Energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms

Abstract

The present invention provides an energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms, including preprocessing historical operating data of the air-cooled air-conditioning units; obtaining current parameters of the computer room to be adjusted, performing filtering operations on the historical operating data based on the current parameters, and generating a candidate data set; vectorizing the current outdoor temperature and humidity and IT equipment energy consumption in the candidate data set to construct a vector database supporting RAG retrieval; constructing historical associated data based on the current set temperature of each air conditioner in the computer room to be adjusted; and judging whether the amount of historical associated data reaches a set threshold: Step S6: converting the control strategy into a control signal to control the air-cooled air conditioner in real time, and generating complete log information including the control process, controlling the energy-saving optimization of the air-cooled air-conditioning system, and predicting the optimal air-conditioning return air temperature based on historical data and real-time environmental changes, thereby maximizing energy efficiency while ensuring the comfort of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic control of air-conditioning systems, and in particular to an energy-saving method for air-cooling air-conditioning units based on the integration of RAG and MPC algorithms. Background Art

[0002] As data center energy consumption continues to rise, extensive research has focused on reducing cooling system energy consumption. This research primarily focuses on three areas: optimizing airflow organization, developing efficient cooling terminal equipment, and designing energy-efficient operation strategies. Operational strategy research primarily focuses on regulating IT equipment power distribution and cooling system operating parameters. However, existing methods often suffer from significant lag, low control accuracy, and poor stability when responding to environmental fluctuations or drastic load changes, making them difficult to meet operational requirements in complex dynamic scenarios.

[0003] Therefore, to address this problem, an energy-saving method for air-cooled air-conditioning units based on the fusion of RAG and MPC algorithms was developed. An air-conditioning operation knowledge base was constructed from historical operation data. Combined with a similarity retrieval mechanism, historical samples closest to the current operating status were extracted, and then the optimal air-conditioning setting parameter recommendations were generated. This allows historical experience to drive current decisions, thereby improving the intelligent control level and energy-saving adaptability of the air-conditioning system. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the present invention proposes an energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms. Each time the system executes the control strategy, complete log information is automatically generated and stored, covering historical related data and corresponding similarities, the control strategy generation method adopted, the return air temperature setting value of the current air-conditioning equipment, and the generated return air temperature setting value content. The air-conditioning return air temperature setting value generated by the output module of the RAG system is based on specific data in the knowledge base. While providing the temperature setting value, the data used as the basis will be output, thereby achieving full-process traceability of the control process and interpretability of key decision nodes.

[0005] To achieve the above objectives, the present invention proposes an energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms, comprising the following steps:

[0006] Step S1: Preprocess the historical operating data of the air-cooled air-conditioning unit. The data is obtained every 30 minutes. The temperature and humidity sensors are used to obtain temperature and humidity data in real time. The obtained data includes:

[0007] Computer room information: including computer room number room id ;

[0008] Time information: including the start time of the collection timeand end time end time ;

[0009] Air conditioning information: including the start and stop status of multiple air conditioners state , Return air set temperature Temp seti , and the total air conditioning energy consumption value e2;

[0010] IT equipment energy consumption information: records the IT system load power consumption e1 within the current time window;

[0011] Outdoor information: including outdoor temperature Temp out , Outdoor humidity Hum out ;

[0012] Sensor information: including the temperature values ​​corresponding to multiple sensors Temp sensori , user-set sensor alarm temperature (updated in real time); The energy consumption values ​​of air conditioners and IT equipment are non-real-time cumulative data. The system uses the difference between the cumulative values ​​of two adjacent time points based on a 30-minute sampling interval to obtain the energy consumption for that period;

[0013] Step S2: Get the current parameters of the room to be adjusted "room number room id 、Air conditioning start and stop status air state 、Start time start time and end time end time ", based on which the historical operation data is filtered to select records with completely consistent parameters to generate candidate data sets;

