Multi-zone air conditioner control method and system based on iterative simulation and fine-tuning of large model
By using iterative simulation and fine-tuning of large models, the adaptability problem of traditional HVAC control methods in dynamic environments was solved, enabling safe and reliable control of multi-zone air conditioning systems, reducing temperature fluctuations and energy consumption risks, and ensuring the system's adaptability and sustainability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG JIANZHU UNIV
- Filing Date
- 2026-05-13
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional HVAC control methods lack adaptability. Control strategies based on fixed rules or classic PID control are difficult to cope with dynamic environmental changes. Reinforcement learning-based methods have limited ability to extract and understand information features in complex environments and lack an immediate physical verification mechanism, which leads to the risk of multi-zone air conditioning systems going out of control in real environments.
The method of iterative simulation and fine-tuning of large models is adopted. The thermal parameters are calibrated by iterative simulation model, and the control strategy is generated by reinforcement learning model. Structured prompt words are constructed in large model. The control strategy is verified by calibrated simulation model, and the feedback of failure results is used for correction, forming a closed-loop iterative optimization.
It enables safe and reliable control of multi-zone air conditioning systems in real-world environments, reduces temperature fluctuations and energy consumption risks, and ensures the system's adaptability and sustainability.
Smart Images
Figure CN122170510A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and building energy systems, and in particular to a multi-zone air conditioning control method and system based on iterative simulation and fine-tuning of large models. Background Technology
[0002] Traditional HVAC control methods lack adaptability. Control strategies based on fixed rules or classic PID control struggle to cope with dynamic environmental information (personnel movement, changes in solar radiation) and complex climatic conditions. While traditional reinforcement learning (RL) methods can optimize strategies through interactive learning, their ability to extract and understand features of complex, high-dimensional environmental information is limited. Furthermore, the training process relies on extensive and insecure online exploration, resulting in a gap in the transfer of simulation to real-world operation.
[0003] In recent years, large language models have demonstrated powerful semantic understanding and reasoning planning capabilities, providing a new paradigm for decision-making in complex systems. However, when directly applied to multi-zone air conditioning control, the unpredictable nature of the control commands generated by large models and the lack of physical constraints in their output space may lead to risks such as frequent equipment start-ups and shutdowns, indoor temperature fluctuations, or surges in energy consumption. Existing technologies lack a mechanism for real-time physical verification of model outputs, making it impossible to assess their safety before command execution, resulting in the risk of system loss of control in real-world environments. Summary of the Invention
[0004] The purpose of this invention is to provide a multi-zone air conditioning control method and system based on iterative simulation and fine-tuning of a large model, which provides an intelligent, safe and continuously iteratively updated adaptive optimization control scheme for multi-zone air conditioning control.
[0005] To address the aforementioned technical problems, this invention provides a multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model, comprising: Obtain current environmental information and use the reinforcement learning model trained by iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building. Based on the first control strategy and current environmental information, a prompt word is constructed; the prompt word is input into the fine-tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building; The second control strategy is verified using the calibrated simulation model. The operation of the air conditioning equipment in the target building's multi-zone system is controlled according to the verified second control strategy. For the second control strategy that fails verification, the verification result is fed back to the fine-tuned large model in the form of feedback prompts for correction. Collect actual operating data of air conditioning in multiple areas of a real building after control execution, add it to historical operating data, and return to the iterative simulation model parameter calibration step for repeated execution.
[0006] Optionally, historical operational data of real buildings is obtained, and the parameters of the iterative simulation model are calibrated using the historical operational data to obtain a calibrated iterative simulation model; a reinforcement learning model is trained using the calibrated iterative simulation model; the parameter calibration of the iterative simulation model using the historical operational data specifically includes: Acquire historical operational data of a real building, including environmental information, current indoor temperature, executed control actions, and measured indoor temperature after execution; The environmental information, current indoor temperature, and control actions from the historical operating data are input into the iterative simulation model, and the indoor temperature at the next moment is calculated by the state-space equation of the iterative simulation model. With the goal of minimizing the mean square error between the indoor temperature predicted by the simulation at the next moment and the measured indoor temperature, the gradient descent method is used to update the thermal parameter matrix in the iterative simulation model, so that the thermodynamic characteristics of the iterative simulation model can approximate the actual thermodynamic characteristics of the real building.
[0007] Optionally, obtain current environmental information and use the reinforcement learning model trained through iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building, including: The acquired current environment information is input into the reinforcement learning model trained by iterative simulation to obtain the control strategy recommended by the reinforcement learning model under the current environment information, and the control action recommended by the reinforcement learning model is used as the first control strategy. The reinforcement learning model is a model that is pre-trained through interactive training with the iterative simulation model and contains recommended control strategies under different environmental information.
[0008] Optionally, prompt words are constructed based on the first control strategy and current environmental information; the prompt words are input into the fine-tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building; specifically: Obtain environmental information of the target building, the current indoor temperature, and pre-defined role definitions, task requirements, and output formats; The environmental information is input into the reinforcement learning model to obtain the first control policy output by the reinforcement learning model. Based on environmental information, current indoor temperature, role definition, task requirements, output format, and first control strategy, prompt words are constructed and input into the finely tuned large model. The finely tuned large model outputs a new multi-zone air conditioning control action group, and the new multi-zone air conditioning control action group is determined as the second control strategy for the multi-zone air conditioning of the target building.
[0009] Optionally, the second control strategy is verified using a calibrated iterative simulation model, and the operation of the multi-zone air conditioning equipment in the target building is controlled according to the verified second control strategy; including: The second control strategy is input into the calibrated iterative simulation model. Based on the current state of the target building and the second control strategy, the iterative simulation model calculates the operating parameters of the multi-zone air conditioning of the target building after the second control strategy is executed. The operating parameters of the multi-zone air conditioning in the target building after executing the second control strategy are compared with the preset target operating parameters to obtain the deviation value; it is then determined whether the deviation value is within the range of the first preset threshold to the second preset threshold. If yes, the second control strategy is verified, and the operation of the air conditioning equipment in the target building's multi-zone system is controlled according to the verified second control strategy; if no, the second control strategy is not verified.
