A Multi-Condition Scheduling Control Method for Heating Systems Based on Large Models and Multi-Agents

Abstract

This invention discloses a multi-condition scheduling and control method for a heating system based on a large model and multiple agents, comprising: establishing a digital twin model of the heating system; setting up a scheduling and control center management agent, and after predicting the heat load demand using a large language model combined with a heat load prediction model, dividing the system into multiple operating conditions, and setting up multiple sub-agents corresponding to the operating conditions; the scheduling and control center management agent determines whether to trigger the operation of the sub-agents under the corresponding operating condition, and simultaneously transmits the predicted heat load demand values ​​for each time period and the heat generation of new energy heating units to the corresponding sub-agents; the sub-agents use the large language model to establish a system optimization scheduling and control knowledge graph based on the large language model, obtain the scheduling and control strategies of each device on the source, grid, and load sides of the heating system, and feed them back to the digital twin model to guide the actual operation and management of the heating system.

Description

Technical Field

[0001] This invention belongs to the field of heating system scheduling and control technology, specifically relating to a multi-condition scheduling and control method for heating systems based on a large model and multiple agents. Background Technology

[0002] Optimal scheduling and control of heating systems involves scheduling and controlling various equipment on the source, grid, and load sides based on terminal heat load demand and the heating characteristics of source-side units. This is crucial for the reliable operation of the heating system and the balance between heat supply and demand. With the widespread application of new energy heating units, optimal scheduling and control of heating systems must consider not only heat load demand but also various uncertainties such as the heat output of new energy sources and meteorological conditions. Currently, traditional scheduling and control methods have certain limitations, such as limited accuracy and efficiency, poor interpretability of results, and fragmented scheduling and control knowledge and rules.

[0003] Furthermore, with the development of large language models and intelligent agent technologies, large language models have demonstrated powerful language understanding capabilities and rich knowledge reserves, while intelligent agent technologies possess the characteristics of perceiving the environment, making decisions, and taking actions, exhibiting autonomy and intelligence. The deep integration of intelligent agents and large language models constructs an intelligent system capable of understanding and responding to complex environments and possessing a high degree of autonomy, marking a new stage in the development of artificial intelligence technology and bringing new opportunities for the optimized scheduling and control of heating systems. Therefore, the deep integration of large language models and intelligent agent technologies, applied to the field of optimized scheduling and control technology for heating systems, is expected to improve the accuracy and efficiency of optimized scheduling and control, and is of great significance to the efficient and reliable operation of heating systems.

[0004] Based on the above technical problems, it is necessary to design a new multi-condition scheduling and control method for heating systems based on large models and multiple agents. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a multi-condition scheduling and control method for heating systems based on a large model and multiple agents. This method makes full use of the powerful natural language processing capabilities and rich pre-trained knowledge of the large language model, combined with the perception and decision-making technology of agents, to realize the scheduling and control management of the heating system under various operating conditions. Moreover, it integrates various scattered scheduling and control knowledge and rules, and visualizes and supports them in the form of knowledge graphs, thereby improving the accuracy and efficiency of the optimized scheduling and control of the heating system.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] This invention provides a multi-condition scheduling and control method for heating systems based on a large model and multiple agents, comprising:

[0008] S1. Establish a digital twin model of the heating system, including source-side new energy heating units, conventional heating units, thermal storage devices, grid-side heating stations, and load-side heat users;

[0009] S2. Using a large language model, based on the multidimensional data collected from the digital twin model of the heating system, heat load description text is generated, and then combined with the heat load prediction model, the heat load demand of the heating system in future periods is predicted.

[0010] S3. Based on the predicted heat load demand for each future period, the system is divided into multiple operating conditions;

[0011] S4. Determine whether the current system meets the conditions for entering the next operating condition. If it does, trigger the corresponding operating condition and obtain the heat load demand value for each future time period and predict the heat output of the new energy heating unit for each future time period.

[0012] S5. Using a large language model and a pre-set intelligent agent, the heat load demand value and the heat output of new energy heating units are analyzed. Combined with historical operating data under various operating conditions, set operating targets, and the scheduling and control rules of each device on the source, grid, and load sides, the scheduling and control strategies of each device on the source, grid, and load sides of the heating system are obtained and fed back to the digital twin model to guide the actual operation and management of the heating system.

[0013] Furthermore, S1 specifically includes:

[0014] Install corresponding sensors on the source side of new energy heating units, conventional heating units, heat storage devices, grid-side heating stations and load-side heat user terminals to obtain actual unit operating parameters, heat storage status information, heating station inlet and outlet water supply parameters and heat user heat consumption data.

[0015] Establish source-side new energy heating unit models, conventional heating unit models, thermal storage device models, grid-side heating station models, heating network pipeline models, and load-side heat user models;

[0016] Determine the interface relationships between the source-side, network-side, and load-side models, and clarify the content and methods of data interaction;

[0017] The source-side, network-side, and load-side models are integrated to form a complete digital twin model of the heating system;

[0018] By setting different operating conditions and parameters, and using the acquired actual data to simulate and verify the digital twin model, the model parameters are adjusted and optimized.

[0019] Furthermore, S2 specifically includes:

[0020] Set up a dispatch control center to manage the intelligent agent, input the heating operation data, heat load data and weather data collected by the digital twin model of the heating system into the big language model, and extract data features through the Transformer structure of the big language model to obtain the features of heating operation data, heat load data and weather data.

[0021] Multiple prompt templates are preset and stored. The similarity between the characteristics of heating operation data, heat load data, and weather data and the preset slot values ​​of the prompt templates is analyzed. The prompt template with the highest similarity is used as the target prompt template.

