Large-scale building refrigeration system intelligent energy saving method based on large language model

Abstract

The invention relates to an intelligent energy-saving method for a large building refrigerating system based on a large language model, and belongs to the field of intelligence of large building refrigerating systems. A reinforcement learning framework is constructed, a multi-dimensional state space formed by building thermal loads, equipment operation states and thermal environment parameters is defined, and a dynamic reward mechanism is designed by taking improvement of energy efficiency, guarantee of thermal safety constraints and maintenance of thermal comfort as optimization objectives; developing a word embedding model special for the refrigeration field and a cross-modal decision conversion model, and realizing bidirectional analysis of a natural language instruction and a physical control parameter through feature coding and embedding mapping; and performing low-rank adaptive fine tuning on the pre-trained large language model based on the reinforcement learning experience set, and constructing a closed-loop verification system. Generalized migration and multi-target collaborative decision-making of equipment control strategies are achieved, the self-adaptive regulation and control capacity of a refrigeration system under dynamic loads is improved, and a universal intelligent solution for energy efficiency optimization, thermal safety and comfort is provided for a large building.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent large-scale building refrigeration systems, and relates to a method for collaborative optimization of multi-level equipment and precise energy consumption control of large-scale building refrigeration systems based on a large language model. Background Art

[0002] With the acceleration of urbanization and the rapid growth of large buildings, building energy consumption has become a core component of my country's energy consumption. Against this backdrop, the issue of building energy consumption has become increasingly prominent, becoming a core component of my country's energy consumption. Relevant surveys show that building energy consumption accounts for a significant proportion of total energy consumption nationwide, nearly half. Among these buildings, over 95% of existing buildings are high-energy-consuming, with significant energy waste. Cooling systems contribute significantly to building energy consumption, exceeding 40% of overall building energy consumption. Therefore, optimizing the energy efficiency of cooling systems has become crucial for building energy conservation. Optimizing the operating efficiency of cooling systems has multiple important implications. First, it can effectively reduce overall building energy consumption and reduce excessive energy consumption, thereby alleviating energy shortages to a certain extent. Second, by introducing a range of measures, such as intelligent control technologies, energy-saving equipment, and advanced energy management systems, energy utilization can be significantly optimized. This not only helps improve energy efficiency but also provides strong technical support for achieving green and low-carbon buildings, driving the construction industry towards a more environmentally friendly and sustainable development.

[0003] The cooling systems of large buildings face numerous complexities. Due to the numerous internal heat sources, such as the heat generated by the continuous operation of numerous electrical devices, coupled with large fluctuations in occupancy density and significant variations in the inflow and outflow of people and the concentration of people at different times of day, the cooling load demand exhibits significant fluctuations and high dynamics over time. Furthermore, the cooling systems of these large buildings typically rely on multiple devices, such as chillers, cooling towers, and water pumps, to meet the cooling load demand. However, the interdependencies and coupling relationships between these devices are complex, making the overall system operation difficult to predict and adjust.

[0004] While the application of variable frequency drive technology has alleviated some operational issues in refrigeration systems to a certain extent, adjusting the operating frequency of equipment based on actual load conditions, thereby achieving energy savings within a certain range, the multi-loop coupled chiller, cooling tower, pump, and other equipment constitutes a nonlinear, strongly coupled system. The complex interactions and constraints between these devices make it difficult to achieve optimal overall system performance using variable frequency drive technology alone.

[0005] Several existing technologies address the precise energy consumption control of large-scale building cooling systems. For example, patent CN114692493A discloses an energy-saving method for building cooling systems. This method utilizes a recurrent neural network to accurately predict cooling load at the end point, which serves as input for subsequent energy-saving optimization. This method then models the hardware parameters of the equipment, using end-point load data as a basis for real-time optimization. However, this method still suffers from issues such as inability to adapt to changing environments and limited generalization capabilities. The method disclosed in patent CN117436586A uses the Deep Deterministic Policy Gradient (DDPG) algorithm for joint control, but its action decision-making process is mismatched with the actual regulation characteristics of the system. Specifically, the power generation adjustment variable of the power supply subsystem is a continuous variable, while the outlet water temperature adjustment variable of the refrigeration subsystem is essentially a discrete adjustment gear (such as a fixed temperature set point). However, there is a natural contradiction between the continuous action space output by the DDPG algorithm and the discrete action space of the refrigeration subsystem. As a result, the continuous action output by the decision maker cannot be directly mapped to the discrete regulation unit of the refrigeration subsystem. This may cause the control signal to be disconnected from the regulation logic of the actual actuator, resulting in insufficient adaptability of the action decision in the discrete action space and loss of control accuracy.

[0006] Therefore, from the current state of the art, existing technologies have the following shortcomings: Traditional control methods, such as proportional-integral-derivative (PID) control, have significant limitations in refrigeration system applications. Refrigeration systems exhibit a response lag, meaning that when demand on the cooling side changes, the system cannot react immediately. Traditional PID control methods, due to their inherent characteristics, cannot promptly adjust the matching between the cooling supply and cooling demand sides. In actual operation, when cooling demand decreases, cooling supply may not decrease in time, resulting in energy waste. Conversely, when cooling demand increases, cooling supply cannot quickly catch up, affecting cooling performance. Deep reinforcement learning (DRL) methods also have their shortcomings in refrigeration system control. These methods typically rely on fixed environments for training, with specific environmental parameters and conditions set during the training process. However, in real-world cooling scenarios for large buildings, the environment is constantly changing, with factors such as outdoor temperature and personnel flow constantly shifting. Because the DRL method is trained based on a fixed environment, it has difficulty adapting quickly to this real-time changing environment, resulting in weak generalization of its strategy. This means that in different real-world scenarios, it may not function effectively and cannot make optimal control decisions based on actual conditions. Although the Large Language Model (LLM) has strong generalization and reasoning capabilities and can handle a variety of general text information and questions, it has obvious shortcomings in the field of refrigeration. The large language model lacks professional knowledge in the refrigeration field, and its general architecture is difficult to adapt to professional problems such as thermodynamic parameter analysis and multi-device collaborative control in refrigeration systems. Refrigeration systems involve complex thermodynamic principles and the collaborative operation of multiple devices, requiring precise analysis and control of various parameters. However, due to the lack of professional knowledge, the large language model has difficulty accurately modeling the system state and cannot formulate effective optimization and control strategies.