[0014] Step S3: The current outdoor temperature and humidity of the candidate data set out 、Hum out The energy consumption of IT equipment is vectorized, and a vector database that supports RAG retrieval is constructed. Through similarity retrieval, data with a high similarity to the current status parameters of the computer room to be adjusted is obtained;

[0015] Step S4: According to the current set temperature Temp of each air conditioner in the room to be adjusted seti , search for the preset adjustment range of each air conditioner's set temperature within the current value [Temp seti-2 , Temp seti+2 ] (i=1,2,…,N) data to construct historical related data;

[0016] Step S5: Determine whether the number of historical associated data reaches a set threshold value num. If the number reaches the threshold value num, the knowledge base is combined with the output of the return air temperature setting value of the air conditioner that can meet the "minimum energy consumption and sensor temperature is lower than the alarm temperature, and the control range of the return air temperature setting value of the air conditioner". The output of the return air temperature setting value of the air conditioner adopts a prescribed format, and the data of the knowledge base based on it is output. If the return air temperature setting value of the air conditioner cannot be generated based on the relevant data, the output is combined with the requirements such as "adopting MPC control" as input and provided to the RAG generation module to output the corresponding control strategy. If the number is not reached or the RAG generation module outputs "adopting MPC control", the system calls the MPC-based backup control strategy;

[0017] Step S6: Convert the control strategy into a control signal to perform real-time control on the air-cooled air conditioner. To achieve the traceability and explainability of the control strategy, the system generates complete log information each time the strategy is executed.

[0018] Furthermore, the reason for selecting 30 minutes as the data acquisition interval is that:

[0019] After adjusting the cooling capacity of an air-cooled air-conditioning system, response variables such as indoor temperature, equipment load, and energy consumption usually have a certain physical response delay. Using a 30-minute sampling period can effectively reflect the impact of the control action on the system operating status, while avoiding data fluctuations and redundancy caused by excessively high sampling frequency. It can also prevent the missing of key operating condition change information due to excessively long sampling periods, thereby achieving a balance between system response and data granularity, and improving the accuracy and robustness of subsequent cooling capacity decisions.

[0020] Furthermore, the preprocessing in S1 includes:

[0021] Missing value processing: perform deletion operations on sample group records with missing data;

[0022] Outlier processing: Use Method for identification, when the data point x satisfies |x-μ|> When , it is determined to be an outlier and is removed, where μ is the mean value of the data set and σ is the standard deviation of the data set.

[0023] Duplicate value processing: The system allows duplicate values ​​for status or control variables. However, when the same value is repeated continuously and exceeds the specified threshold, the system should automatically trigger an abnormal alarm and generate an alarm log containing the time, frequency and specific value of the duplicate value.

[0024] The start and stop status of the air conditioner is coded using 0 and 1;

[0025] Calculate the cumulative energy consumption of air conditioning and IT equipment.

[0026] Furthermore, the specific implementation steps in S2 are:

[0027] S21: Extract "room" from historical operation data id , air state , start time , end time "Parameters; S22: Get the "room" of the current room to be adjusted id , air state , start time , end time "Parameters, by comparing the parameters in historical data, filter and return the structured data that exactly matches.

[0028] Furthermore, the vector database construction in S3 specifically includes:

[0029] Step S31: The "Temp out , Hum out , the values ​​of e1" are organized into a two-dimensional array to meet the input format requirements of FAISS vector index construction;

[0030] Step S32: The "Temp out , Hum out , e1" constructs the query vector, and performs Top-K similarity search using the L2 distance (Euclidean distance) in the established FAISS index to obtain the index numbers of the K candidate samples closest to the query vector. The L2 distance is calculated as follows:

[0031] ;

[0032] Among them, x is the parameter corresponding to the search value in the vector database, y is the parameter currently searched for in the room to be adjusted, and n is the number of adjustable air conditioners.

[0033] Step S33: Using the index position returned by FAISS, backtrack and extract the corresponding complete data from the original structured data.

