regulation and control knowledge corresponding to the running mode of each hot user, use deep learning method to extract entities, relationships and attributes respectively; the relationship extraction includes the relationship between the hot user and the regulation and control equipment, the relationship between the hot user and the running mode, and the relationship between the regulation and control equipment and the running mode;

[0142] Establish a sub-household regulation and control knowledge graph according to the extracted entities, relationships and attributes, including: determining the structure of the sub-household regulation and control knowledge graph of the secondary network, using the form of directed graph, taking nodes to represent entities and edges to represent the relationship between entities; add entities as nodes to the knowledge graph, and create corresponding edges according to the relationship, and add corresponding attribute information to each node and edge.

[0143] It should be noted that the entities include heat user entities, control equipment entities, and operation mode entities; attribute extraction includes: for heat user entities, extracting heat consumption, temperature, etc. from their heat metering data as attributes; for control equipment entities, extracting opening degree, adjustment time, etc. from their control data as attributes; for operation mode entities, extracting target temperature range, energy saving rate, etc. from control knowledge as attributes.

[0144] like Figure 5 As shown, in this embodiment, S5 specifically includes:

[0145] The data on the operating modes selected by hot users are linked and integrated with the hot user profile data;

[0146] Extract knowledge information related to the operation mode and control equipment of heat users from the individual household control knowledge graph, transform it into a vector form suitable for model input, and fuse the processed knowledge graph vector with heat user data;

[0147] The knowledge of the hydraulic balance mechanism of the heating network is transformed into a mathematical model and constraints, and the relevant parameters of the hydraulic balance of the heating network are extracted.

[0148] Define the training objective: With the goal of generating the optimal secondary network household control strategy, the training objective function is set according to the operating mode requirements of heat users, minimizing the energy consumption of the heating network, and ensuring the hydraulic balance of the heating network.

[0149] The data on the correlation between the operation mode and the heat user profile, the knowledge graph data, and the hydraulic balance data of the heating network are integrated as training samples and input into the large model for training. This generates a large model of the secondary network of the heating system for individual household control. The parameters of the large model are continuously adjusted so that the large model can learn the relationship between the heat user operation mode, the heat user profile, the individual household control knowledge graph, and the hydraulic balance of the heating network.

[0150] After training and optimization, the large model is fed with new hot user operation modes, hot user profiles, and real-time operation data of the secondary network. Based on the learned knowledge and relationships, the large model generates a secondary network household control strategy, including the adjustment parameters of the household control equipment and the switching time of the operation mode.

[0151] In actual application, the training and optimization of the large model of the secondary network household regulation and control of the heating system adopts a deep reinforcement learning algorithm, specifically including: representing the secondary network household regulation and control environment state through the operation mode selected by the heat user, the heat user portrait, the household regulation and control knowledge graph, and the hydraulic balance working condition of the heat network where the heat user building is located; taking the regulation and control parameters of the household regulation and control equipment of the secondary network heat user as the action space, and designing a reward mechanism according to the regulation and control effect of the action; the regulation and control effect is preset as minimizing the energy consumption of the heat network and ensuring the hydraulic balance of the heat network on the premise of meeting the operation mode of each heat user; by continuously executing the action, observing the result, obtaining the reward, and updating the value function Q value of each action, the large model can learn the optimal secondary network household regulation and control strategy.

[0152] In several embodiments provided in the present application, it should be understood that the disclosed system and method can also be implemented in other ways. The system embodiments described above are only illustrative, for example, the flowcharts and block diagrams in the drawings show the possible implementation architecture, function and operation of the system, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment or a part of code, which includes one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different order from that noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0153] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. If the functions are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art, or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various program code storage media.

[0154] The above-described embodiments according to the present application are merely exemplary and are not intended to limit the present application. A person skilled in the art can make various changes and modifications of the present application without departing from the spirit and scope of the present application. The technical scope of the present application should be determined by the following claims.