[0007] Therefore, there is an urgent need to develop more efficient and intelligent energy-saving methods. These methods need to be able to adapt to the complex and fluctuating cooling load demands of large buildings, adjusting cooling capacity in real time based on factors such as time of day and environmental conditions. This approach should achieve coordinated and optimized control of multiple groups of equipment, overcoming the traditional control methods' reliance on fixed environments and preset models. This approach will improve the cooling system's responsiveness in dynamic environments, reduce energy waste caused by load fluctuations and device coupling, and ultimately improve overall energy efficiency. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide an intelligent energy-saving method for large-scale building refrigeration systems based on a large language model, so as to provide key technical support for solving the problem of collaborative optimization of multi-level refrigeration equipment in complex dynamic environments, break through the dependence of traditional control strategies on fixed environment training, and realize adaptive energy consumption precision control of refrigeration systems in large-scale building scenarios such as commercial complexes and data centers, providing professional and general intelligent solutions for the green and low-carbon operation of large buildings.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] An intelligent energy-saving method for a large building refrigeration system based on a large language model comprises the following steps:

[0011] S1. Build a reinforcement learning framework for large-scale building cooling systems, including: defining reinforcement learning states, determining action decision parameters, establishing optimization goals, and setting reward mechanisms;

[0012] S2. Design an input-adaptive word embedding model and a cross-modal decision transformation model. This includes: designing a specialized word embedding model for the large language model in the refrigeration field based on the input characteristics of the large language model; and establishing a bidirectional mapping relationship between the natural language semantic space and the physical parameter space of the refrigeration system.

[0013] S3. Fine-tuning and training the large language model, including: based on the large-scale building refrigeration system reinforcement learning framework described in S1, using the refrigeration strategy to interact with the simulation software to generate a reinforcement learning experience set as a fine-tuning training set for the large model; using this data set to fine-tune the large language model to learn professional knowledge in the field of large-scale building refrigeration systems, and at the same time training the decision conversion model.

[0014] Furthermore, the reinforcement learning state in S1 includes the load condition, thermal state and equipment operating state of the large building; the operating state includes: chiller outlet water temperature, air handling unit outlet air temperature, cooling tower cooling water outlet water temperature, building temperature, air handling unit outlet air volume, chilled water pump and cooling water pump flow rate;

[0015] The action decision parameters in S1 include: chilled water outlet temperature of the chiller, air flow rate of the air handling unit, water outlet temperature of the cooling tower and water outlet flow rate of the cooling chilled water pump;

[0016] The optimization objectives described in S1 include: improving energy efficiency, ensuring thermal safety constraints, maintaining thermal comfort, and coordinated operation among multiple devices;

[0017] The reward mechanism described in S1 is: setting a reinforcement learning reward mechanism based on the optimization objective to guide the large-scale building refrigeration system to make the best decision between energy conservation, stable operation of equipment and user comfort requirements.

[0018] Furthermore, the word embedding model in S2 is composed of a feature encoding layer and an embedding mapping layer, wherein the feature encoding layer performs preliminary feature extraction on the original sequence state, and the embedding mapping layer fuses the extracted thermodynamic parameters and equipment operation state features;

[0019] The establishment of a bidirectional mapping relationship between the natural language semantic space and the physical parameter space of the refrigeration system described in S2 is specifically as follows: developing a cross-modal decision conversion model to parse the natural language decisions generated by the large language model into a structured instruction set containing decision parameters such as the chilled water outlet temperature of the chiller and the air flow rate of the air handling unit.

[0020] Furthermore, the definition of the reinforcement learning state in S1 is specifically:

[0021] The state vector is set as a time series set containing thermal state, equipment operating state and load characteristics, specifically expressed as:

[0022] s=(s thl ,s equip ,P buliding )

[0023] Among them, s represents the state of reinforcement learning; s thl Thermal status of large buildings; equip P is the equipment operation status of large buildings; buliding The overall power consumption of large buildings;

[0024] Each state takes N historical values, such as P for load conditions. buliding express:

[0025] P buliding =(P buliding (t),P buliding (t-1),…,P buliding (tN))

[0026] Among them, t represents the current time;

[0027] The thermal state s thl Designed to include a time series set of large building temperatures and multiple sets of equipment temperatures, thermal state s thl Specifically expressed as:

[0028]

[0029] Among them, T building is the time series temperature in a large building; The mth group of chillers are the chilled water outlet temperature, air handling unit outlet temperature, and cooling tower cooling water outlet temperature, and the temperatures are all sequential temperatures with N historical values, m = 1, 2, ..., M, where M is the total number of equipment groups;

[0030] Set the device operating status equip The design is a time series set of multiple sets of equipment operating parameters, specifically expressed as:

[0031] s equip =(Q CRAH,1 ,Q CHW,1 ,Q CW,1 ,…,Q CRAH,M ,Q CHW,M ,Q CW,M )

[0032] Among them, Q CRAH,m , Q CHW,m , Q CW,m They are the air output volume of the air handling unit of group m, the flow rate of the chilled water pump, and the flow rate of the cooling water pump.