[0010] Optionally, the feedback prompt is the reason why the second control strategy failed verification; the feedback prompt is composed of role definition, task requirements, output format, environmental information, current indoor temperature, room air conditioning control action group, and feedback suggestion.
[0011] Optionally, the environmental information includes: human heat output, ground temperature, outdoor temperature, and solar radiation intensity.
[0012] To address the above problems, this invention also provides a multi-zone air conditioning control system based on iterative simulation and fine-tuning of a large model, comprising: The first control strategy generation module acquires current environmental information and uses a reinforcement learning model trained by iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building. The second control strategy generation module constructs prompt words based on the first control strategy and current environmental information; inputs the prompt words into the fine-tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building; The verification module verifies the second control strategy using the calibrated simulation model, and controls the operation of the air conditioning equipment in the target building's multi-zone system based on the verified second control strategy; for the second control strategy that fails verification, the verification result is fed back to the fine-tuned large model in the form of feedback prompts for correction. The closed-loop iterative module collects actual operating data of air conditioning in multiple areas of the real building after control execution, adds it to historical operating data, and returns the iterative simulation model parameter calibration steps for repeated execution.
[0013] To address the aforementioned technical problems, this invention also provides a multi-zone air conditioning control device based on iterative simulation and fine-tuning of a large model, comprising: A memory for storing instructions, the instructions including the steps of the multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model as described in any of the above-mentioned methods; A processor for executing the instructions.
[0014] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model as described in any of the preceding claims.
[0015] The above technical solution has the following advantages or beneficial effects: To address the issue that in multi-zone air conditioning control scenarios, the unpredictable and physically constrained room air conditioning control actions generated by large-scale models can lead to deviations from target indoor temperatures, imbalances in regional comfort, or abnormal energy consumption, this invention implements a one-step rapid pre-run of each set of room air conditioning control actions generated by the large-scale model before actual execution. The safety is assessed by combining the deviation between the predicted and target temperatures in each region. Control actions that fail verification are intercepted, and structured feedback prompts are generated to guide the fine-tuned large-scale model to be regenerated, thus balancing temperature comfort and energy consumption in multi-zone air conditioning control. This mechanism provides verifiable, interventionist, and error-correctable operational safeguards for the deployment of large-scale models in multi-zone air conditioning control scenarios, reducing the risk of temperature fluctuations and increased energy consumption due to improper control actions. Attached Figure Description
[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0017] Figure 1 A flowchart of a multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model provided in an embodiment of the present invention; Figure 2 This is a control process diagram of a multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of a multiphysics simulation environment for multi-zone air conditioning control provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of PPO model training for multi-zone air conditioning control provided in an embodiment of the present invention. Figure 5 This is a flowchart of the PPO model training process for multi-zone air conditioning control provided in an embodiment of the present invention. Figure 6This is a simplified schematic diagram of fine-tuning a large model for multi-zone air conditioning control provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of PPO_LoRA supervised fine-tuning for multi-zone air conditioning control provided in an embodiment of the present invention. Detailed Implementation
[0018] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0019] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0020] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.
[0021] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.
[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0023] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.
[0024] Example 1 Multi-zone air conditioning: The target building contains multiple zones, each with an independent indoor temperature sensor and independently controllable air conditioning equipment, with heat exchange between zones through the building envelope.
[0025] This embodiment takes multi-zone air conditioning as an example. In the energy-saving optimization control method of multi-zone air conditioning equipment, when the large model (LLM) is directly applied to the actual control task, many problems will be encountered, resulting in the output of multi-zone air conditioning failing to simultaneously meet comfort and energy saving.
[0026] While Large General Models (LLMs) possess semantic understanding capabilities, they cannot grasp specialized building thermodynamics knowledge and energy-saving optimization strategies solely through contextual learning. Existing fine-tuning methods face a dilemma: full-parameter fine-tuning of models with billions of parameters is computationally extremely costly and requires a large amount of labeled data; while parameter-efficient fine-tuning methods often employ simple behavioral imitation (such as supervised learning), which can only reproduce actions from examples and cannot effectively inject long-term optimization goals and decision preferences from reinforcement learning into large models, making it difficult for the model to learn strategies that surpass those of the training data.
[0027] Furthermore, current physical simulation environments in the building HVAC field have significant technical limitations: one type of simulation model oversimplifies building thermodynamic processes, using only fixed parameters or empirical formulas to construct simulation logic, making it difficult to realistically reproduce the dynamic thermal changes of actual buildings; another type of simulation environment is built using a data-driven approach, which, while possessing high simulation accuracy, does not significantly affect human perception of minute indoor temperature fluctuations (such as temperature adjustments within a 0.5℃ range). Frequent interactive training between this type of high-precision simulation environment and large language models significantly increases computational complexity and processing time, failing to meet the efficient interactive requirements of large models during HVAC control strategy training. More importantly, both types of simulation environments lack a linkage calibration mechanism with the real environment. Once deployed, they become disconnected from the physical world, unable to track thermodynamic drift caused by factors such as building envelope aging and changes in usage patterns. This results in upper-level control strategies being trained in a gradually distorted "static island," ultimately making it difficult to migrate to real-world scenarios.
[0028] The control commands generated by large models are unpredictable, and their output space lacks physical constraints. Directly applying them to multi-zone air conditioning systems may lead to risks such as frequent equipment start-ups and shutdowns, indoor temperature fluctuations, or surges in energy consumption. Existing technologies lack a mechanism for real-time physical verification of model outputs, making it impossible to assess their safety before command execution, resulting in the risk of the system going out of control in real-world environments.