Claims

1. A personalized control method for a secondary heating network based on a large model and multimodal profiling, characterized in that, It includes: S1. Obtain multimodal data of each heat user in the secondary network of the heating system, and after feature extraction and importance analysis using a large model, construct a profile of each heat user in the secondary network, specifically including: The system acquires multimodal data from each heat user in the secondary network of the heating system, including text data, voice data, and image data. The text data includes basic information such as heat user type, age, building orientation, building area, floor level, user's daily routine, heat consumption during different time periods, flexible heat load response information, complaint information, maintenance work order information, and preferred operating modes. The voice data includes voice evaluations of heating performance between heat users and the heating company, voice descriptions of faults, and voice information regarding special heating needs. The image data includes images of indoor heating equipment status, outdoor heating pipelines, building exteriors, building layouts, and images of people moving around the building. The heat user types include residential heating and industrial / commercial heating. The acquired multimodal data is preprocessed and fused to form a structured dataset, which is then input into a pre-trained large model. The large model and prompt words are combined, and the structured dataset and prompt words are used as input. An attention mechanism is then used to extract features from the data, including features from text, speech and images, to obtain basic features of hot users and user hot behavior features. After feature importance analysis is performed using a feature importance evaluation method, the output includes a secondary network profile of each hot user with multiple profile dimensions. S2. Extract information on the operation modes of each household in the secondary network of the heating system, combine this information with the heat user profile of the selected household operation mode, and generate an operation mode profile for that household operation mode using a large model. Specifically, this includes: Obtain the definition of each household's operating mode in the secondary network of the heating system, the heat user information of each household's operating mode selected in history, and the operating parameters of the operating mode; Set prompts to guide the generation of operational mode profiles from large models, and clarify their function in describing the generation of operational mode profiles; The name of the operating mode, the description of its characteristics, the correlation information between each operating mode and the hot user profile, and the operating parameters of the operating mode are used as operating mode data. These are then combined with prompt words input into the large model to generate an operating mode profile, represented as follows: ； For the first Operation mode data for each operation mode; Prompt text when generating a profile for the operating mode; For the first A profile of each operating mode; The large model used when generating operational mode profiles; S3. Input the profiles of each hot user and the profile of each operating mode of the secondary network into the large model for matching analysis, and output personalized operating mode recommendation results and recommendation words for each hot user; S4. Based on the individual heat metering data, control data, and control knowledge corresponding to the operation mode of each heat user, establish an individual control knowledge graph; S5. Based on the recommended operating modes and recommended messages for each heat user, recommend the selected operating mode to the heat user, and then, based on the selected operating mode, heat user profile, and household control knowledge graph, combined with the knowledge of the hydraulic balance mechanism of the heating network, train and optimize the large-scale model of household control of the secondary network of the heating system, and output the household control strategy of the secondary network.

2. The personalized control method for a secondary heating network according to claim 1, characterized in that, The prompts used when constructing the profiles of each hot user in the secondary network include: Role setting: Clearly define the large model as an expert in the heating field during the dialogue process, and guide the large model to output the behavioral patterns and response methods of the heat user profile; Requirements Description: Describe the tasks required for the large model to perform, including feature extraction and feature importance analysis; Multi-dimensional user profile definition: including user basic feature definition and user heat usage behavior definition, to guide the large model to correctly understand specific meanings; wherein, the user basic feature is defined as the static information of heat users; the user heat usage behavior is defined as the dynamic information of heat users, including heat user work and rest behavior information, heat user heat usage preference information at different times, flexible response information, operation mode preference information, and heat usage feedback information.

3. The personalized control method for a secondary heating network according to claim 1, characterized in that, The profiles of each hot user in the secondary network are defined as follows: ； For the first A structured dataset of hot users; Prompt words used when building hot user profiles; For the first A profile of a popular user; The large model used to build hot user profiles.

4. The personalized control method for a secondary heating network according to claim 1, characterized in that, In S2, the specific operating modes of each household in the secondary network of the heating system include: Normal operating mode: The indoor temperature reaches the preset temperature at any time during the heating season; Energy-saving operation mode: It has a time period division mechanism and a customized temperature selection mechanism. It can select to save energy by setting an indoor temperature lower than the conventional preset temperature during a fixed time period according to the user's heating behavior, while providing economic compensation. Intelligent adaptive mode: Real-time monitoring of indoor temperature, room activity time, and number of people moving around, intelligently and automatically adjusting heating supply; Economic optimization mode: Based on energy price fluctuation information, it provides the optimal operating plan while meeting basic comfort requirements, thereby reducing heating costs.