[0033] Furthermore, the action decision in S1 is specifically:

[0034] The action decision is set as a mixed action space including discrete and continuous action spaces; for the mth group of devices, m = 1, 2, ..., M, where M is the total number of device groups; the discrete action vector a d,m Expressed as:

[0035]

[0036] a d,m Including chiller chilled water outlet temperature adjustment Cooling tower cooling water outlet temperature adjustment in, Continuous motion vector a c,m Expressed as:

[0037] a c,m =(ΔQ CRAH ,ΔQ CHW,m ,ΔQ CW,m )

[0038] a c,m Contains the air volume adjustment Q of the mth group of air handling units CRAH,m , Chilled water pump flow adjustment Q CHW,m , Cooling water pump flow adjustment Q CW,m ; The climbing constraint is expressed as:

[0039]

[0040]

[0041] Furthermore, the establishment of optimization goals and setting of reward mechanisms described in S1 are specifically as follows:

[0042] The optimization goal is set to minimize the comprehensive objective function, and the energy efficiency is expressed as:

[0043]

[0044] P cooling (t) = P EC (t)+P pump (t)+P CT (t)+P CRAH (t)

[0045] Among them, P cooling (t) represents the energy consumption of the refrigeration system in period t, P building (t) represents the energy consumption of large buildings in period t; P EC (t), P pump (t), P CT (t), P CRAH (t) represents the energy consumption of the chiller, water pump, cooling tower, and air handling unit in period t respectively;

[0046] The objective function is expressed as:

[0047]

[0048] Where, Indicates the constraints on thermal safety and comfort, where T building represents the room temperature of a large building, ψ T represents the safety / comfort threshold of the temperature in a large building, α1 and α2 are weight factors, and their sum is 1;

[0049] The reward mechanism for reinforcement learning is set as:

[0050] r=r1-α1PUE-α2ln(1+exp(T room -ψ T ))-τ1

[0051] Where r1 is a large constant to ensure that the reward is positive; α1 and α2 are weight factors; ln(1+exp(T room -ψ T )) represents the penalty for overheating of large buildings, ψ T represents the safety / comfort threshold of the temperature in a large building, and τ1 is the penalty for violating the constraint conditions in the action decision.

[0052] Furthermore, the establishment of optimization goals and setting of reward mechanisms in S1 also include setting constraints for the refrigeration system, specifically:

[0053] The constraints of the refrigeration system are expressed as:

[0054]

[0055] Where, Indicates the rated air output of the air handling unit;

[0056] The constraints of the chilled water outlet temperature of the chiller and the cooling water outlet temperature of the cooling tower are expressed as:

[0057]

[0058] In the formula and Respectively represent the maximum and minimum values of the chilled water outlet temperature, and Respectively represent the maximum and minimum values of the cooling water outlet temperature;

[0059] The limiting constraint condition of the pump rated flow is expressed as:

[0060]

[0061] Where, Indicates the rated flow of the chilled water pump. Indicates the rated flow rate of the cooling water pump.

[0062] Furthermore, the S2 specifically includes:

[0063] S21. Design the feature encoding layer: Use a one-dimensional convolutional neural network to extract features from the original time series state. The state output dimension after feature extraction by the feature encoding layer is [batch size, sequence length, feature dimension].

[0064] S22. Design the embedding mapping layer: Set a trainable linear layer in the embedding mapping layer;

[0065] S23. Construct a cross-modal decision conversion model to issue cooling decisions: set up two trainable linear layers to make specific cooling decisions based on the output features of the large language model, one linear layer is used to output control decisions in the continuous action space, and the other linear layer is used to output control decisions in the discrete action space.

[0066] Furthermore, the S3 specifically includes:

[0067] S31. Establish a reinforcement learning environment: Build a simulation system for the cooling system of a large building, import historical load data of the large building, simulate the operation of the cooling system of the large building under different load scenarios, and collect environmental data to establish a reinforcement learning environment;

[0068] S32. Generate a large language model fine-tuning training set: Use deep reinforcement learning strategies to obtain professional experience data through online interaction;

[0069] S33, domain-adapted large language model fine-tuning training: Use low-rank adaptation to perform domain-adapted fine-tuning training on large language models;

[0070] S34. Verification of the domain-adapted large language model: Integrate the large language model, word embedding model and cross-modal decision conversion module into the simulation environment established in S31 to build a closed-loop verification system.

[0071] Furthermore, the environmental data in S31 includes: thermal state data, equipment state data, and energy consumption data;

[0072] The S32 specifically includes: utilizing the reinforcement learning agent to collect environmental states, execute action decisions to complete action delivery, and obtain rewards from environmental feedback according to a reward function;

[0073] The S33 is specifically as follows: freeze the pre-trained parameters |Φ0| of the large language model, introduce additional low-rank matrices to approximate the changes required by the specific network layer of the large language model; for the pre-trained matrix W with a dimension of d×k in Φ0 o , assuming that there is a low rank r that is much smaller than min{d,k}, construct two low rank matrices A and B of dimensions: d×r and r×k to approximate W o During the fine-tuning training process, only the attention layer is low-rank decomposition;

[0074] The S34 is specifically as follows: real-time collection of environmental status, input into the large language model after pre-processing by the embedding layer, and output of natural language instructions which are parsed into chiller chilled water outlet temperature, cooling tower cooling water outlet temperature, air handling unit air volume, and chilled cooling water pump flow adjustment value through the decision conversion model.

[0075] The beneficial effects of the present invention are:

[0076] Enhanced multi-device collaborative optimization and dynamic control capabilities: A reinforcement learning framework is used to define multidimensional state spaces, such as building heat load, equipment operating parameters, thermal environment response, and mixed action spaces. This, for example, combines discrete and continuous actions, effectively resolving the complex device coupling and response lag issues inherent in traditional control methods. Discrete action design, such as the three-state adjustment of chiller temperature, avoids energy consumption fluctuations caused by frequent starts and stops. Continuous actions, such as gradual flow rate regulation of pumps, enable refined control. Combined with time series state modeling, this approach captures load fluctuations and thermodynamic parameter trends, compensates for control delays, and enhances the system's adaptability to dynamic cooling load demands, transcending the reliance of traditional PID or single DRL methods on fixed environments.

[0077] Domain Knowledge Embedding and Accurate Cross-Modal Decision Parsing: Developing a specialized word embedding model and cross-modal decision transformation model for the refrigeration domain, the team extracts thermodynamic parameters and equipment operating characteristics through a feature encoding layer, and embeds a mapping layer to achieve bidirectional mapping between natural language commands and physical control parameters. This design bridges the semantic gap between general-purpose LLMs and refrigeration systems, accurately parsing natural language commands, such as increasing the cooling tower outlet water temperature, into structured control parameters and embedding physical constraints such as flow thresholds and temperature limits. This prevents equipment malfunctions caused by misinterpreted commands and significantly improves control stability and safety.