[0034] Furthermore, the specific steps in S4 are:

[0035] Step S41: Calculate the adjustment range of each air conditioner according to the current set temperature of each air conditioner in the room to be adjusted [Temp seti-2 , Temp seti+2 ];

[0036] Step S42: Search the data obtained in S3 to find the corresponding air-conditioning temperature in the range [Temp seti-2 ,Temp seti+2] data to obtain historical related data.

[0037] Furthermore, the generation module output control strategy using RAG in S5 includes the following steps:

[0038] Step S511: Textualize the historical related data to form a knowledge base, wherein the historical related data includes "room id , air state , start time , end time , Temp out , Hum out ,e1,e2,Temp seti , Temp sensori ", build templates and convert each piece of data into a natural language sentence;

[0039] Step S512: The knowledge base is combined with the output of the return air temperature setpoint of the air conditioner that satisfies the "minimum energy consumption and sensor temperature below the alarm temperature, and the control range of the return air temperature setpoint of the air conditioner" requirement. The return air temperature setpoint of the air conditioner is output in the specified JSON format, and the data based on the knowledge base is output. If the return air temperature setpoint of the air conditioner cannot be generated based on the relevant data, the requirement of "using MPC control" is output. These are combined and used as the input of the RAG generation module.

[0040] Step S513: The RAG generation module generates and outputs the corresponding control strategy based on the received input. Specifically, during the fusion process, the search results are first sorted in ascending order based on the total air conditioning energy consumption parameter e2, prioritizing the retention of the relevant documents of the first num samples with the lowest energy consumption values. A complete input context is then constructed, guided by multiple constraint prompts and the ordered concatenation of the sorted document contents, to form an enhanced model input that includes energy-saving preference indications.

[0041] Step S514: The final control strategy is generated by calling a pre-trained language generation model (deepseek) and combining it with enhanced context information. The language model does not require retraining or fine-tuning, and reasoning is performed directly through the program interface. To ensure that the generated results meet the high requirements of air conditioning control scenarios for output consistency, parsability, and reliability, the temperature coefficient (temperature) is set to 0.4-0.6 during the reasoning phase to improve the certainty and stability of the output. The cumulative probability truncated sampling (Nucleus Sampling) is set to 0.7-0.8 to limit the range of generated vocabulary and improve output quality. The repetition penalty factor (repetition_penalty) is set to 1 to maintain neutrality to retain necessary field repetitions in the JSON structured output.

[0042] Furthermore, the use of MPC to generate a control strategy in S5 includes the following steps:

[0043] Step S521: Construct the indoor temperature state equation based on the system operation data and the outdoor temperature and humidity:

[0044] ;

[0045] Among them, k1, k2, k3, k4, k5, k 6、 k i+6 is the coefficient, Temp seti is the control quantity, which sets the return air temperature of each air conditioner. in Temp out 、Hum out ,e1, and e2 are state quantities, which are indoor temperature, outdoor temperature, outdoor humidity, IT equipment energy consumption, and total internal consumption of air conditioner, respectively;

[0046] Step S522: Solve the coefficients of the indoor temperature state equation, minimize the residual sum of squares as the optimization goal, use the least squares method to solve and obtain the optimal regression coefficient, and use indicators such as R², MAE, and RMSE to evaluate the performance of the model;

[0047] Step S523: Design a reasonable loss function, which includes three parts:

[0048] The first part is the square of the difference between the predicted value and the target temperature value Temp_target

[0049] The second part is the square of the difference between the current control amount and the last control amount

[0050] The third part is the square of the difference between the actual energy consumption after regulation and the energy-saving target energy consumption

[0051] The loss function expression is: L = α1·(predicted temperature - target temperature)² + α2·(current control amount - last control amount)² + α3·(actual energy consumption - energy saving target)²