[0022] The slot values ​​of the target prompt template are updated based on the characteristics of heating operation data, heat load data, and weather data to generate heat load description text;

[0023] Keyword extraction is performed on the heat load description text. Based on the keywords, the characteristics affecting the heat load of the heating system are obtained and then input into a machine learning algorithm for processing. A heat load prediction model is then established to obtain the heat load demand data of the heating system in various future periods.

[0024] Furthermore, S3 specifically includes:

[0025] The scheduling and control center management agent obtains the heat load demand data for each future time period output by the heat load prediction model, and performs statistical analysis to obtain the maximum, minimum, average, and rate of change of the heat load, thereby clarifying the trend and fluctuation range of the heat load.

[0026] Based on the changing trend, fluctuation range, and rate of change of heat load, as well as the actual operating characteristics of the system, the system is divided into multiple operating conditions, and clear boundary conditions, characteristic descriptions, and scheduling and control objectives are set for each operating condition.

[0027] Furthermore, in step S4, it is determined whether the current system meets the conditions for entering the next operating condition. If it does, the corresponding operating condition is triggered, including:

[0028] The management agent of the scheduling and control center determines whether the system's operating conditions have changed based on the current system's operating data. If the operating conditions have changed, it matches the input data with various operating condition characteristics based on the learned knowledge of different operating condition characteristics, identifies the operating condition type, and switches the operating conditions according to the identified operating conditions, triggering the corresponding operating condition sub-agent to start working.

[0029] Among them, the operating condition sub-agent is created for each operating condition. Each operating condition sub-agent has independent simulation and scheduling control capabilities, and has a knowledge base and algorithm model related to that operating condition. Each operating condition sub-agent is also configured with relevant initial parameters. The scheduling control center management agent is responsible for coordinating the work between the operating condition sub-agents, switching operating conditions and triggering the work of each operating condition sub-agent. At the same time, each operating condition sub-agent feeds back the simulation and scheduling control effect of the heating system under its operating condition to the scheduling control center management agent.

[0030] Furthermore, in S4, predicting the heat output of the new energy heating units for future time periods includes:

[0031] Based on the historical operating data, heat output, and meteorological data of the new energy heating units, key data affecting the heat output of the new energy heating units are extracted, and prompt word templates are constructed based on the key data.

[0032] Obtain structured historical operating data, heat production, and meteorological data; extract key data; map them with the corresponding key data in the prompt word template; and construct a text description of the structured data using natural language to serve as a textualized unit heat production dataset.

[0033] A large language model is selected as the basic model for fine-tuning, and the textualized unit heat production dataset is used as the training data for supervised fine-tuning. Supervised fine-tuning training is performed on the large language model so that it learns the data knowledge of new energy unit heat production.

[0034] By using natural language, the data features affecting the heat production of new energy heating units are input into a large language model that is fine-tuned and trained, and the predicted heat production values ​​of new energy heating units for each future time period are output.

[0035] Furthermore, during the supervised fine-tuning training of the large language model, the low-rank adaptive LoRA technique is employed. For the original weight matrix in the training model, two low-rank matrices A and B are introduced to reduce the number of trainable parameters, as shown below:

[0036] ;

[0037] This is the original weight matrix of the model; The scaling factor is A; A and B are low-rank matrices. This is the updated model weight matrix;

[0038] The gradient descent algorithm is used to calculate the loss function. With respect to the gradients of A and B, update the low-rank matrices A and B, as follows:

[0039] ;

[0040] ;

[0041] , This is the updated low-rank matrix; The learning rate; , Loss functions The gradients of A and B.

[0042] Furthermore, in step S5, before obtaining the scheduling and control strategies for each device on the source, network, and load sides of the heating system, it is necessary to establish a system optimization scheduling and control knowledge graph based on a large language model, specifically including:

[0043] At the data layer: The created sub-agents collect historical operating data, scheduling and control strategies, operating objectives, and text data of scheduling and control rules of each device on the source, grid, and load sides under their respective operating conditions, and perform data preprocessing.

[0044] Large Language Model Training: Select a large language model, convert the preprocessed text data into an input format acceptable to the large language model, and use the text data to conduct supervised training on the large language model, so that the model can learn the relationship between heat load demand, heat output and operating conditions of new energy heating units, scheduling and control strategies, and operating objectives, so that the model can understand and generate knowledge related to system optimization scheduling and control.

[0045] At the conceptual level: large language models are used to mine and classify entities and relationships in text data, as well as to annotate the corpus;

[0046] At the instance layer: knowledge extraction and instantiation are performed using a large language model;

[0047] Based on the entities, relations, and corpus annotations at the concept layer, the ontology structure of the knowledge graph is defined. Based on the knowledge extraction and instantiation at the instance layer, specific system scheduling and control knowledge graph instances are formed, and a system optimization scheduling and control knowledge graph based on a large language model is established.

[0048] Furthermore, the historical operating data includes heat load demand, heat output of new energy heating units, operating parameters of conventional heating units, operating parameters of heating stations, and heat metering data of heat users; the scheduling and control strategy includes the operating output of source-side new energy heating units and conventional heating units, the heat storage and release status and heat storage and release status of heat storage devices, as well as the pump and valve adjustment parameters of grid-side heating stations and the pump and valve adjustment parameters of load-side heat users at the building entrance and household entrance; the operating objectives include economic objectives, energy-saving objectives, environmental protection objectives, and comfort objectives; the scheduling and control rules include the scheduling and control logic principles of each device on the source, grid, and load sides and the scheduling and control coupling between devices.

[0049] Furthermore, in the conceptual layer, a large language model is used to obtain all nouns in the text data. After identifying the entity objects, prompt words are introduced for corpus clustering and classification. The large language model and machine learning algorithms are used to mine the relationships between entities from the text data. In the instance layer, the large language model is used to extract scheduling control computational knowledge rules from the text data based on the ontology structure and corpus annotations. After instantiating them into entity-relation-entity triples, a graph database is called to store and visualize the triples generated by the large language model.