[0078] Efficient fine-tuning and lightweight deployment adaptability: Low-rank adaptive technology is used to fine-tune the LLM for domain adaptation. Only the attention layer parameters are adjusted and the pre-trained weights are frozen. This approach infuses refrigeration domain expertise while retaining general semantic understanding capabilities. Low-rank matrix approximation significantly reduces the computational cost of fine-tuning, adapting to the real-time decision-making requirements of edge computing devices. Combined with reinforcement learning experience set training, knowledge distillation and dynamic optimization strategy transfer are achieved, supporting rapid adaptation to different building scenarios such as data centers and commercial complexes, enhancing the solution's industry universality.

[0079] Multi-objective optimization and comprehensive system energy efficiency improvement: A dynamic reward mechanism integrates energy consumption indicators, thermal safety thresholds, and comfort constraints to balance energy conservation, stability, and user experience through multi-objective optimization. Discrete action design reduces equipment start-up and shutdown losses, extending the life of key components. Continuous action smooth adjustment reduces pipeline pressure fluctuations and improves system stability. Cross-modal analysis and lightweight deployment reduce heterogeneous system integration costs and shorten technology implementation cycles.

[0080] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0082] Figure 1 This is a diagram of a multi-device refrigeration architecture of a refrigeration system according to an embodiment of the present invention;

[0083] Figure 2 This is an LLM fine-tuning training graph according to an embodiment of the present invention;

[0084] Figure 3 FIG. 1 is a diagram of a method architecture according to an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0086] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0087] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0088] See also Figure 1, which is a multi-device refrigeration architecture diagram of a refrigeration system according to an embodiment of the present invention, illustrates the workflow of multiple cold source devices in a data center refrigeration system. The diagram includes four cooling towers, which are connected to four chillers (chillers 1-4) via cooling water pumps (cooling water pumps 1-4). The chillers are then connected to four CRAHs (computer room air conditioners, CRAH1-4) via chilled water pumps (chillers 1-4). Chilled water and cooling water are distinguished by supply and return water using lines of different colors, and the hot air in the data center is processed and converted into cold air for circulation. The entire system achieves cooling of the data center and ensures its stable operation through the coordinated operation of cooling towers, chillers, water pumps, terminal air conditioners and other equipment.

[0089] This paper proposes an intelligent energy-saving method for large-scale building refrigeration systems based on a large language model (LLM). This method can provide key technical support for solving the problem of collaborative optimization of multi-level refrigeration equipment in complex dynamic environments. It effectively breaks through the reliance of traditional control strategies on fixed environment training, realizes adaptive energy consumption precision control of refrigeration systems in large-scale building scenarios such as commercial complexes and data centers, and provides a professional, universal and intelligent solution for the green and low-carbon operation of large buildings. The method includes the following steps:

[0090] S1: Construct a reinforcement learning framework for large-scale building cooling systems. This mainly includes defining states, action decisions, optimization objectives, and reward mechanisms. States include large-scale building load conditions, thermal conditions, and equipment operating conditions. Specifically, they include the chiller (Electric Chiller, EC) outlet water temperature, the computer room air handler (CRAH) outlet air temperature, the cooling tower cooling water outlet temperature, the building temperature, the CRAH air volume, and the flow rates of the chilled water pump and cooling water pump. This fully reflects the system's operating status and cooling demand. Action decisions are set as the chiller chilled water outlet temperature, CRAH air volume, cooling tower outlet water temperature, and cooling water pump outlet flow. The EC chilled water outlet temperature affects the cooling capacity output, the CRAH air volume directly affects the thermal environment in the building and is related to comfort and energy consumption, the cooling tower outlet water temperature has a significant impact on the efficiency of the entire refrigeration system, and the outlet flow of the cooling and chilled water pumps affects system stability and equipment operating efficiency. The optimization objective comprehensively considers energy utilization efficiency, thermal safety and thermal comfort constraints, while taking into account the coordinated operation of multiple devices. Based on the optimization objective, a reinforcement learning reward mechanism is established to encourage the system to make decisions that are conducive to energy saving, stable operation, and comfort assurance.

[0091] S2: Design an input-adaptive word embedding model and a cross-modal decision conversion model. Based on the input characteristics of the LLM, a word embedding model specifically for the LLM in the refrigeration field is designed. The word embedding model consists of a feature encoding layer and an embedding mapping layer. The feature encoding layer is used to perform preliminary feature extraction on the original sequence state, while the embedding mapping layer integrates the extracted thermodynamic parameters, equipment operating status and other professional features to achieve accurate mapping of unstructured environmental data to a high-dimensional semantic space. At the same time, a cross-modal decision conversion model is developed to parse the natural language decisions generated by the LLM into a structured instruction set containing control parameters such as the chilled water outlet temperature of the chiller and the CRAH air flow rate. A bidirectional mapping relationship is established between the natural language semantic space and the physical parameter space of the refrigeration system, eliminating the semantic gap between the general language model and the domain control logic.

[0092] S3: Fine-tune the training of a large language model. Figure 2 , which is an LLM fine-tuning training graph for an embodiment of the present invention. Based on the reinforcement learning framework in step S1, an existing high-efficiency cooling strategy is used to interact with the simulation software. The resulting reinforcement learning experience set serves as the fine-tuning training set for the large model. This dataset is then used to fine-tune the large language model, enabling it to learn specialized knowledge in the field of large-scale building cooling systems. Simultaneously, this dataset is used to train the decision-making transformation model, improving its ability to accurately transform semantic information into cooling system decision outputs.