5. The personalized control method for a secondary heating network according to claim 1, characterized in that, The large model generates a profile of its operational mode, including: The input sequence data is divided into words using a word segmenter, and each word is mapped to a low-dimensional vector space to obtain word embeddings. At the same time, positional encoding is added to allow the large model to learn the positional information of the words in the sequence. The multi-layer transformer encoder structure based on a large model utilizes the multi-head self-attention mechanism of each encoder to enable the large model to learn the relationship between lexical units in different representation subspaces and capture semantic information. Then, a feedforward neural network is used to perform information transformation and feature extraction on the output of the multi-head self-attention. Different abstract representations of the input data are learned at different levels. By continuously stacking, feature representations containing rich information of the input data are obtained. In the output layer, a linear transformation is used to map this feature representation to a dimension of the same size as the vocabulary, obtaining the prediction score for each word. A decoding strategy is then used to select the word with the highest probability based on the prediction score as the next word to be generated, and the text content of the running mode profile is generated step by step.

6. The personalized control method for a secondary heating network according to claim 1, characterized in that, S3 specifically includes: The profiles of hot users and operating modes of each secondary network are integrated into the input dataset; Set up prompts to guide the large model to output personalized operating mode recommendation results and recommended dialogue, and clarify the role, task requirements and thought process of the large model when recommending operating modes and dialogue. The input dataset and prompt words are fed into the large model, which uses its internal attention mechanism and neural network layers to calculate the similarity between the hot user profile features and the profile features of each operating mode. Based on the similarity calculation results, the large model selects the operating mode with the highest similarity as the personalized recommendation result output for popular users; The large model recommends results based on the personalized operating mode output, and combines them with a pre-built corpus of terms to output differentiated recommended terms for different popular users.

7. The personalized control method for a secondary heating network according to claim 1, characterized in that, S4 specifically includes: Acquire heat metering data for each heat user, including heat consumption, flow rate, and temperature at different time periods; Obtain individual control data for each heat user, including operating mode switching time and control equipment adjustment parameters; Acquire the control knowledge corresponding to the operating mode, including: the control mechanism of the control equipment and the target temperature range under different operating modes; Based on the individual heat metering data, control data, and control knowledge corresponding to the operation mode of each heat user, deep learning methods are used to extract entities, relationships, and attributes respectively; the relationship extraction includes the relationship between heat users and control equipment, the relationship between heat users and operation mode, and the relationship between control equipment and operation mode. Based on the extracted entities, relationships, and attributes, a household-specific regulation knowledge graph is established, including: determining the structure of the secondary network household-specific regulation knowledge graph, adopting a directed graph form, with nodes representing entities and edges representing relationships between entities; adding entities as nodes to the knowledge graph, creating corresponding edges according to relationships, and adding corresponding attribute information to each node and edge.

8. The personalized control method for a secondary heating network according to claim 1, characterized in that, S5 specifically includes: The data on the operating modes selected by hot users are linked and integrated with the hot user profile data; Extract knowledge information related to the operation mode and control equipment of heat users from the individual household control knowledge graph, transform it into a vector form suitable for model input, and fuse the processed knowledge graph vector with heat user data; The knowledge of the hydraulic balance mechanism of the heating network is transformed into a mathematical model and constraints, and the relevant parameters of the hydraulic balance of the heating network are extracted. Define the training objective: With the goal of generating the optimal secondary network household control strategy, the training objective function is set according to the operating mode requirements of heat users, minimizing the energy consumption of the heating network, and ensuring the hydraulic balance of the heating network. The data on the correlation between the operation mode and the heat user profile, the knowledge graph data, and the hydraulic balance data of the heating network are integrated as training samples and input into the large model for training. This generates a large model of the secondary network of the heating system for individual household control. The parameters of the large model are continuously adjusted so that the large model can learn the relationship between the heat user operation mode, the heat user profile, the individual household control knowledge graph, and the hydraulic balance of the heating network. After training and optimization, the large model is fed with new hot user operation modes, hot user profiles, and real-time operation data of the secondary network. Based on the learned knowledge and relationships, the large model generates a secondary network household control strategy, including the adjustment parameters of the household control equipment and the switching time of the operation mode.

Images