[0029] It is evident that existing technologies separate the simulation environment, control strategy, and real physical environment: the simulation environment cannot evolve with the real environment, the control strategy cannot efficiently learn from the simulation and transfer to the real scenario, and the model output lacks physical safety constraints. This results in the system facing multiple technical bottlenecks throughout its entire lifecycle, from initial construction to long-term operation, including simulation distortion, difficulty in strategy transfer, unsafe control, and inability to adapt to environmental changes. Therefore, this invention proposes a multi-zone air conditioning control method that can solve the above-mentioned technical problems.
[0030] like Figure 1-2 As shown, the multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model provided in this embodiment of the invention includes: Taking three rooms in a company as an example, the air conditioning in rooms 1, 2 and 3 is controlled by a multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model, so as to achieve both comfort and energy saving.
[0031] S101: Obtain historical operating data of the real building, use the historical operating data to calibrate the parameters of the iterative simulation model to obtain the calibrated iterative simulation model; train the reinforcement learning model with the calibrated iterative simulation model; use the historical operating data to calibrate the parameters of the iterative simulation model.
[0032] First, historical operational data of the real building is acquired. This historical operational data includes environmental information, the current indoor temperature, executed control actions, and the measured indoor temperature after execution. The environmental information, current indoor temperature, and control actions from the historical operational data are input into the iterative simulation model. The simulated predicted indoor temperature for the next moment is calculated using the state-space equations of the iterative simulation model. With the optimization objective of minimizing the mean square error between the simulated predicted indoor temperature and the measured indoor temperature, the gradient descent method is used to update the thermal parameter matrix in the iterative simulation model, making the thermodynamic characteristics of the iterative simulation model approximate the actual thermodynamic characteristics of the real building. Specifically: S1011: Construct an iterative simulation model. Based on the basic environmental information of the target building (personnel heat generation, outdoor temperature, solar radiation, ground temperature), set the initial parameter matrix of the state-space equations: A matrix: Thermal conductivity coefficients between rooms; B matrix: The influence coefficient of air conditioning control actions on temperature; C matrix: The influence coefficient of external disturbances (personal heat generation, outdoor temperature, solar radiation, ground temperature) on room temperature.
[0033] S1012: Collect historical operational data of the actual building. Collect real operational data for the past week using sensors or data platforms deployed in the target building; the sensors include temperature sensors, power meters, personnel sensors, etc.; each data point includes: environmental information, initial indoor temperature, actual executed control action, and indoor temperature after execution. Environmental information includes: ground temperature, outdoor temperature, solar radiation, and personnel heat output. S1013: Calibrate iterative simulation model parameters based on historical data. With the optimization objective of minimizing the mean square error between the simulated predicted temperature and the actual observed temperature, the gradient descent method is used to iteratively update the B matrix (matrices A and C can be updated synchronously as needed): For each historical sample, calculate the simulated predicted temperature:
[0034] in, This represents the current room temperature, with the initial room temperature set at 18℃; A represents the thermal conductivity between rooms, for example: The first row represents room 1, the second row represents room 2, and the third row represents room 3. The diagonal elements represent the room's temperature retention capacity; the closer the value is to 1, the better the heat retention. The numbers following the diagonal elements represent the temperature exchange relationship with rooms 2 and 3, respectively; the larger the value, the greater the influence of that temperature on the temperature of room 1. The meanings of the second and third rows are similar to those of the first row, and will not be repeated here.
[0035] This represents the effect of air conditioning control actions on temperature (the matrix dimension is 3×3, calibrated according to the actual effect of the air conditioner).
[0036] This represents the current air conditioning control actions for the three rooms (matrix dimension 3×1, [air conditioning control actions for room 1, air conditioning control actions for room 2, air conditioning control actions for room 3]).
[0037] This represents the influence coefficient of external disturbances on room temperature; the matrix is 3×4, with rows corresponding to rooms and columns corresponding to the influence coefficients of four types of environmental information on each room: ([personal heat generation power], [ground temperature], [outdoor temperature], [solar radiation intensity]).
[0038] The matrix represents the specific value of the impact of external disturbances on room temperature (4*1), and its elements are ([personal heat output], [ground temperature], [outdoor temperature], [solar radiation intensity]), which are read from data collected from real buildings.
[0039] also, , and The value of changes as the time step of the multiphysics simulation environment progresses.
[0040] Calculate the loss function:
[0041] in, This refers to the actual temperature at the next moment.
[0042] Calculate the gradient for each element Bi,j in matrix B:
[0043] The parameters are updated as follows:
[0044] Where η is the learning rate.
[0045] Repeat the iterations until the loss function converges or the preset number of iterations is reached, and complete the parameter calibration.
[0046] S1014: Save the calibrated iterative simulation model as the multiphysics simulation environment for subsequent steps (S102 and beyond).
[0047] S1015: Establish an iterative simulation model continuous calibration mechanism. During the actual operation of the system, repeat steps S1002 to S1004 every fixed period (3 hours) to recalibrate the iterative simulation model using newly collected real operating data, so that the simulation environment always closely approximates the current thermodynamic characteristics of the real building.
[0048] (1) Real-time operation data collection and storage. During the actual control of building equipment, the system continuously collects the following data: environmental information, actual indoor temperature of each room, executed control actions (air conditioning power adjustment value), equipment energy consumption, etc. Each data point is accompanied by a timestamp, forming a historical operation database.