[0093] S11: Define the reinforcement learning state. Considering the thermal inertia characteristics of the cooling system, the dynamic changes in building loads, and the coordinated operation characteristics of multiple devices, the state vector is designed as a time series set containing thermal state, equipment operating state, and load characteristics. Specifically, it is expressed as:

[0094] s=(s thl ,s equip ,P buliding )

[0095] Among them, s represents the state of reinforcement learning; s thl Thermal status of large buildings; equip P is the equipment operation status of large buildings; buliding The overall power consumption of large buildings;

[0096] Each state takes N historical values, such as P for load conditions. buliding express:

[0097] P buliding =(P buliding (t),P buliding (t-1),…,P buliding (tN))

[0098] Among them, t represents the current time;

[0099] The thermal state s thl Designed to include a time series set of large building temperatures and multiple sets of equipment temperatures, thermal state s thl Specifically expressed as:

[0100]

[0101] Among them, T building is the time series temperature in a large building; They are the chilled water outlet temperature of the mth group EC, the CRAH air outlet temperature, and the cooling tower cooling water outlet temperature, and the temperatures are all sequential temperatures with N historical values, m = 1, 2, ..., M, where M is the total number of equipment groups.

[0102] Set the device operating status equip The design is a time series set of multiple sets of equipment operating parameters, specifically expressed as:

[0103] s equip =(Q CRAH,1 ,Q CHW,1 ,Q CW,1 ,…,Q CRAH,M ,Q CHW,M ,Q CW,M )

[0104] Among them, Q CRAH,m , Q CHW,m , Q CW,m They are the air output volume, chilled water pump flow rate, and cooling water pump flow rate of the mth group of CRAH respectively.

[0105] Introducing historical time step values of multi-dimensional state parameters can improve the model's predictive ability to cope with complex working conditions such as load fluctuations and cooling delays, effectively capture the changing trends of thermodynamic parameters and load periodic characteristics, compensate for the time difference between control decisions and system responses, provide a complete environmental context for reinforcement learning agents, and support the formulation of precise control strategies.

[0106] S12: Determine the action decision. Based on the requirements of collaborative control of multiple groups of devices, the action decision is designed as a hybrid action space, covering both discrete and continuous action spaces. For the mth group of devices (m = 1, 2, ..., M, where M is the total number of device groups), the discrete action vector can be expressed as:

[0107]

[0108] a d,m Specifically including EC chilled water outlet temperature adjustment Cooling tower cooling water outlet temperature adjustment in, The design of discrete actions relies on the thermodynamic characteristics of the equipment. Discrete gear adjustment is used to avoid frequent start and stop of the equipment and maintain a reasonable coefficient of performance (COP). The continuous action vector is expressed as:

[0109] a c,m =(ΔQ CRAH ,ΔQ CHW,m ,ΔQ CW,m )

[0110] a c,m Specifically includes the mth group CRAH air volume adjustment amount Q CRAH,m , Chilled water pump flow adjustment Q CHW,m , Cooling water pump flow adjustment Q CW,m , in order to ensure the stability of the system, a climbing constraint is introduced, which is expressed as:

[0111]

[0112] Through fine adjustment of continuous variables, the needs of building thermal environment control and energy consumption optimization can be met, and the impact of drastic parameter fluctuations on system operation can be avoided.

[0113] S13: Establish optimization goals and set up reward mechanisms. The optimization goal is set to minimize the comprehensive objective function to achieve a balance between maximizing energy efficiency (Power Usage Effectiveness, PUE) and thermal safety\comfort constraints. PUE can be used to evaluate the energy efficiency of large buildings. It refers to the ratio of the total energy consumption consumed during the operation of a large building to the overall energy consumption in the building (excluding cooling energy consumption). The closer the PUE value is to 1, the higher the energy efficiency of the large building. PUE is described by the following formula:

[0114]

[0115] P cooling (t) = P EC (t)+P pump (t)+P CT (t)+ CRAH (t)

[0116] Among them, P cooling (t) represents the energy consumption of the refrigeration system in period t, P building (t) represents the energy consumption of large buildings in period t; P EC (t), P pump (t), P CT (t), P CRAH (t) represents the energy consumption of the chiller, water pump (including chilled water pump and cooling water pump), cooling tower, and air handling unit in period t.

[0117] The objective function is specifically expressed as:

[0118]

[0119] In the formula represents the penalty for overheating of large buildings, i.e., the constraint on thermal safety and comfort, where T building represents the room temperature of a large building, ψ T represents the safety / comfort threshold of temperature in large buildings, α1 and α2 are weight factors, and their sum is 1.

[0120] According to the optimization goal, the reward mechanism of reinforcement learning is set, and the immediate reward obtained from the environment is set to:

[0121] r=r1-α1PUE-α2ln(1+exp(T room -ψ T ))-τ1

[0122] Where r1 is a large constant to ensure that the reward is positive; α1 and α2 are weight factors; ln(1+exp(τ room -ψ T )) represents the penalty for overheating of large buildings, ψ T represents the safety / comfort threshold of the temperature in a large building, and τ1 is the penalty for violating the constraint conditions in the action decision.

[0123] The constraints of the cooling system, first of all, consider that the CRAH is limited by its fan rated air volume and motor rated capacity, and there are the following constraints:

[0124]

[0125] Where, Indicates the rated air output of the CRAH.

[0126] Based on the operating characteristics of the chiller and cooling tower equipment, the chilled water outlet temperature of the EC and the cooling water outlet temperature of the cooling tower are subject to the following constraints:

[0127]

[0128] In the formula and Respectively represent the maximum and minimum values of the chilled water outlet temperature, and Respectively represent the maximum and minimum values of the cooling water outlet temperature.

[0129] In addition, due to the limitation of the rated flow of the pump, the following constraints exist:

[0130]

[0131] In the formula Indicates the rated flow of the chilled water pump. Indicates the rated flow rate of the cooling water pump.

[0132] Furthermore, the above step S2 specifically includes:

[0133] S21: Design the feature encoding layer. Since the one-dimensional convolutional neural network (1D Convolutional Neural Network, 1D-CNN) is specifically used to process time series and sequence data, it is particularly suitable for the reinforcement learning state composed of historical state sequences proposed in S1. Therefore, the feature encoding layer uses 1D-CNN to extract features from the original time series state. Through the convolution operation of 1D-CNN, the temporal features and patterns in the time series can be effectively extracted. The state output dimension after feature extraction by the feature encoding layer is [Batch_size, seq_len, feature_dim], that is, [batch size, sequence length, feature dimension], where they are batch length, sequence length, and feature output dimension respectively.