[0052] in, 、 、 α is a weight coefficient that is dynamically adjusted according to the real-time operating status of the system. It is used to balance temperature control accuracy, energy efficiency optimization, and regulation stability. In actual applications, these weight coefficients will be dynamically adjusted according to the real-time operating status of the system: when the system load is high, the weight of α3 is increased to prioritize energy efficiency; when the temperature control accuracy requirement is high, the weight of α1 is increased to improve the accuracy of temperature control; the weight of α2 can be adjusted according to the stability of the change of the regulation amount to avoid excessive regulation fluctuations;

[0053] Step S524: Define the adjustment range of the control amount as [Temp seti-2 , Temp seti+2 ], and as a constraint condition, take the loss function designed in step S523 as the optimization objective function, use the SLSQP algorithm to perform constrained multivariable optimization problems, solve the nonlinear minimization problem, set the maximum number of iterations to 50, and the convergence condition is that the change of the objective function in three consecutive iterations is less than 0.1℃, and determine the optimal control strategy for the current system state.

[0054] Furthermore, the control strategy conversion in S6 includes:

[0055] Step S61: After the MPC generates a control strategy, the system receives the control quantity sequence output by the MPC and dynamically adjusts the return air temperature set point of the corresponding terminal air conditioner accordingly;

[0056] Step S62: After the RAG generation module generates the control strategy, it calls the JSON parser to parse the JSON text and convert the temperature setting value to a numeric type. It then extracts the device number from the point identifier and sorts it. Ultimately, it generates an ordered sequence of temperature control instructions and inputs it into the air conditioning control execution module to complete the dynamic adjustment of the terminal equipment. During this process, the system also implements a data verification mechanism. If the generated control strategy content is non-standard or ambiguous, such as missing required fields or data parsing errors, the system returns the backup instruction of "using MPC control" to ensure stable system operation.

[0057] Step S63: Generate log information, including the historical correlation data and corresponding similarity searched in S4, the adopted control strategy generation method, the current return air temperature setting value of the air conditioning equipment, and the generated return air temperature setting value.

[0058] Furthermore, the method monitors and collects data on the operating status of the air-cooled air-conditioning system in real time, and adjusts the set temperature of the air-conditioning unit in real time based on the user's dynamic requirements for operating parameters such as temperature and load, thereby achieving reasonable scheduling and on-demand supply of cooling capacity among multiple air conditioners.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This paper provides an energy-saving method for air-cooled air conditioning units based on the integration of RAG and MPC algorithms. Using Retrieval-Augmented Generation (RAG) technology, this method constructs relevant data and forms a knowledge base. By inputting the retrieved, valid information into a large model, a control strategy is generated. This method effectively mitigates the hallucination problem that can easily occur during the strategy generation process with large models, improving the accuracy and reliability of the control strategy.

[0061] 2. The present invention provides an energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms. When the knowledge base data generated by RAG retrieval is insufficient and the RAG generation module outputs "adopt MPC control", MPC control is adopted. By combining advanced RAG and traditional MPC control, an innovative solution is provided for energy-saving optimization control of air-cooled air-conditioning systems.

[0062] 3. This paper provides an energy-saving method for air-cooled air conditioning units based on the integration of RAG and MPC algorithms. This method uses a retrieval-augmented generation (RAG)-based approach to optimize energy-saving control of air-cooled air conditioning systems. By combining innovative control strategies and technical means, this system can more accurately control the return air setpoint temperature of the air conditioner, achieving a low-energy operation mode. The RAG model generates control strategies by building a knowledge base based on historical operating data, which provides strong interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 It is a structural diagram of the overall process of the present invention;

[0065] Figure 2 It is a process of predictive regulation based on RAG;

[0066] Figure 3 It is the process of MPC algorithm prediction and control. DETAILED DESCRIPTION

[0067] The technical solutions of the present invention will be more clearly and completely explained below through description of preferred embodiments of the present invention in conjunction with the accompanying drawings.