[0049] (2) Iterative simulation model calibration. The calibration process is triggered at a preset cycle (3 hours): Extract real operational data for the current period from the historical operational database; The extracted environmental information is input into the current iterative simulation model, and the physics engine is run to obtain the simulated predicted indoor temperature. With the goal of minimizing the mean square error (MSE) between the simulated predicted temperature and the actual collected temperature, gradient descent is used to update the key thermal parameters in the simulation model, mainly including: the influence coefficient of air conditioning control action on temperature in the state space equation B matrix, the heat transfer coefficient between rooms in matrix A, and the influence coefficient of external disturbance in matrix C. The updated parameters are saved to obtain the calibrated new iterative simulation model.
[0050] Furthermore, the calibrated iterative simulation model is used as an interactive environment to train a reinforcement learning model and generate an interactive dataset of multi-zone air conditioning in the target building. The interactive dataset of multi-zone air conditioning in the target building is constructed as follows: like Figure 4-5As shown, during the training phase, the multiphysics simulation environment outputs real-world environmental information to the PPO model to be trained. The PPO model generates control actions and sends them to the multiphysics simulation environment. The multiphysics simulation environment calculates and executes these control actions to obtain the indoor temperature and reward value. Then, it inputs the next set of real-world environmental information into the PPO model to be trained. This continuous interaction ultimately preserves the interactive dataset of the target building's multi-zone air conditioning and the trained PPO model. Specifically: Step 1: The multiphysics simulation environment acquires environmental information of Room 1, Room 2, and Room 3. The PPO model to be trained receives the current environmental information of Room 1, Room 2, and Room 3 transmitted from the multiphysics simulation environment interaction interface.
[0051] Step 2: The PPO model to be trained outputs the air conditioning control actions for rooms 1, 2, and 3 respectively, based on the environmental information of rooms 1, 2, and 3 at the current time. The PPO model to be trained will randomly output a 1*3 array to the interactive interface of the multiphysics simulation environment. For example, ([0.2], [0.3], [0.2]) indicates that the air conditioning control action for room 1 is 0.2, the air conditioning control action for room 2 is 0.3, and the air conditioning control action for room 3 is 0.2. The numerical limit for each control action is [-1, 1], where -1 indicates that the air conditioning cooling power is adjusted to the maximum, and 1 indicates that the air conditioning heating power is adjusted to the maximum. The specific power value conversion method is: value * air conditioning power. For example, a control action of 0.5 means that when the air conditioning power is 1400W, the heating power is adjusted to 700W.
[0052] Step 3: Calculate the reward value for executing the air conditioning control action through the multiphysics simulation environment; the PPO model to be trained receives the indoor temperature after executing the air conditioning control action and the reward value after executing the air conditioning control action from the interactive interface of the multiphysics simulation environment, and saves them in the format of [environmental information, control action, reward value, indoor temperature after executing the air conditioning control action].
[0053] Step 4: Determine if the current step count has reached the number of steps set in the parameters. If not, repeat steps 1-3. After reaching the number of steps set in the parameters, calculate the saved information and obtain the advantage value using the default formula for near-end policy optimization. Update the policy network of the PPO model to be trained based on the advantage value.
[0054] Step 5: Determine if the current step count has reached the total number of steps set in the parameters. If not, repeat steps 1-4. After reaching the total number of steps set in the parameters, save the trained PPO model; and save the interaction data as an array of [current step count, environmental information, room air conditioning control action group, reward value, indoor temperature after the action is executed].
[0055] In this embodiment, the number of steps is set to 2048, and the total number of steps is set to 500000.
[0056] The PPO model to be trained interacts with a multiphysics simulation environment, enabling it to learn multi-zone air conditioning control strategies. A high-quality interactive dataset is constructed, which comprehensively records the current step count, environmental information, room air conditioning control action groups, reward values, and indoor temperature after executing control actions under different environmental conditions. After the PPO model training is complete, this dataset is saved to provide reference actions for larger models.
[0057] The specific process for calculating the reward value for executing the air conditioning control action based on the multiphysics simulation environment is as follows: The specific power value is obtained by multiplying the room air conditioning control action corresponding to the current environmental information with the air conditioning power; the temperature difference is obtained by subtracting the indoor temperature setpoint from the indoor temperature value after executing the room air conditioning control action; the energy consumption penalty is calculated by multiplying the specific power value by the room coefficient, and the comfort penalty is calculated by the temperature difference. The energy consumption penalty value and the comfort penalty value are added together to obtain the total penalty value.
[0058] The indoor temperature value after executing the room air conditioning control action is calculated by the physics engine in the multiphysics simulation environment. The physics engine is responsible for calculating the indoor temperature in the multiphysics simulation environment. It should be noted that in the process of calculating the reward value for executing the air conditioning control action, only the power value is obtained, and this power value is only used to calculate the reward value energy consumption.
[0059] Assuming there are n rooms, i.e. n zones, and the air conditioner's rated power Pac = 1400 W, what is the control action for the i-th room? Where negative values represent cooling and positive values represent heating; based on this condition, the actual power consumption of the air conditioner is, i.e., the specific power value: ; Energy consumption penalty is: ; in, It is the energy consumption penalty weight, i.e., the room coefficient. Where 0.05 is a fixed coefficient and n is the number of rooms.
[0060] The comfort penalty is a piecewise function: ;
[0061] The total reward value obtained based on the energy consumption penalty value and the comfort penalty value is: .
[0062] The changes in room temperature and building energy consumption following the execution of room air conditioning control actions are quantified into specific reward values to reflect the effectiveness of the current air conditioning control actions. An adaptive weight initialization strategy based on the number of rooms is adopted, where the energy consumption penalty weight is set to 0.05 divided by the number of rooms. This method considers the aggregation effect of multi-room control actions and temperature errors, ensuring that the reward scale remains within a reasonable range as the number of rooms increases, and providing a unified weight initialization benchmark for buildings of different sizes.