[0050] The beneficial effects of this invention are:

[0051] (1) This invention uses a large language model to predict heat load demand and heat output of new energy heating units. It can more accurately grasp the trend of heat load change and the trend of heat output of new energy heating units. Compared with traditional prediction methods, it can understand various image factors in a language-based way, consider factors more comprehensively, and predict with higher accuracy. Furthermore, it can quickly model through language-based fine-tuning, improving prediction efficiency. In addition, the natural language process makes the entire prediction process interpretable.

[0052] (2) By setting up a dispatch control center management intelligent agent, the present invention realizes centralized management and unified dispatch of the entire heating system. At the same time, each working condition sub-intelligent agent is responsible for the simulation and dispatch control under its own operating conditions, forming a management mode that combines centralized and decentralized approaches, improving the efficiency of dispatch control management and making the system operation management more orderly. In addition, the dispatch control center management intelligent agent divides the operating conditions according to the heat load demand prediction results, and triggers the working condition sub-intelligent agents to work when the operating condition switching conditions are met, so that the system can carry out refined management according to different heat load conditions, flexibly adapt to various changes in actual operation, and improve the reliability and stability of system operation.

[0053] (3) The working condition sub-intelligent agent establishes an optimized scheduling and control knowledge graph based on a large language model, integrates various scattered scheduling and control knowledge and rules in the heating system, stores them in the knowledge graph in a structured form and displays them visually, provides intuitive and clear decision support, and can comprehensively consider multiple factors to formulate a more scientific and reasonable scheduling and control strategy, so as to realize the coordinated and optimized operation of various equipment on the source, grid and load sides.

[0054] 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.

[0055] 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. Attached Figure Description

[0056] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a multi-condition scheduling and control method for a heating system based on a large model and multiple agents, according to the present invention.

[0058] Figure 2 This is a flowchart of the heat load demand prediction method using a large language model in this invention;

[0059] Figure 3 A flowchart is used to establish the knowledge graph for the optimized scheduling and control of the heating system of this invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] like Figure 1 As shown, this embodiment provides a multi-condition scheduling and control method for a heating system based on a large model and multiple agents, which includes:

[0062] S1. Establish a digital twin model of the heating system, including source-side new energy heating units, conventional heating units, thermal storage devices, grid-side heating stations, and load-side heat users;

[0063] S2. Using a large language model, based on the multidimensional data collected from the digital twin model of the heating system, heat load description text is generated, and then combined with the heat load prediction model, the heat load demand of the heating system in future periods is predicted.

[0064] S3. Based on the predicted heat load demand for each future period, the system is divided into multiple operating conditions;

[0065] S4. Determine whether the current system meets the conditions for entering the next operating condition. If it does, trigger the corresponding operating condition and obtain the heat load demand value for each future time period and predict the heat output of the new energy heating unit for each future time period.

[0066] S5. Using a large language model and a pre-set intelligent agent, the heat load demand value and the heat output of new energy heating units are analyzed. Combined with historical operating data under various operating conditions, set operating targets, and the scheduling and control rules of each device on the source, grid, and load sides, the scheduling and control strategies of each device on the source, grid, and load sides of the heating system are obtained and fed back to the digital twin model to guide the actual operation and management of the heating system.

[0067] In this embodiment, S1 specifically includes:

[0068] Install corresponding sensors on the source side of new energy heating units, conventional heating units, heat storage devices, grid-side heating stations and load-side heat user terminals to obtain actual unit operating parameters, heat storage status information, heating station inlet and outlet water supply parameters and heat user heat consumption data.

[0069] Establish source-side new energy heating unit models, conventional heating unit models, thermal storage device models, grid-side heating station models, heating network pipeline models, and load-side heat user models;

[0070] Determine the interface relationships between the source-side, network-side, and load-side models, and clarify the content and methods of data interaction;

[0071] The source-side, network-side, and load-side models are integrated to form a complete digital twin model of the heating system;

[0072] By setting different operating conditions and parameters, and using the acquired actual data to simulate and verify the digital twin model, the model parameters are adjusted and optimized.

[0073] In practical applications, multi-condition operating data of the heating system are integrated into an established digital twin model. A reverse identification method is then used to adaptively identify and correct the simulation results of the digital twin model, resulting in a corrected digital twin model. Specifically, based on the simulation results of the digital twin model and the actual collected operating data, the mean square error of both is set as the objective function to measure the difference between them. By adjusting the parameters in the digital twin model to minimize the objective function, the gradient descent method can be used to search for the optimal parameter combination.

[0074] In practical applications, new energy heating units include solar thermal collection systems, which utilize solar energy to convert light energy into heat energy. The heat absorbed by the absorber is conducted to the collector tubes, which then transfer the heat to the water in the storage tank, causing the water temperature to rise. The heat balance model of the absorber is expressed as:

[0075] ;

[0076] ρ is the mass of water; c is the specific heat capacity of water; The temperature of the solar absorber; The temperature of the heat collector tube; The surface area of ​​the absorber; Solar energy absorption rate; Input solar energy; The overall heat transfer coefficient of the system; For heat loss;

[0077] The heat balance model of the heat collector tube is expressed as:

[0078] ;

[0079] The area of ​​the heat collection tube; The temperature of the water in the storage tank;

[0080] The heat balance model of water in the storage tank is expressed as: .