[0134] S22: Design the embedding mapping layer. However, the features extracted by the feature encoding layer may not be aligned with the LLM's token space. For example, the state output dimension after feature extraction is feature_dim = 512, while large language models such as GPT-2 require a 768-dimensional input. To address this problem, trainable linear layers are designed for the embedding mapping layer. These linear layers can automatically learn an efficient mapping method to convert the feature space into the token space. In addition, the projected embeddings are further normalized through layer normalization to ensure the stability of training, and ultimately a set of token-like embeddings are generated, which can be effectively used by large language models.

[0135] S23: Construct a cross-modal decision transformation model. Through the word embedding model, LLM can extract high-level features containing important task-specific information from the historical state sequence. These features are then directly output to the cooling decision through the cross-modal decision transformation model. Specifically, the cross-modal decision transformation model is designed as two trainable linear layers, which are used to make specific cooling decisions based on the LLM output features. One linear layer is used to output the control decision in the continuous action space, and the other linear layer is used to output the control decision in the discrete action space. Unlike the native LLM output model head, the cross-modal decision transformation model can map the natural language semantic space to the physical parameter space of the cooling system to complete the distribution of cooling decisions.

[0136] Furthermore, the above step S3 specifically includes:

[0137] S31: Establish a reinforcement learning environment. Using simulation software such as the Transient System Simulation Tool (TRNSYS) or the EnergyPlus tool, first establish a simulation system for the cooling system of a large building. Import historical load data for the large building and simulate the operation of the cooling system under different load scenarios. Collect environmental data including thermal status (including water temperature and room temperature), equipment status (including air volume and flow), and energy consumption data (power consumption of each device) as the basis for establishing a reinforcement learning environment. Integrate the defined reinforcement learning elements to establish a reinforcement learning environment.

[0138] S32: Generate LLM fine-tuning training set. Based on the reinforcement learning environment established in S31, existing deep reinforcement learning strategies such as Deep Deterministic Policy Gradient (DDPG) and Deep Q-Network (DQN) are adopted to obtain professional experience data through online interaction. Specifically, the reinforcement learning agent collects environmental states, executes action decisions, completes action dispatch, and obtains rewards from environmental feedback according to the reward function. In this way, the agent continuously interacts with the environment to obtain experience data such as states, actions, and rewards, and continuously accumulates a sufficient number of [states s, continuous actions a c , discrete action a d , reward r] quadruple experience pool, which forms a standardized training set after cleaning and denoising, providing a domain-specific fine-tuning training set for subsequent LLM.

[0139] S33: Domain-adapted LLM fine-tuning training. Based on the domain-specific fine-tuning training set generated by S32, the low-rank adaptation (LoRA) efficient fine-tuning technology is used to perform domain-adapted fine-tuning training on the LLM. Due to the large number of LLM parameters, such as the parameter volume |Φ0| of the Pathways Language Model (PaLM) reaching 540 billion, directly training all parameters will incur an unbearable computational cost. For this reason, the LoRA technology is used to first freeze the pre-trained parameters |Φ0| of the LLM, and then introduce an additional low-rank matrix to approximate the changes required for the specific network layer of the LLM, and then learn domain-specific knowledge. The core is that the parameter changes |Φ0| in the fine-tuning process (indicating the parameter update amount required when fine-tuning the original pre-trained parameters Φ0 of the LLM) are low-rank.

[0140] For the pre-trained matrix W of dimension d×k in Φ0 o, assuming that there is a low rank r that is much smaller than min{d,k}, construct two low rank matrices A and B of dimensions: d×r and r×k to approximate W o The update is:

[0141] W=W o +ΔW=W o +AB

[0142] When fine-tuning, W o is frozen, and all parameter updates only act on matrices A and B. During the LoRA fine-tuning process, only the attention layer of LLM is trained, that is, only the attention layer is low-rank decomposed, because this layer plays a key role in processing the input sequence and capturing the dependencies and context information in the sequence. This fine-tuning training method can not only significantly reduce the fine-tuning cost, but also because each W in Φ0 o Neither is updated, and the pre-trained knowledge of LLM can be retained, which means that the same LLM can be used as the basic model for different scenarios, and domain-adapted expertise can be acquired by training different low-rank matrices.

[0143] S34: Verification of the Domain-Adapted LLM. The trained LLM, word embedding model, and cross-modal decision transformation module are integrated into the simulation environment established in S31 to construct a closed-loop verification system. The specific process is as follows: Environmental conditions are collected in real time, pre-processed by the embedding layers (S21 and S22), and then input into the LLM. The natural language instructions output by the model are parsed by the decision transformation model (S23) into EC chilled water outlet temperature, cooling tower cooling water outlet temperature, CRAH air volume, and chilled cooling water pump flow adjustment values. During the verification phase, multiple different test scenarios are set up, including normal operating conditions, extreme operating conditions (such as high temperature and high load, low temperature and low load), and abnormal operating conditions such as simulated equipment failures. For each test scenario, the system is run continuously for a period of time to collect various environmental feedback data, such as thermal conditions (water temperature, room temperature), equipment status, and energy consumption data (power consumption of each device, PUE value). This feedback data is compared and analyzed with the expected ideal data to evaluate the accuracy and effectiveness of the model's decisions.

[0144] The present invention integrates multi-source heterogeneous data such as historical temperature series of chillers, building temperature field distribution, and real-time power of terminal equipment through multi-dimensional state perception and hybrid action space design, constructs a dynamic perception network with temporal and spatial correlation, and combines the attention mechanism of the large language model (LLM) to achieve forward-looking capture of the periodic laws of cooling load fluctuations and sudden abnormal signals, providing advanced decision-making support for equipment group control. A three-state adjustment strategy (heating / maintaining / cooling) is designed for discrete equipment such as chillers, and smooth gradual control is adopted for continuous variables such as water pump flow, taking into account the physical characteristics of the equipment and the response speed of the system, significantly optimizing the overall energy efficiency ratio (PUE) of the refrigeration system, reducing the comprehensive energy consumption cost of large buildings, and at the same time reducing equipment losses caused by frequent start-up and shutdown, extending the service life of key components, and coordinating the operating status of the equipment group in advance in the scenario of sudden change in cooling load to avoid thermal environment fluctuations caused by the disconnection between cooling supply and demand.