[0063] S102: Obtain current environmental information and determine the first control strategy for the multi-zone air conditioning of the target building using the reinforcement learning model trained by iterative simulation, including: inputting the obtained current environmental information into the reinforcement learning model trained by iterative simulation to obtain the control strategy recommended by the reinforcement learning model under the current environmental information, and using the control action recommended by the reinforcement learning model as the first control strategy; The reinforcement learning model is a model that is pre-trained through interactive training with the iterative simulation model and contains recommended control strategies under different environmental information.
[0064] In this embodiment, the reinforcement learning model adopts the PPO model, specifically: after receiving actual environmental information, the PPO model outputs the air conditioning control actions for the room based on the environmental information. The air conditioning control actions output by the PPO model for multiple areas of the target building based on the environmental information are used as the first control strategy for the air conditioning in multiple areas of the target building.
[0065] For example, if Room 1, Room 2, and Room 3 are multiple areas, the trained PPO model will output a 1*3 array ([0.2], [0.3], [0.2]), indicating that the air conditioning control action of Room 1 is 0.2, the air conditioning control action of Room 2 is 0.3, and the air conditioning control action of Room 3 is 0.2.
[0066] This invention uses a reinforcement learning model (PPO model) as a high-quality demonstration generator, enabling it to interact with a multiphysics simulation environment to produce a large amount of high-reward state-action pairing data covering diverse scenarios. Based on this, cue word pairs are automatically constructed for fine-tuning large models. This method automatically distills energy-saving and comfort-oriented control logic from the data, replacing inefficient and subjective expert experience coding and ensuring the objectivity and comprehensiveness of knowledge injection.
[0067] S103: Construct prompt words based on the first control strategy and current environmental information; input the prompt words into the fine-tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building; specifically: Obtain environmental information of the target building, the current indoor temperature, and pre-defined role definitions, task requirements, and output formats; The environmental information is input into the reinforcement learning model to obtain the first control policy output by the reinforcement learning model. Based on environmental information, current indoor temperature, role definition, task requirements, output format, and first control strategy, prompt words are constructed and input into the finely tuned large model. The finely tuned large model outputs a new multi-zone air conditioning control action group, and the new multi-zone air conditioning control action group is determined as the second control strategy for the multi-zone air conditioning of the target building.
[0068] The fine-tuned large model is a large language model equipped with a LoRA adapter and a value function. It is fine-tuned and trained using a pre-collected interactive dataset of multi-zone air conditioning in the target building. Each data point in the interactive dataset contains [current step count, environmental information, room air conditioning control action group, reward value, and indoor temperature after the action is executed]; To ensure that the large language model can generate high-precision and high-stability control actions, this invention designs a structured prompt word template. Since the output of the large language model is random, it may give different magnitude actions for the same state, leading to system oscillations or increased energy consumption. The structured prompt words, by clearly defining the value range, unit, and calculation basis of the control actions, effectively constrain the output space, significantly improving repeatability and stability. Furthermore, the original large model lacks expert knowledge guidance in the field of air conditioning control. The structured prompt words use the recommended actions of the PPO model as a reference benchmark and provide environmental information, providing prior knowledge guidance for the model, thereby improving control accuracy and energy-saving effects without the need for retraining.
[0069] The constructed prompts combine the environmental information, the current indoor temperature, the role definition, the task requirements, the output format, and the first control strategy, and input them into the finely tuned large model.
[0070] It's important to note that the role definition, task requirements, and output format in the prompts are pre-set. The role definition specifies the AI's identity, domain, and responsibilities. For example: You are a building energy management AI, responsible for controlling the air conditioning system to maintain the room temperature close to the target temperature. Task requirements: 1. Analyze the current environmental state; 2. Output the air conditioning action to maintain the environment at 22℃. A positive value indicates heating, and a negative value indicates cooling. The current indoor temperature is obtained in real-time through sensors.
[0071] The output format forces the model to output a brief reasoning process (approximately 20 words) and a normalized array of actions, in the format "Actions: [a1, a2, a3]", ensuring output parsability. For example: Output format: 1. Your analysis and reasoning process: approximately 20 words. 2. Must be output in the following action format: Actions: [0.2, 0.2, 0.2], please replace 0.2 with your actions.
[0072] Here is a complete prompt: Role definition: You are a building energy management AI, responsible for controlling the air conditioning system to maintain the room temperature close to the target temperature.
[0073] Task Requirements: 1. Analyze the current environmental state. 2. Output the actions of the air conditioner to maintain the environment at 22℃. Positive values indicate heating, and negative values indicate cooling. Output Format: 1. Your analysis and reasoning process: approximately 20 words. 2. Must be output according to the following action format. Actions: [0.2, 0.2, 0.2], please replace 0.2 with your actions. Please make a decision based on the following information: Current environmental information: [Personnel heat output], [Ground temperature], [Outdoor temperature], [Solar radiation intensity]; Current indoor temperature: [18.2, 19.3, 17.6]; PPO recommended actions [Action 1, Action 2, Action 3].
[0074] Through the above-described structured design, this invention enables the large language model to accurately and consistently map environmental states into highly reliable control commands, providing a reliable input basis for subsequent simulation verification and iterative correction, thereby achieving more precise control while satisfying comfort and energy efficiency.
[0075] Because the format of the text information output by large models is unstable, a multi-strategy parsing method is used to extract control actions from the natural language text returned by the large models. Specifically: (1) Try the following parsing strategies in sequence: (1.1) Matches compound action formats without parentheses (e.g., "0.2, 0.2, 0.3"); (1.2) Parse the last line as a JSON array; (1.3) Scan the entire text for bracket structures to find numerical pairs.
[0076] (2) Limit the control actions to a reasonable physical range [-1, 1].