[0081] Conventional heating units include electric boilers and combined heat and power (CHP) units. Electric boilers can convert electrical energy into heat energy, and under the guidance of time-of-use electricity pricing, they can optimize the heat load curve. The output heat power model is expressed as: ; Let t be the output thermal power of the electric boiler. For heating efficiency; Input electrical power to the electric boiler;

[0082] The output cost of an electric boiler is expressed as follows: ; Let t be the cost of the electric boiler's output. This is the unit output cost coefficient for electric boilers.

[0083] A combined heat and power (CHP) unit includes a gas turbine and a waste heat recovery system. The waste heat recovery system recovers waste heat, achieving cascaded energy utilization. The model is represented as follows:

[0084] ;

[0085] in, This represents the electrical power output of the combined heat and power unit at time t. To provide power output for the combined heat and power unit at time t; For the power generation efficiency of combined heat and power units; The thermal power output of the combined heat and power unit at time t; For the heat conversion efficiency of a combined heat and power unit; This represents the natural gas consumption of the combined heat and power unit at time t. This is the lower heating value of natural gas.

[0086] The thermal storage device model is represented as follows:

[0087] ;

[0088] Let t be the thermal power stored in the thermal storage device at time t; The heat loss rate of the heat storage device; , These refer to the heat storage and heat release power of the heat storage device, respectively. , These refer to the heat storage and heat release efficiencies of the heat storage device, respectively. For time intervals.

[0089] The heating network pipeline model is represented as follows:

[0090] ;

[0091] The outlet temperature; Inlet temperature; , These are the pipeline outlet flow rate and inlet flow rate, respectively. The process ends at node i; Let's start from node i.

[0092] The heating station model is represented as follows:

[0093] ;

[0094] ;

[0095] This refers to the primary side fluid mass flow rate; The specific heat capacity of the primary fluid; , These are the primary water supply temperature and the return water temperature, respectively. The heat transfer coefficients are those between the primary and secondary fluids. This refers to the heat exchange area between the primary and secondary fluids. This represents the total mass of the primary fluid. , These are the secondary side supply water temperature and return water temperature, respectively. This refers to the secondary side fluid mass flow rate; The specific heat capacity of the secondary fluid; This represents the total mass of the secondary fluid.

[0096] Heat users on the load side refer to heat-consuming building objects, represented in the model as follows:

[0097] ;

[0098] , These are the inflow and outflow temperatures on the load side, respectively. For the inlet hot water mass flow rate of the building object; Specific heat capacity of hot water at the building entrance; The hourly heat load of the building object.

[0099] like Figure 2 As shown, in this embodiment, S2 specifically includes:

[0100] Set up a dispatch control center to manage the intelligent agent, input the heating operation data, heat load data and weather data collected by the digital twin model of the heating system into the big language model, and extract data features through the Transformer structure of the big language model to obtain the features of heating operation data, heat load data and weather data.

[0101] Multiple prompt templates are preset and stored. The similarity between the characteristics of heating operation data, heat load data, and weather data and the preset slot values ​​of the prompt templates is analyzed. The prompt template with the highest similarity is used as the target prompt template.

[0102] The slot values ​​of the target prompt template are updated based on the characteristics of heating operation data, heat load data, and weather data to generate heat load description text;

[0103] Keyword extraction is performed on the heat load description text. Based on the keywords, the characteristics affecting the heat load of the heating system are obtained and then input into a machine learning algorithm for processing. A heat load prediction model is then established to obtain the heat load demand data of the heating system in various future periods.

[0104] It should be noted that heating operation data includes pipeline network operation data between the heating system's heating stations and heat users, as well as heat metering data for heat users. This includes secondary side supply and return water temperatures, pressures, and flow rates, pump frequencies, valve openings, and indoor temperatures. Heat load data mainly consists of historical heat load data corresponding to different weather conditions and system operating conditions, including heat load variation curves, peak and valley values, and heat load variation rates. The heat load variation curve represents the heat load changing over time during historical heating periods (hourly, daily, weekly, or monthly timescales), allowing for an understanding of the volatility of heat load changes. Weather data mainly includes outdoor temperature, humidity, wind speed, and solar irradiance.

[0105] By combining large language models and machine learning algorithms for heat load demand forecasting, the advantages of each can be fully utilized to improve the accuracy and reliability of heat load demand forecasting results. Large language models can generate descriptions of heating operation data, heat load data, and weather data. These descriptions are related to changes in heating operation status, heat load, and weather data during historical heating periods. By analyzing this data, the accuracy of forecasting results can be improved.

[0106] Data feature extraction is performed using the Transformer structure of a large language model: the data is converted into a format acceptable to the model, such as combining numerical data with text descriptions to form text such as water supply temperature 80 degrees, return water temperature 60 degrees, flow rate 50 cubic meters / hour, heat load 500 megawatts, outdoor temperature 5 degrees, and relative humidity 60%. Then, the text is converted into word vectors through the model's embedding layer, and then features are extracted through a multi-layer Transformer encoder. In this process, the self-attention mechanism in the Transformer structure captures the correlation between different data, such as focusing on the relationship between outdoor temperature and heat load.

[0107] Example template 1: "Under [weather conditions], the heating system's supply water temperature is [supply water temperature value], return water temperature is [return water temperature value], flow rate is [flow rate value], current heat load consumption is [heat load value], and its main influencing factor is [influencing factor]."

[0108] Example template 2: "When the outdoor temperature is [outdoor temperature value] and the relative humidity is [humidity value], the relationship between the heating operation parameters [heating parameters] and the heat load [heat load condition] is [relationship description]."

[0109] When given input including heating operation data, heat load data, weather data, and prompts, the large language model first analyzes the input information to understand its meaning and context. BERT captures the complex relationships between words and their contexts through the self-attention mechanism in the Transformer model structure, thereby constructing bidirectional contextual representations that can capture deep semantics. These representations can better understand each word in a sentence, especially key entities and information in intent recognition and slot filling.