[0145] In terms of domain knowledge embedding and cross-modal decision transformation, by developing a multimodal attention mechanism, the natural language instructions output by the large language model (such as "increase the cooling tower outlet water temperature to reduce energy consumption") are dynamically parsed into structured control parameters and embedded into the equipment physical constraint knowledge base, automatically avoiding instructions that exceed the physical limits of the equipment (such as excessive chilled water pump flow), achieving a deep integration of semantic understanding and physical constraints, significantly reducing the risk of equipment malfunction caused by misinterpretation of instructions, and demonstrating excellent control stability under complex working conditions. By intelligently parsing energy-saving oriented instructions, energy waste caused by redundant cooling is reduced, and combined with the seamless docking design of mainstream industrial protocols, the customized development investment of heterogeneous equipment systems is reduced, the technology implementation cycle is shortened, and control accuracy and operation and maintenance efficiency are improved.

[0146] For low-rank, efficient fine-tuning technology, a parameter freezing strategy is used to lock the core parameters of the large language model, leaving only the weights related to the attention mechanism open for adaptive adjustment. This not only retains the general capabilities of natural language understanding, but also injects expert strategy experience into the model through knowledge distillation technology, strengthening the understanding of the dynamic characteristics of the cooling system and achieving continuous improvement in energy efficiency during long-term operation. The lightweight fine-tuning solution is adapted to edge computing devices, supporting real-time decision-making in low-computing environments and reducing hardware deployment costs. At the same time, through knowledge transfer capabilities, it quickly adapts to the cooling needs of different building types (such as data centers and commercial complexes), improving the industry universality of the solution and significantly optimizing the economic efficiency and environmental adaptability of the cooling system throughout its life cycle.

[0147] The comprehensive technical advantages of this invention include: significantly improving the overall PUE of a building through dynamic coordinated control, achieving industry-leading energy efficiency levels; reducing cooling system operation and maintenance costs, including equipment maintenance, energy consumption, and manual intervention costs; and maintaining a stable thermal environment in extreme climates and sudden load fluctuations, ensuring building functional safety and user comfort. Typical application scenarios include dynamically optimizing traffic density in commercial complexes, maximizing the use of free cooling in data centers, and adapting to cyclical heat loads in industrial plants, demonstrating its broad applicability and practical engineering value.

[0148] See also Figure 3 , is a method architecture diagram of an embodiment of the present invention; the figure shows an architecture that integrates a refrigeration system and an LLM intelligent control model. The refrigeration system on the upper left side includes a cooling tower, a cooling water pump, a chiller, a chilled water pump and a CRAH, which is connected to a large building to realize heat and cold exchange. In the LLM intelligent control model part below, the state is first acquired and processed by the word embedding model, and then the LLM fine-tuning technology based on Lora is combined with the trained LLM, and then the action is issued (including continuous and discrete action issuance). Chilled water and cooling water circulate in the system, and hot air and cold air are also involved in the exchange. This architecture optimizes the operation of the refrigeration system through the LLM intelligent control model to create a suitable temperature environment for large buildings.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent energy-saving method for large-scale building refrigeration systems based on a large language model, characterized in that: The following steps are involved: S1. Build a reinforcement learning framework for large-scale building cooling systems, including: defining reinforcement learning states, determining action decision parameters, establishing optimization goals, and setting reward mechanisms; S2. Design an input-adaptive word embedding model and a cross-modal decision transformation model. This includes: designing a specialized word embedding model for the large language model in the refrigeration field based on the input characteristics of the large language model; and establishing a bidirectional mapping relationship between the natural language semantic space and the physical parameter space of the refrigeration system. S3. Fine-tuning and training the large language model, including: based on the large-scale building refrigeration system reinforcement learning framework described in S1, using the refrigeration strategy to interact with the simulation software to generate a reinforcement learning experience set as a fine-tuning training set for the large model; using this data set to fine-tune the large language model to learn professional knowledge in the field of large-scale building refrigeration systems, and at the same time training the decision conversion model.

2. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 1 is characterized in that: The reinforcement learning state in S1 includes the load condition, thermal state and equipment operating state of the large building; the operating state includes: chiller outlet water temperature, air handling unit outlet air temperature, cooling tower cooling water outlet water temperature, building temperature, air handling unit outlet air volume, chilled water pump and cooling water pump flow rate; The action decision parameters in S1 include: chilled water outlet temperature of the chiller, air flow rate of the air handling unit, water outlet temperature of the cooling tower and water outlet flow rate of the cooling chilled water pump; The optimization objectives described in S1 include: improving energy efficiency, ensuring thermal safety constraints, maintaining thermal comfort, and coordinated operation among multiple devices; The reward mechanism described in S1 is: setting a reinforcement learning reward mechanism based on the optimization objective to guide the large-scale building refrigeration system to make the best decision between energy conservation, stable operation of equipment and user comfort requirements.

3. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 2 is characterized in that: The word embedding model in S2 is composed of a feature encoding layer and an embedding mapping layer. The feature encoding layer performs preliminary feature extraction on the original sequence state, and the embedding mapping layer integrates the extracted thermodynamic parameters and equipment operation status features; The establishment of a bidirectional mapping relationship between the natural language semantic space and the physical parameter space of the refrigeration system described in S2 is specifically as follows: developing a cross-modal decision conversion model to parse the natural language decisions generated by the large language model into a structured instruction set containing decision parameters such as the chilled water outlet temperature of the chiller and the air flow rate of the air handling unit.

4. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 3 is characterized by: The definition of the reinforcement learning state in S1 is specifically: The state vector is set as a time series set containing thermal state, equipment operating state and load characteristics, specifically expressed as: s=(s thl ,s equip ,P buliding ) Among them, s represents the state of reinforcement learning; s thl Thermal status of large buildings; equip P is the equipment operation status of large buildings; buliding The overall power consumption of large buildings; Each state takes N historical values, such as P for load conditions. buliding express: P buliding =(P buliding (t),P buliding (t-1),…,P buliding (t-N)) Among them, t represents the current time; The thermal state s thl Designed to include a time series set of large building temperatures and multiple sets of equipment temperatures, thermal state s thl Specifically expressed as: Among them, T building is the time series temperature in a large building; The mth group of chillers are the chilled water outlet temperature, air handling unit outlet temperature, and cooling tower cooling water outlet temperature, and the temperatures are all sequential temperatures with N historical values, m = 1, 2, ..., M, where M is the total number of equipment groups; Set the device operating status equip The design is a time series set of multiple sets of equipment operating parameters, specifically expressed as: s equip =(Q CRAH,1 ,Q CHW,1 ,Q CW,1 ,…,Q CRAH,M ,Q CHW,M ,Q CW,M ) Among them, Q CRAH,m , Q CHW,m , Q CW,m They are the air output volume of the mth group of air handling units, the chilled water pump flow rate, and the cooling water pump flow rate.

5. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 4 is characterized in that: The specific action decision in S1 is: The action decision is set as a mixed action space including discrete and continuous action spaces; for the mth group of devices, m = 1, 2, ..., M, where M is the total number of device groups; the discrete action vector a d,m Expressed as: a d,m Including chiller chilled water outlet temperature adjustment Cooling tower cooling water outlet temperature adjustment in, Continuous motion vector a c,m Expressed as: a c,m =(ΔQ CRAH ,ΔQ CHW,m ,ΔQ CW,m ) a c,m Contains the air volume adjustment Q of the mth group of air handling units CRAH,m , Chilled water pump flow adjustment Q CHW,m , Cooling water pump flow adjustment Q CW,m ; The climbing constraint is expressed as:

6. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 5 is characterized in that: The establishment of optimization goals and setting of reward mechanisms in S1 are specifically as follows: The optimization goal is set to minimize the comprehensive objective function, and the energy efficiency is expressed as: P cooling (t)=P EC (t)+P pump (t)+P CT (t)+P CRAH (t) Among them, P cooling (t) represents the energy consumption of the refrigeration system in period t, P building (t) represents the energy consumption of large buildings in period t; P EC (t), P pump (t), P CT (t), P CRAH (t) represents the energy consumption of the chiller, water pump, cooling tower, and air handling unit in period t respectively; The objective function is expressed as: Where, Indicates the constraints on thermal safety and comfort, where T building represents the room temperature of a large building, ψ T represents the safety / comfort threshold of the temperature in a large building, α1 and α2 are weight factors, and their sum is 1; The reward mechanism for reinforcement learning is set as: r=r1-α1PUE-α2ln(1+exp(T room -ψ T ))-t1 Where r1 is a large constant to ensure that the reward is positive; α1 and α2 are weight factors; ln(1+exp(T room -ψ T )) represents the penalty for overheating of large buildings, ψ T represents the safety / comfort threshold of the temperature in a large building, and τ1 is the penalty for violating the constraint conditions in the action decision.

7. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 6 is characterized in that: Establishing the optimization goal and setting the reward mechanism in S1 also includes setting the constraints of the refrigeration system, specifically: The constraints of the refrigeration system are expressed as: Where, Indicates the rated air output of the air handling unit; The constraints of the chilled water outlet temperature of the chiller and the cooling water outlet temperature of the cooling tower are expressed as: In the formula and Respectively represent the maximum and minimum values of the chilled water outlet temperature, and Respectively represent the maximum and minimum values of the cooling water outlet temperature; The limiting constraint condition of the pump rated flow is expressed as: Where, Indicates the rated flow of the chilled water pump. Indicates the rated flow rate of the cooling water pump.

8. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 3 is characterized by: The S2 specifically includes: S21. Design the feature encoding layer: Use a one-dimensional convolutional neural network to extract features from the original time series state. The state output dimension after feature extraction by the feature encoding layer is [batch size, sequence length, feature dimension]. S22. Design the embedding mapping layer: Set a trainable linear layer in the embedding mapping layer; S23. Construct a cross-modal decision conversion model to issue cooling decisions: set up two trainable linear layers to make specific cooling decisions based on the output features of the large language model, one linear layer is used to output control decisions in the continuous action space, and the other linear layer is used to output control decisions in the discrete action space.

9. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 3 is characterized by: The S3 specifically includes: S31. Establish a reinforcement learning environment: Build a simulation system for the cooling system of a large building, import historical load data of the large building, simulate the operation of the cooling system of the large building under different load scenarios, and collect environmental data to establish a reinforcement learning environment; S32. Generate a large language model fine-tuning training set: Use deep reinforcement learning strategies to obtain professional experience data through online interaction; S33, domain-adapted large language model fine-tuning training: Use low-rank adaptation to perform domain-adapted fine-tuning training on large language models; S34. Verification of the domain-adapted large language model: Integrate the large language model, word embedding model and cross-modal decision conversion module into the simulation environment established in S31 to build a closed-loop verification system.

10. The intelligent energy-saving method for large-scale building refrigeration systems based on a large language model according to claim 9, characterized in that: The environmental data in S31 includes: thermal status data, equipment status data, and energy consumption data; The S32 specifically includes: utilizing the reinforcement learning agent to collect environmental states, execute action decisions to complete action delivery, and obtain rewards from environmental feedback according to a reward function; The S33 is specifically as follows: freeze the pre-trained parameters |Φ0| of the large language model, introduce additional low-rank matrices to approximate the changes required by the specific network layer of the large language model; for the pre-trained matrix W with a dimension of d×k in Φ0 o , assuming that there is a low rank r that is much smaller than min{d,k}, construct two low rank matrices A and B of dimensions: d×r and r×k to approximate W o During the fine-tuning training process, only the attention layer is low-rank decomposition; The S34 is specifically as follows: real-time collection of environmental status, input into the large language model after pre-processing by the embedding layer, and output of natural language instructions which are parsed into chiller chilled water outlet temperature, cooling tower cooling water outlet temperature, air handling unit air volume, and chilled cooling water pump flow adjustment value through the decision conversion model.

Images