[0077] The fine-tuning process for the large model is as follows: Prompt words are constructed using the interactive dataset of multi-zone air conditioning in the target building, prompting the model to output control actions. Then, the action loss is calculated based on the PPO loss function. Finally, backpropagation is used to update the LoRA adapter, enabling the large model to learn control actions that yield high reward values. During backpropagation, gradients simultaneously apply to the low-rank matrix parameters of the LoRA adapter and the linear layer parameters of the newly added value function, updating both. After training, the LoRA adapter and value function parameters are saved, forming a large-scale intelligent building control model that can be deployed independently. Figure 6-7 As shown, the specific process is as follows: S1031: Read the saved interactive dataset of multi-zone air conditioning in the target building. Each entry includes: [current step number, environmental information, room air conditioning control action group, reward value, indoor temperature after the action is executed].
[0078] S1032: Obtain the pre-set role definition, task requirements, and output format. Concatenate the role definition, task requirements, output format, environmental information, current indoor temperature, and room air conditioning control action group into a prompt word, and input it into the large model equipped with a LoRA adapter and value function. The large model outputs a new room air conditioning control action group based on the prompt word.
[0079] S1033: Calculate the loss between the large model room air conditioning control action group and the PPO model room air conditioning control action group using the default loss function of the near-end policy optimization algorithm, and use the backpropagation algorithm to synchronously update the LoRA adapter parameters and the parameters of the value function.
[0080] S1034: Determine whether all data information has been used. If not, repeat S1031-S1033; if yes, training ends, and the updated parameters of the LoRA adapter and value function are saved.
[0081] This training process distills the decision-making capabilities of the PPO model into a larger model, enabling the model to retain its general language understanding capabilities while gaining the ability to generate low-energy, high-comfort room air conditioning control actions based on building environment information.
[0082] The finely tuned large model is equipped with a LoRA adapter and value function, transforming the general large model into a professional large model with expertise in building control while maintaining the original language understanding ability. This ensures efficient deployment while guaranteeing that the model can generate stable and reliable air conditioning control actions.
[0083] S104: Verify the second control strategy using the calibrated simulation model, and control the operation of the air conditioning equipment in the target building's multi-zone system according to the verified second control strategy; for the second control strategy that fails verification, the verification result is fed back to the fine-tuned large model in the form of feedback prompts for correction.
[0084] The second control strategy is verified using a calibrated iterative simulation model, and the operation of the multi-zone air conditioning equipment in the target building is controlled according to the verified second control strategy; including: The second control strategy is input into the calibrated iterative simulation model. Based on the current state of the target building and the second control strategy, the iterative simulation model calculates the operating parameters of the multi-zone air conditioning of the target building after the second control strategy is executed. The operating parameters of the multi-zone air conditioning in the target building after executing the second control strategy are compared with the preset target operating parameters to obtain the deviation value; it is then determined whether the deviation value is within the range of the first preset threshold to the second preset threshold. If yes, the second control strategy is verified, and the operation of the air conditioning equipment in the target building's multi-zone system is controlled according to the verified second control strategy; if no, the second control strategy is not verified.
[0085] If the second control strategy fails the verification, the failure result will be fed back to the fine-tuned large model in the form of feedback prompts, and the fine-tuned large model will make corrections based on the feedback prompts.
[0086] The feedback prompt is the reason why the second control strategy failed verification. The feedback prompt consists of role definition, task requirements, output format, environmental information, current indoor temperature, room air conditioning control action group, and feedback suggestion. Feedback prompt: [role definition, task requirements, output format, environmental information, current indoor temperature, room air conditioning control action group, feedback suggestion].
[0087] Feedback suggestions are based on pre-validation or historical control effects, providing qualitative recommendations. For example, "If the temperature in some rooms is lower than the target, it is recommended to increase heating appropriately; if the temperature in some rooms is higher than the target, it is recommended to reduce heating appropriately; if the temperature is close to the target, fine-tuning can be done," guiding the large language model to adjust the magnitude of actions after fine-tuning.
[0088] The feedback prompt word structure effectively constrains the output space of a large language model by simultaneously considering the synergistic effects of role positioning, task instructions, output format constraints, and feedback suggestions, thereby improving the reliability and efficiency of control decisions.
[0089] Furthermore, to ensure the safety of the room air conditioning control action group output by the fine-tuned large model, each control action generated by the fine-tuned large model will be rapidly simulated in the simulation model before being sent to the real device. If it fails to meet the standard, an interception will be triggered immediately, and the specific reason for the failure will be injected as a structured feedback prompt to guide the large model to generate corrective actions.
[0090] After the large model is fine-tuned and the second control strategy is determined, the second control strategy is input into the simulation model (multiphysics simulation environment) for verification. The verified second control strategy is then output to control the actual equipment operation of the target building. For second control strategies that fail verification, the failure results are fed back to the fine-tuned large model in the form of feedback prompts, and the large model is corrected based on these prompts. The specific process is as follows: S1041: After the fine-tuned large model generates the air conditioning control action, the system first inputs the action into the multiphysics simulation environment for single-step pre-simulation. The multiphysics simulation environment calculates the predicted temperature value of each room after execution based on the current building status and action parameters.
[0091] S1042: The system sets a temperature deviation threshold (±2℃) as the judgment standard, and calculates the deviation between the predicted temperature value and the target temperature value of each room. Among them, the first preset threshold is -2℃ and the second preset threshold is +2℃. If the temperature is within the range of the first preset threshold to the second preset threshold, i.e. (-2℃, +2℃), the action is determined to be safe and sent to the real energy system for execution. If the temperature exceeds the threshold range, a safety warning is triggered, and the system generates a feedback prompt, such as "The action just now caused the temperature of room 1 to be lower than the target temperature. It is recommended to adjust the high control action." The system guides the large model to generate a correction action based on the feedback prompt.