[0110] Large language models extract features from the input text, including vocabulary, syntactic structure, and semantic relationships. Based on the extracted features and the knowledge learned by the large language model, a probability is assigned to each possible output word or sentence, which can reflect the likelihood of the word or sentence appearing in the context. Appropriate words or sentences are then selected to generate descriptive text.

[0111] In practical applications, the machine learning algorithm used to establish the heat load prediction model is the BP neural network model. The specific process includes:

[0112] 1) Determine the number of layers, the number of nodes in each layer, the activation function, the initial weights, and the threshold of the BP neural network; The core of the BP neural network is to process the input signal from the input layer and hidden layer to the output layer through forward propagation. When the difference between the output result and the expected output value is greater than the limit value, back propagation will be performed to correct the network, and the weights and thresholds will be continuously adjusted to minimize the error.

[0113] 2) Before model training, preprocess the heat load data samples to check for missing or abnormal data, and then divide the sample data into training and test sets.

[0114] 3) Input the training set into the BP neural network model for training. During the model training process, the validation set will evaluate the training results. Adjust the parameters of the BP neural network model according to the evaluation results and continue training until the ideal output is achieved.

[0115] In this embodiment, S3 specifically includes:

[0116] The scheduling and control center management agent obtains the heat load demand data for each future time period output by the heat load prediction model, and performs statistical analysis to obtain the maximum, minimum, average, and rate of change of the heat load, thereby clarifying the trend and fluctuation range of the heat load.

[0117] Based on the changing trend, fluctuation range, and rate of change of heat load, as well as the actual operating characteristics of the system, the system is divided into multiple operating conditions, and clear boundary conditions, characteristic descriptions, and scheduling and control objectives are set for each operating condition.

[0118] In practical applications, the classification is based on the following criteria: considering the operating characteristics of the heating system equipment, such as when the heat load demand exceeds 350 MW, the heating equipment needs to operate at full load; when the heat load is between 300 and 350 MW, the equipment can operate stably with some adjustment range; when the heat load is below 300 MW, some equipment can operate at reduced power to save energy. The classification is combined with the trend and range of heat load changes. The numerical values ​​and ranges for heat load classification are set based on the actual operating status of the heating system, heat load demand, and the heating capacity of the units.

[0119] High-load operating conditions: The boundary condition is that the heat load demand is greater than 350 MWh and continues to rise. Characteristics include heating equipment operating near or at full load, high pipeline pressure, and rapid heat load growth. The scheduling and control objective is to ensure stable equipment operation to meet heat load demand, while simultaneously monitoring equipment operating status and preventing equipment failure.

[0120] Medium-load condition: The boundary condition is a heat load between 300 and 350 MWh, with a relatively stable rate of change. The characteristic description is that the heating equipment is operating normally with a certain adjustment margin. The scheduling and control objective is to optimize equipment operating parameters, improve heating efficiency, and reduce energy consumption.

[0121] Low-load operating condition: The boundary condition is a heat load below 300 MWh, with a gradual or decreasing rate of change. Characteristics include some heating equipment operating at reduced power and lower network pressure. The scheduling and control objective is to minimize the number of operating devices while meeting heat load demands, thus achieving energy-saving operation.

[0122] The scheduling and control of heating systems can usually be done at the hourly level. Therefore, the operating conditions can be divided according to the hourly scale. For example, the operating conditions of the next day can be divided based on the predicted heat load demand values ​​for different time periods of the next day, and the time periods of low, medium and high heat load demand can be identified.

[0123] In this embodiment, in step S4, it is determined whether the current system meets the conditions for entering the next operating condition. If it does, the corresponding operating condition is triggered, including:

[0124] The management agent of the scheduling and control center determines whether the system's operating conditions have changed based on the current system's operating data. If the operating conditions have changed, it matches the input data with various operating condition characteristics based on the learned knowledge of different operating condition characteristics, identifies the operating condition type, and switches the operating conditions according to the identified operating conditions, triggering the corresponding operating condition sub-agent to start working.

[0125] Among them, the operating condition sub-agent is created for each operating condition. Each operating condition sub-agent has independent simulation and scheduling control capabilities, and has a knowledge base and algorithm model related to that operating condition. Each operating condition sub-agent is also configured with relevant initial parameters. The scheduling control center management agent is responsible for coordinating the work between the operating condition sub-agents, switching operating conditions and triggering the work of each operating condition sub-agent. At the same time, each operating condition sub-agent feeds back the simulation and scheduling control effect of the heating system under its operating condition to the scheduling control center management agent.

[0126] It should be noted that creating the working condition sub-agent includes:

[0127] 1) Define the goals and functions of the intelligent agent.

[0128] Define objectives: Based on the operational requirements of the heating system, determine the objectives of the sub-agents for each operating condition. For example, the objective of the high-load sub-agent is to ensure stable operation of the equipment under high load and meet the heat load demand; the objective of the medium-load sub-agent is to optimize operating parameters and improve heating efficiency; and the objective of the low-load sub-agent is to achieve energy-saving operation.

[0129] Define functions: Clearly define the functions that the intelligent agent needs to possess. For example, data monitoring function, used to acquire real-time operating data and heat load data of heating equipment; analysis and decision-making function, to formulate reasonable scheduling and control strategies based on data and operating conditions; command execution function, to send control commands to heating equipment, etc.

[0130] 2) Choosing a development platform and technology

[0131] Platform selection: Choose a suitable development platform based on the actual needs and technical capabilities of the heating system;

[0132] Technology Determination: A large language model is adopted as the core language processing and decision engine of the intelligent agent, combined with machine learning algorithms and intelligent optimization algorithms for data processing and analysis decision-making; among them, machine learning algorithms include BP, LSTM, and CNN, and intelligent optimization algorithms include particle swarm optimization and genetic optimization algorithms. The optimal algorithm is determined by comparative analysis based on the actual calculation results of the algorithms.