[0068] like Figure 1 As shown, this embodiment provides an energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms, the method comprising the following steps:

[0069] Get historical operating data of air-cooled units: Use temperature and humidity sensors to obtain real-time outdoor temperature Temp out , outdoor humidity Hum out , the temperature values ​​corresponding to multiple sensors Temp sensori ; Use various sensors and instruments to record the room number room id , the start time of collectiontime and end time end time , the start and stop status of multiple air conditioners air state and return air set temperature Temp set , total energy consumption of air conditioning e2, IT system load power consumption e1, user-set sensor alarm temperature (updated in real time).

[0070] Data preprocessing includes:

[0071] (1) Instantaneous quantity: outdoor temperature Temp out , outdoor humidity Hum out , the temperature value Temp corresponding to each sensor sensori , the start and stop status of multiple air conditioners air_state and the return air setting temperature Temp of each air conditioner seti For this type of data, the average value within 30 minutes is used as the integrated data. Accumulated amount: total energy consumption of air conditioners e2, IT system load power consumption e1, this type of data takes the difference between the first and last 30 minutes of data as the integrated data.

[0072] The sensor alarm temperature set by the user uses real-time update data. id , the start time of collection time and end time end time , no further processing is required.

[0073] (2) For missing values, perform sample group deletion; for outliers, use The method is used to identify and eliminate duplicate values. For repeated value processing, an alarm log is generated when the number of repetitions exceeds the threshold, and the start and stop status of the air conditioner is coded with 0 and 1.

[0074] Receive user query data, that is, the parameters of the current room to be adjusted, and extract "room id , air state , start time , end time " parameter, and then traverse the "room id , air state , start time , end time " parameter to filter and return the structured data that exactly matches.

[0075] The "Temp out , Hum out ,e1" values ​​are organized into a two-dimensional array as the FAISSS index vector; the IndexFlatL2 indexer is used to create a FAISSS index based on "Temp out , Hum out, e1" L2 distance index; use the "Temp out , Hum out , the value vector of “e1” is used as the query, and the indexes of the top 15 most similar samples are returned; according to the retrieved index I, the corresponding complete data is extracted from the original structured data.

[0076] Search the corresponding air-conditioning temperature in the range [Temp seti-2 , Temp seti+2 ] data to obtain historical related data.

[0077] Set the threshold num=10 to determine whether the number of historical correlation data reaches the threshold. If not, use MPC to generate the control strategy, such as Figure 2 As shown, the specific steps are:

[0078] (1) Based on the system operation data and outdoor temperature and humidity, the indoor temperature state equation is constructed, and the least squares method is used to optimize the solution with the goal of minimizing the sum of squares of the residuals between the predicted value and the actual value of the indoor temperature, and the optimal coefficient is obtained;

[0079] (2) The energy-saving optimization goal of this embodiment is to minimize the total energy consumption of the air-conditioning system while ensuring that the system operates at a comfortable ambient temperature;

[0080] (3) Construction of loss function: Specifically, the target workshop temperature is set to 25℃, the energy saving target is set to reduce energy consumption by 10%, and the target is set to 200kwh. The loss function is constructed as follows: .

[0081] in,

[0082] ;

[0083] (4) Set the maximum number of iterations to 50, the convergence condition is that the objective function change is less than 0.1°C for three consecutive iterations, and the control adjustment is performed every 30 minutes. The prediction step is set to 20 steps and the control step is 1 step.

[0084] (5) According to the system model, predict the system state for the next 20 steps and use the SLSQP algorithm to solve the optimal control input combination at the next control moment [Temp set1 , Temp set2 ,…,Temp setN ] and input into the system for regulation;

[0085] (6) Repeat step (5) to calculate the control input combination for the next cycle to ensure stable operation of the system, and adjust the control parameters within the control cycle to perform rolling optimization.