[0092] S1043: Verify whether the air conditioning control action generated by the fine-tuned large model meets the standard. If it does, execute it; otherwise, repeat steps S1041-S1042 until it meets the standard or the maximum number of repetitions (3 times) is reached.
[0093] S105: Obtain the actual operating data of the real building after the current control is executed, add the actual operating data to the historical operating data, and return to the parameter calibration step for repeated execution.
[0094] To achieve performance improvement and climate adaptation throughout the system's lifecycle, this invention establishes a weekly iterative optimization process. An update process is initiated weekly: new data is combined with the updated simulation environment to drive fine-tuning of the large model, thus updating the large model. This closed loop enables the entire system to achieve continuous iterative updates. Details are as follows: Based on the newly collected data, the system performs a model update process; The PPO model and LoRA adapter are retrained to fine-tune the deployed large model, adapting it to environmental information after seasonal changes.
[0095] This invention adds a runtime safety decision-making closed loop: each set of control commands generated by the large model must be injected into a multiphysics simulation environment for single-step rapid pre-running and physical safety verification before being sent to the real device. Commands are only released if the results meet safety constraints; otherwise, structured feedback is generated to guide the model to re-decision. Simultaneously, the system establishes a long-term system evolution closed loop: new data from the real system is collected periodically (e.g., weekly), driving both the recalibration of the simulation model's parameters and the incremental fine-tuning of the large control model. This dual-model co-evolution closed loop enables the entire system to have sustainable iterative update capabilities, adapting to changes in climate change and building usage patterns, fundamentally alleviating the problem of continuous optimization after deployment using traditional methods.
[0096] By deeply integrating multiphysics simulation models, reinforcement learning optimization capabilities, and large language model reasoning and generalization capabilities, and establishing a multi-level closed-loop mechanism that runs through the entire system lifecycle and interacts with the real environment, we have achieved full-process intelligence, safety, and adaptability from perception and decision-making to safe execution and continuous evolution, providing a better systemic solution for the field of building energy conservation.
[0097] In one specific implementation, the control of a heating, ventilation, and air conditioning system in three rooms is taken as an example.
[0098] First, the system automatically collects current environmental information by deploying temperature sensors, solar radiation sensors, and pyroelectric sensors for personnel within the room, and downloading the data through a data platform. The collected environmental information includes: ground temperature 26℃, outdoor temperature 34℃, solar radiation intensity 600 W / m², and personnel heating power 225W.
[0099] The aforementioned environmental information is input into the trained and saved PPO model, which outputs a recommended control action value [-0.3, -0.2, -0.2]. This recommended control action value is defined as: -0.3 × the current operating power of the air conditioner, used only to calculate the energy consumption penalty. Subsequently, the system combines the environmental information, the current indoor temperature, role definition, task requirements, output format, and the recommended control action into a prompt word vector, namely: [Ground temperature: 26℃, Outdoor temperature: 34℃, Room indoor temperature: 18℃, 17℃, 19℃, Solar radiation: 600 W / m², Human heat generation power: 225W, Recommended control action: [-0.3, -0.2, -0.2], Role definition: You are a building energy management AI, responsible for controlling the air conditioning system to maintain the room temperature close to the target temperature. Task requirements: 1. Analyze the current environmental state. 2. Output the air conditioning action to maintain the environment at 22℃. Positive values for the action indicate heating, and negative values indicate cooling. Output format: 1. Your analysis and reasoning process: approximately 20 words. 2. You must output according to the following action format. Actions: [-0.3,-0.2,-0.2], please replace with your actions.
[0100] The prompt word vector is input into the fine-tuned large language model, which generates what it considers the most suitable control action based on the received environmental information and the room temperature. For example, the large language model considers the outdoor temperature to be high, so it generates a control action value of -0.5.
[0101] To ensure the safety and effectiveness of the control actions, the system inputs the control action value of -0.5 generated by the large language model into the pre-built simulation model to perform simulation derivation. Based on the current environmental conditions, the simulation model calculates the predicted indoor temperature after executing the control action, which is 20℃. The system further compares this predicted indoor temperature with the preset target temperature value of 25℃, calculating a deviation of -5℃ (i.e., the predicted temperature is 5℃ lower than the target temperature). If this deviation exceeds a preset threshold (e.g., -2℃), it is determined that the control action may result in an excessively low room temperature, failing to meet comfort requirements.
[0102] In response to the above judgment, the system automatically generates feedback suggestions: "The previous action caused the room temperature to fall below the target temperature. It is recommended to increase the control action." The feedback suggestions are combined with the role definition, task requirements, output format, environmental information, current indoor temperature, and room air conditioning control action group to form feedback prompt words. The feedback prompt words are input into the large language model to guide the large language model to correct its previously generated control actions based on physical constraints. For example, the cooling action may be reduced from -0.5 to -0.3, thereby iteratively generating better control instructions.
[0103] Example 2 This embodiment provides a multi-zone air conditioning control system based on iterative simulation and fine-tuning of a large model, including: The first control strategy generation module acquires current environmental information and uses a reinforcement learning model trained by iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building. The second control strategy generation module constructs prompt words based on the first control strategy and current environmental information; inputs the prompt words into the fine-tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building; The verification module verifies the second control strategy using the calibrated simulation model, and controls the operation of the air conditioning equipment in the target building's multi-zone system based on the verified second control strategy; for the second control strategy that fails verification, the verification result is fed back to the fine-tuned large model in the form of feedback prompts for correction. The closed-loop iterative module collects actual operating data of air conditioning in multiple areas of the real building after control execution, adds it to historical operating data, and returns the iterative simulation model parameter calibration steps for repeated execution.