[0133] 3) Design the intelligent agent architecture

[0134] Input module: Receives data information transmitted by the management agent of the scheduling control center and acquires relevant scheduling and control data from the system;

[0135] Processing module: Constructs a knowledge graph for the scheduling and control of the heating system based on large language models and machine learning algorithms, providing a basis for the decision-making of intelligent agents;

[0136] Output module: Based on the decision results of the processing module, output scheduling and control instructions for relevant equipment on the source, grid and load sides, such as controlling the start and stop of heating units, adjusting the operating power of heating units, adjusting the adjustment parameters of pumps and valves in heating stations, etc., and output system operation status reports and other information.

[0137] 4) Configure the knowledge base and algorithm model

[0138] Build a knowledge base: Collect knowledge and information related to the scheduling and control of this operating condition, organize this knowledge into a structured or semi-structured data format, and store it in the knowledge base so that the corresponding knowledge graph can be built later and provided to the operating condition sub-intelligent agent for quick query and call;

[0139] Configure the algorithm model: Select and configure the appropriate algorithm model according to the functional requirements of the agent.

[0140] In this embodiment, step S4, predicting the heat output of the new energy heating unit in future time periods, includes:

[0141] Based on the historical operating data, heat output, and meteorological data of the new energy heating units, key data affecting the heat output of the new energy heating units are extracted, and prompt word templates are constructed based on the key data.

[0142] Obtain structured historical operating data, heat production, and meteorological data; extract key data; map them with the corresponding key data in the prompt word template; and construct a text description of the structured data using natural language to serve as a textualized unit heat production dataset.

[0143] A large language model is selected as the basic model for fine-tuning, and the textualized unit heat production dataset is used as the training data for supervised fine-tuning. Supervised fine-tuning training is performed on the large language model so that it learns the data knowledge of new energy unit heat production.

[0144] By using natural language, the data features affecting the heat production of new energy heating units are input into a large language model that is fine-tuned and trained, and the predicted heat production values ​​of new energy heating units for each future time period are output.

[0145] In this embodiment, when performing supervised fine-tuning training on the large language model, the low-rank adaptive LoRA technique is used. For the original weight matrix in the training model, two low-rank matrices A and B are introduced to reduce the number of trainable parameters, as shown below:

[0146] ;

[0147] This is the original weight matrix of the model; The scaling factor is A; A and B are low-rank matrices. This is the updated model weight matrix;

[0148] The gradient descent algorithm is used to calculate the loss function. With respect to the gradients of A and B, update the low-rank matrices A and B, as follows:

[0149] ;

[0150] ;

[0151] , This is the updated low-rank matrix; The learning rate; , Loss functions The gradients of A and B.

[0152] When using the LoRA technique to fine-tune large language models, a low-rank structure is introduced into the model's weight matrix to reduce the number of trainable parameters, thereby achieving effective fine-tuning of the model without significantly increasing computational costs.

[0153] like Figure 3 As shown, in this embodiment, before obtaining the scheduling and control strategies for each device on the source, network, and load sides of the heating system in step S5, it is necessary to establish a system optimization scheduling and control knowledge graph based on a large language model, specifically including:

[0154] At the data layer: The created sub-agents collect historical operating data, scheduling and control strategies, operating objectives, and text data of scheduling and control rules of each device on the source, grid, and load sides under their respective operating conditions, and perform data preprocessing.

[0155] Large Language Model Training: Select a large language model, convert the preprocessed text data into an input format acceptable to the large language model, and use the text data to conduct supervised training on the large language model, so that the model can learn the relationship between heat load demand, heat output and operating conditions of new energy heating units, scheduling and control strategies, and operating objectives, so that the model can understand and generate knowledge related to system optimization scheduling and control.

[0156] At the conceptual level: large language models are used to mine and classify entities and relationships in text data, as well as to annotate the corpus;

[0157] At the instance layer: knowledge extraction and instantiation are performed using a large language model;

[0158] Based on the entities, relations, and corpus annotations at the concept layer, the ontology structure of the knowledge graph is defined. Based on the knowledge extraction and instantiation at the instance layer, specific system scheduling and control knowledge graph instances are formed, and a system optimization scheduling and control knowledge graph based on a large language model is established.

[0159] In this embodiment, the historical operating data includes heat load demand, heat output of new energy heating units, operating parameters of conventional heating units, operating parameters of heating stations, and heat metering data of heat users; the scheduling and control strategy includes the operating output of source-side new energy heating units and conventional heating units, the heat storage and release status and heat storage and release status of heat storage devices, as well as the pump and valve adjustment parameters of grid-side heating stations and the pump and valve adjustment parameters of load-side heat users at the building entrance and household entrance; the operating objectives include economic objectives, energy-saving objectives, environmental protection objectives, and comfort objectives; the scheduling and control rules include the scheduling and control logic principles of each device on the source, grid, and load sides and the scheduling and control coupling between devices.

[0160] In this embodiment, in the concept layer, a large language model is used to obtain all nouns in the text data. After identifying the entity objects, prompt words are introduced for corpus clustering and classification. The large language model and machine learning algorithms are used to mine the relationships between entities from the text data. In the instance layer, the large language model is used to extract scheduling control computational knowledge rules from the text data based on the ontology structure and corpus annotations. After instantiating them into entity-relationship-entity triples, a graph database is called to store and visualize the triples generated by the large language model.