[0086] If it is achieved, the generation module of the RAG system is used to output the corresponding control strategy, such as Figure 3 As shown, the specific steps are:

[0087] (1) The historical related data obtained include "room id , air stateN , start time , end time , Temp out , Hum_ out ,e1,e2,Temp seti , Temp sensori ", build a template and convert each piece of data into a natural language sentence to form a related document. Among them, the template sentence is: "Room {room id}, the collection start time is {start time}, and the end time is {end time}, the outdoor temperature is {Temp out}℃, humidity is {Hum out}%, the open state of air conditioner N is {air stateN}, the return air set temperature is {Temp seti}℃, the temperature of sensor N is {Temp sensori}℃, the energy consumption of IT equipment is {e1}, and the total energy consumption of air conditioning is {e2}".

[0088] (2) Construct the prompt words as follows to generate the return air temperature set value that meets the following conditions: the air conditioning energy consumption must be minimized and the return air temperature set value of each air conditioner must be lower than the alarm temperature. The output return air temperature set value is limited to the range of ±2 based on the current set value. In the standard JSON format: {"air conditioning point": "N", "return air temperature set value of air conditioner N": "Temp seti Return the return air temperature setpoint for each air conditioner and output the data from the knowledge base. If the return air temperature setpoint for the air conditioner cannot be generated based on the relevant data, output "MPC control is used";

[0089] (3) First, the search results are sorted in ascending order according to the total energy consumption parameter e2 of the air conditioner, and the relevant documents of the top 10 samples with the lowest energy consumption values ​​are retained first. Then, the prompt words are used as the guide words and the complete input context of the document content after orderly splicing and sorting is formed to form an enhanced model input containing energy-saving preference indications.

[0090] (4) The final control strategy is generated by calling the pre-trained language generation model (DeepSeek-v3) and combining it with enhanced context information. The language model does not need to be retrained or fine-tuned, and reasoning is performed directly through the program interface. To ensure that the generated results meet the high requirements of the air-conditioning control scenario for output consistency, parsability, and reliability, the temperature coefficient (temperature) is set to 0.5 and the cumulative probability truncation sampling (Nucleus Sampling) is set to 0.75 during the reasoning phase to limit the range of generated vocabulary to improve output quality, and the repetition penalty factor (repetition penalty factor) is set to 0.5. penalty ) is set to 1.

[0091] (5) By parsing the Temp in each record seti Field, extract the return air set temperature of all air conditioners in the order of air conditioner points, and obtain the numerical control quantity output sequence [Temp set 1. Temp set2 ,…,Temp setN ].

[0092] Generate log information, including the searched historical correlation data and the corresponding similarity, the control strategy generation method adopted, the current return air temperature setting value of the air conditioning equipment, and the generated return air temperature setting value.