[0104] It should be noted that the first control strategy generation module, the second control strategy generation module, the verification module, and the closed-loop iteration module mentioned above correspond to the multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model in Embodiment 1. The examples and application scenarios implemented by the above modules and their corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.
[0105] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0106] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.
[0107] Example 3 This embodiment also provides a multi-zone air conditioning control device based on iterative simulation and fine-tuning of a large model, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to make the electronic device perform the method described in Embodiment 1 above.
[0108] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.
[0109] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.
[0110] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.
[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A multi-zone air conditioning control method based on iterative simulation and fine-tuning of large models, characterized by, include: Obtain current environmental information and use the reinforcement learning model trained by iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building. Based on the first control strategy and current environmental information, a prompt word is constructed; The prompt words are input into the finely tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building. The second control strategy is verified using the calibrated simulation model. The operation of the air conditioning equipment in the target building's multi-zone system is controlled according to the verified second control strategy. For the second control strategy that fails verification, the verification result is fed back to the fine-tuned large model in the form of feedback prompts for correction. Collect actual operating data of air conditioning in multiple areas of a real building after control execution, add it to historical operating data, and return to the iterative simulation model parameter calibration step for repeated execution.
2. The multi-zone air conditioning control method based on large model iteration simulation and fine-tuning according to claim 1, wherein, It also includes acquiring historical operational data of real buildings, using the historical operational data to calibrate the parameters of the iterative simulation model, and obtaining the calibrated iterative simulation model; The reinforcement learning model was trained using the calibrated iterative simulation model; The iterative simulation model is calibrated using the historical running data, specifically including: Acquire historical operational data of a real building, including environmental information, current indoor temperature, executed control actions, and measured indoor temperature after execution; The environmental information, current indoor temperature, and control actions from the historical operating data are input into the iterative simulation model, and the predicted indoor temperature for the next moment is calculated through the state-space equations of the iterative simulation model. With the goal of minimizing the mean square error between the indoor temperature predicted by the simulation at the next moment and the measured indoor temperature, the gradient descent method is used to update the thermal parameter matrix in the iterative simulation model, so that the thermodynamic characteristics of the iterative simulation model can approximate the actual thermodynamic characteristics of the real building. 3.The multi-zone air conditioning control method based on large model iteration simulation and fine tuning according to claim 1, wherein, Obtain current environmental information and use the reinforcement learning model trained through iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building, including: The acquired current environment information is input into the reinforcement learning model trained by iterative simulation to obtain the control strategy recommended by the reinforcement learning model under the current environment information, and the control action recommended by the reinforcement learning model is used as the first control strategy. The reinforcement learning model is a model that is pre-trained through interactive training with the iterative simulation model and contains recommended control strategies under different environmental information.
4. The multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model according to claim 1, characterized in that, Based on the first control strategy and current environmental information, a prompt word is constructed; the prompt word is input into the fine-tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building; Specifically: The system acquires environmental information of the target building, the current indoor temperature, and pre-defined role definitions, task requirements, and output formats; it then inputs the environmental information into a reinforcement learning model to obtain the first control strategy output by the reinforcement learning model. Based on environmental information, current indoor temperature, role definition, task requirements, output format, and first control strategy, prompt words are constructed and input into the finely tuned large model. The finely tuned large model outputs a new multi-zone air conditioning control action group, and the new multi-zone air conditioning control action group is determined as the second control strategy for the multi-zone air conditioning of the target building.
5. The multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model according to claim 1, characterized in that, The second control strategy is verified using a calibrated iterative simulation model, and the operation of the multi-zone air conditioning equipment in the target building is controlled according to the verified second control strategy; including: The second control strategy is input into the calibrated iterative simulation model. Based on the current state of the target building and the second control strategy, the iterative simulation model calculates the operating parameters of the multi-zone air conditioning of the target building after the second control strategy is executed. The operating parameters of the multi-zone air conditioning in the target building after executing the second control strategy are compared with the preset target operating parameters to obtain the deviation value; it is then determined whether the deviation value is within the range of the first preset threshold to the second preset threshold. If yes, the second control strategy is verified, and the operation of the multi-zone air conditioning equipment in the target building is controlled according to the verified second control strategy; if no, the second control strategy is not verified.
6. The multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model according to claim 1, characterized in that, The feedback prompt is the reason why the second control strategy failed the verification; the feedback prompt is composed of role definition, task requirements, output format, environmental information, current indoor temperature, room air conditioning control action group, and feedback suggestion.
7. The multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model according to claim 6, characterized in that, The environmental information includes: human heat output, ground temperature, outdoor temperature, and solar radiation intensity.
8. A multi-zone air conditioning control system based on iterative simulation and fine-tuning of a large model, characterized in that, include: The first control strategy generation module acquires current environmental information and uses a reinforcement learning model trained by iterative simulation to determine the first control strategy for the multi-zone air conditioning of the target building. The second control strategy generation module constructs prompt words based on the first control strategy and the current environment information; The prompt words are input into the finely tuned large model to determine the second control strategy for the multi-zone air conditioning of the target building. The verification module verifies the second control strategy using the calibrated simulation model, and controls the operation of the air conditioning equipment in the target building's multi-zone system based on the verified second control strategy; for the second control strategy that fails verification, the verification result is fed back to the fine-tuned large model in the form of feedback prompts for correction. The closed-loop iterative module collects actual operating data of air conditioning in multiple areas of the real building after control execution, adds it to historical operating data, and returns the iterative simulation model parameter calibration steps for repeated execution.
9. A multi-zone air conditioning control device based on iterative simulation and fine-tuning of a large model, characterized in that, include: A memory for storing instructions, the instructions comprising the steps of the multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model as described in any one of claims 1 to 7; A processor for executing the instructions.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-zone air conditioning control method based on iterative simulation and fine-tuning of a large model as described in any one of claims 1 to 7.