[0161] It should be noted that the large language model used to build the knowledge graph is DeepSeek. Specifically, when mining relationships between entities from text data, the large language model used is DeepSeek, and the machine learning algorithm is a convolutional neural network (CNN). The loaded DeepSeek model is used to extract entity relationships from the text. By inputting specific prompts into the model, it is guided to output entity and relationship information. The relationships mined by the large language model are verified using a trained CNN. The relational text is input into the CNN model, and the model's output is used to determine whether the relationship is correct. Specifically, the embedding layer in the CNN converts word indices into word vectors, a one-dimensional convolutional layer performs convolution operations to extract text features, a global max pooling layer pools the convolution results to reduce dimensionality, two fully connected layers are used, and the last layer uses an activation function for classification.

[0162] In practical applications, the conceptual layer implementation process includes:

[0163] Noun Extraction: The DeepSeeK API is called, taking text data as input, with the prompt "Please extract all nouns from the following text". For example, if the text is "Under high load conditions, the operating power of the water pump in the heating system needs to be increased to meet the heat load demand", the model returns a list of nouns: ["Heating system", "High load conditions", "Water pump", "Operating power", "Heat load demand"].

[0164] Corpus Clustering and Classification: Introduce prompt words, such as "Classify the following nouns according to heating equipment, operating conditions, and heat load relevance." Input the extracted nouns into a large language model, which classifies them based on semantic understanding. For example, the classification results might be heating equipment ["water pump"], operating conditions ["high load condition"], and heat load relevance ["operating power", "heat load demand"]. Here, the K-Means machine learning clustering algorithm can also be used to assist the DeepSeeK large model in achieving more accurate clustering. First, convert the nouns into word vectors (e.g., using Word2Vec), then use the K-Means algorithm to cluster the word vectors, and finally use the clustering results as a reference to perform the final classification using DeepSeeK.

[0165] Relationship mining: Utilizing large language models and machine learning algorithms to uncover relationships between entities. For example, dependency parsing tools can be used to perform syntactic analysis on text to obtain grammatical dependencies between words. Combining the large language model DeepSeeK, given the input prompt "Analyze the relationship between the following nouns: heating system, water pump, heat load demand," the model, based on the syntactic analysis results and its own semantic understanding capabilities, outputs the relationship "The heating system includes water pumps, and the water pumps operate to meet the heat load demand." Alternatively, convolutional neural networks (CNNs) can be used for feature extraction and relationship classification to further validate and supplement the relationships mined by the large language model.

[0166] The instance layer implementation process includes:

[0167] Knowledge rule extraction: The DeepSeeK large language model is used to extract scheduling control computation knowledge rules based on ontology structure and corpus annotation. For example, the ontology structure defines concepts and relationships such as "equipment", "operating parameters", and "control strategy". Given the text "When the heat load exceeds the set threshold, increase the water pump operating power by 20%", combined with the ontology structure and corpus annotation (annotating heat load as "operating parameters", water pump as "equipment", and increasing operating power as "control strategy"), the DeepSeeK large language model extracts the knowledge rule "When the operating parameters (heat load) exceed the set threshold, execute the control strategy (increase operating power by 20%) on the equipment (water pump)".

[0168] Triple Instantiation and Storage: Extracted knowledge rules are instantiated as entity-relation-entity triples. For example, the above example could be instantiated as (heat load, exceeded, set threshold), (water pump, executed, increase operating power by 20%). A graphical database (such as Neo4j) is then used for storage. For visualization, Neo4j provides a graphical interface that displays entities and relations within the triples through nodes and edges, allowing users to intuitively see the structure and relationships of scheduling control knowledge.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0170] Furthermore, the functional modules in the various embodiments of this invention 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 function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the steps of the methods in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0171] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A large model and multi-agent based multi-working condition scheduling control method for a heating system, characterized in that, It comprises: S1, a digital twin model of a heating system including a source-side new energy heating unit, a conventional heating unit, a heat storage device, a network-side heat station and a load-side heat user is established; S2, a large language model is used to generate a heat load description text according to the multi-dimensional data collected by the digital twin model of the heating system, and then a heat load prediction model is used to predict the heat load demand of the heating system in each future period; S3, according to the heat load demand prediction results of each future period, the system is divided into multiple operating conditions, including: The dispatching control center management agent obtains the future heat load demand data output by the heat load prediction model, and performs statistical analysis to obtain the maximum, minimum, average and change rate of the heat load, and clearly determines the change trend and fluctuation range of the heat load; According to the change trend, fluctuation range, change rate and actual operation characteristics of the system, the system is divided into multiple operating conditions, and each operating condition is set with clear boundary conditions, feature description and dispatching control target; S4, judge whether the current system meets the condition of entering the next operating condition, if yes, trigger the corresponding operating condition, and obtain the heat load demand value of each future period and the heat production of the new energy heating unit in each future period; In S4, if the condition of entering the next operating condition is met, the corresponding operating condition is triggered, including: The dispatching control center management agent determines whether the system operating condition changes according to the current system operating data, if the operating condition changes, the input data is matched with various operating condition characteristics according to the learned operating condition characteristic knowledge, the operating condition type is identified, and the operating condition is switched according to the identified operating condition, and the corresponding operating condition sub-agent starts to work; Among them, the operating condition sub-agent is created for each operating condition, each operating condition sub-agent has independent simulation and dispatching control capability, and has a knowledge base and algorithm model related to the operating condition, and each operating condition sub-agent is configured with related initial parameters; The dispatching control center management agent is responsible for coordinating the work between the operating condition sub-agents, switching the operating conditions and triggering the work of the operating condition sub-agents, and the operating condition sub-agents feed back the simulation and dispatching control effect of the heating system in their operating conditions to the dispatching control center management agent; S5, a large language model and a preset agent are used to analyze the heat load demand value and the heat production of the new energy heating unit, and combined with the historical operating data of each operating condition, the set operating target and the mining of the dispatching control rules of the source, network and load side equipment, the dispatching control strategy of the source, network and load side equipment of the heating system is obtained, and is fed back to the digital twin model for guiding the actual operation management of the heating system.

2. The method of claim 1, wherein, S1 specifically comprises: Install corresponding sensors on the source-side new energy heating unit, conventional heating unit, heat storage device, network-side heat station and load-side heat user end to obtain actual unit operating parameters, heat storage state information, heat supply parameters of heat station inlet and outlet and heat user heat consumption data; Establishing a source-side new energy heat supply unit model, a conventional heat supply unit model, a heat storage device model, a network-side heat station model, a heat network pipeline model, and a load-side heat user model; Determining the interface relationship between the source-side, network-side, and load-side models, and clarifying the content and method of data interaction; Integrating the source-side, network-side, and load-side models to form a complete digital twin model of the heat supply system; Simulating and verifying the digital twin model by setting different operating conditions and parameters, and using actual data obtained to adjust and optimize the model parameters.

3. The method of claim 1, wherein, The S2 specifically includes: Setting a scheduling control center to manage the intelligent agent, inputting the heat supply operation data, heat load data, and weather data collected by the digital twin model of the heat supply system into a large language model, extracting data features through the Transformer structure of the large language model, and obtaining heat supply operation data features, heat load data features, and weather data features; Pre-setting and storing multiple prompt templates, analyzing the similarity between the heat supply operation data features, heat load data features, and weather data features and the pre-set slot values of the prompt templates, and taking the prompt template with the maximum similarity as the target prompt template; Updating the slot values of the target prompt template according to the heat supply operation data features, heat load data features, and weather data features, and generating a heat load description text; Extracting keywords from the heat load description text, obtaining features affecting the heat load of the heat supply system according to the keywords, and inputting them into a machine learning algorithm for processing to establish a heat load prediction model and obtain heat load demand data of the heat supply system in future periods.

4. The method of claim 1, wherein, In the S4, predicting the heat output of new energy heat supply units in future periods includes: Extracting key data affecting the heat output of new energy heat supply units from historical operation data, heat output, and meteorological data of the new energy heat supply units, and constructing a prompt word template according to the key data; Obtaining structured historical operation data, heat output, and meteorological data, extracting key data, and mapping them with corresponding key data in the prompt word template, and constructing a text description of the structured data in a natural language manner as a textual unit heat data set; Selecting a large language model as a basic model for fine-tuning, and using the textual unit heat data set as supervised fine-tuning training data to perform supervised fine-tuning training on the large language model, so that the large language model learns the data knowledge of new energy unit heat output; Inputting the data features affecting the heat output of new energy heat supply units into the fine-tuned large language model in a natural language manner, and outputting the heat output prediction value of new energy heat supply units in future periods.

5. The method of claim 4, wherein, When performing supervised fine-tuning training on the large language model, the low-rank adaptive LoRA technology is used to reduce the number of trainable parameters by introducing two low-rank matrices A and B to the original weight matrix in the training model, represented as: ； is the original weight matrix of the model; is the scaling coefficient; A and B are low-rank matrices; is the updated weight matrix of the model; using a gradient descent algorithm by computing a loss function With respect to the gradients of A and B, the low-rank matrices A and B are updated, denoted as: ； ； , is an updated low-rank matrix; is a learning rate; , are loss functions gradients with respect to A and B.

6. The method of claim 1, wherein, Before obtaining the scheduling control strategy of the source, network, and load sides of the heat supply system in the S5, a system optimization scheduling control knowledge graph based on a large language model needs to be established, specifically including: In the data layer: collect the historical operation data of each device on the source network load side under the respective working conditions, the text data of the scheduling control strategy, the operation target and the scheduling control rule through the created working condition sub-agent, and perform data preprocessing; Large language model training: select a large language model, convert the preprocessed text data into an input format acceptable to the large language model, use the text data to supervise the training of the large language model, let the model learn the relationship between the heat load demand, the heat output of the new energy heating unit, the operation condition, the scheduling control strategy and the operation target, and make the model understand and generate the knowledge related to the system optimization scheduling control; In the concept layer: use the large language model to mine and classify entities and relationships in the text data, and perform corpus annotation; In the instance layer: use the large language model to extract and instantiate knowledge; According to the entities, relationships and corpus annotation in the concept layer, define the ontology structure of the knowledge graph, and according to the knowledge extraction and instantiation in the instance layer, form specific system scheduling control knowledge graph instances, and establish a system optimization scheduling control knowledge graph based on the large language model.

7. The method of claim 6, wherein, The historical operation data includes heat load demand values, heat output of new energy heating units, operation parameters of conventional heating units, operation parameters of heat stations and heat metering data of heat users; the scheduling control strategy includes the operation output of source side new energy heating units and conventional heating units, and the heat storage and release state and amount of heat storage devices, as well as the pump and valve adjustment parameters of the network side heat station and the building front and house front pump and valve adjustment parameters of the load side heat users; the operation target includes economic target, energy saving target, environmental protection target and comfort target; the scheduling control rule includes the scheduling control logic principle of each device on the source network load side and the scheduling control coupling between devices.

8. The method of claim 6, wherein, In the concept layer, all nouns in the text data are obtained by using the large language model, the entity objects are determined, the prompt words are introduced for corpus clustering and classification, and the relationships between entities are mined from the text data by using the large language model and machine learning algorithm; in the instance layer, the scheduling control calculation knowledge rule is extracted from the text data according to the ontology structure and corpus annotation by using the large language model, and after instantiation as entity-relation-entity triple, the triple generated by the large language model is stored and visualized by calling the graph database.

Images