[0093] The above-described specific embodiments merely describe preferred embodiments of the present invention and do not limit the scope of protection of the present invention. Any modifications, substitutions, and improvements made to the technical solution of the present invention by a person skilled in the art based on the textual description and drawings provided herein, without departing from the design concept and spirit of the present invention, shall fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. An energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms is characterized in that: The following steps are involved: Step S1: pre-processing historical operation data of the air-cooled air-conditioning unit, wherein the historical operation data includes computer room information, time information, air-conditioning information, IT equipment energy consumption information, outdoor information, and sensor information; Step S2: Obtain the current parameters of the computer room to be adjusted, including the computer room number, the start and stop status of the air conditioner, the start time and the end time, and perform a filtering operation on the historical operation data based on the current parameters to generate a candidate data set; Step S3: Vectorize the current outdoor temperature and humidity and IT equipment energy consumption in the candidate data set to build a vector database that supports RAG retrieval. Use similarity retrieval to obtain a preset number of data that are closest to the current state parameters of the computer room to be adjusted. Step S4: Based on the current set temperature of each air conditioner in the computer room to be adjusted, search for data in which the set temperature of each air conditioner is within a preset adjustment range of the current value, and construct historical correlation data; Step S5: Determine whether the amount of historical related data reaches a set threshold: If the threshold is reached, the historical correlation data is used to construct a knowledge base, which is combined with the control constraints as input and provided to the RAG generation module to output the control strategy; If the threshold is not reached or the RAG generation module outputs "adopt MPC control", the MPC-based backup control strategy is called; The method of using the RAG generation module to output the control strategy in step S5 includes: Step S511: text-processing the historical related data to build a knowledge base; Step S512: Construct a prompt word containing the following constraints: Output the air conditioning return air temperature setpoint that meets the minimum energy consumption and the sensor temperature is lower than the alarm temperature; The control range of the return air temperature setting value is limited to ±2℃ of the current setting value; The output uses the specified JSON format; The output must include the knowledge base data on which it is based; When no effective strategy can be generated, output "adopt MPC control"; Step S513: combining the knowledge base with the prompt words and inputting them into the RAG generation module to output the control strategy; The method of using MPC to generate a control strategy in step S5 includes: Step S521: Constructing an indoor temperature state equation based on system operation data and outdoor temperature and humidity; Step S522: using the least squares method to solve the state equation coefficients; Step S523: Design a multi-objective loss function including temperature control accuracy, energy efficiency optimization, and control stability. The expression of the loss function is: L = α1·(predicted temperature - target temperature)² + α2·(current control amount - last control amount)² + α3·(actual energy consumption - energy saving target)² in, 、 、 is the weight coefficient, which is dynamically adjusted according to the real-time operating status of the system; Step S524: In the control amount adjustment interval [Temp seti-2 , Temp seti+2 ], an optimization algorithm is used to solve the optimal control strategy; Step S6: converting the control strategy into a control signal to control the air-cooled air conditioner in real time, and generating complete log information including the control process; The control signal conversion in step S6 includes: For the control strategy generated by MPC: directly receive the control quantity sequence and adjust the return air temperature set value of the corresponding air conditioner; For the control strategy generated by RAG: parse the JSON format output, extract the device number and temperature set value, and generate an ordered sequence of temperature control instructions; Set up a data validation mechanism to return the backup instruction of "adopt MPC control" when the generated content is not standardized.

2. The energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms according to claim 1 is characterized in that: The historical operation data collection in step S1 is specifically as follows: The computer room information includes the computer room number; Time information includes the start time and end time of the collection; Air conditioning information includes the start and stop status of multiple air conditioners, return air set temperature, and total air conditioning energy consumption; IT equipment energy consumption information includes IT system load power consumption within the current time window; Outdoor information includes outdoor temperature and outdoor humidity; Sensor information includes temperature values ​​corresponding to multiple sensors and sensor alarm temperatures set by the user; The data collection time interval is 30 minutes, and the energy consumption value of air conditioners and IT equipment is obtained by calculating the difference between the cumulative values ​​of two adjacent time points to obtain the energy consumption increment of the period.

3. The energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms according to claim 1 is characterized in that: The pre-processing in step S1 includes: Missing value processing, outlier processing, duplicate value processing, data coding and energy consumption calculation; The missing value processing is used to delete sample group records with missing data; The outlier processing is used to identify and eliminate abnormal data points using statistical methods; The repeated value processing is used to monitor the frequency of repeated values ​​and generate an alarm log when the frequency exceeds a threshold; The data encoding is used to encode the start and stop status of the air conditioner using 0 and 1; The energy consumption calculation is used to calculate the time period increment of air conditioning energy consumption and IT equipment energy consumption.

4. The energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms according to claim 1 is characterized in that: The candidate data set generation method in step S2 is: Filter out the records in the historical operation data whose computer room number, air conditioner start / stop status, start time, and end time are completely consistent with the parameters of the current computer room to be adjusted.

5. The energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms according to claim 1 is characterized in that: The vector database construction in step S3 includes: Use FAISS to build a vector database; Calculate vector similarity using Euclidean distance; Perform Top-K similarity search to obtain the K candidate samples closest to the query vector, where K is a preset positive integer.

6. The energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms according to claim 1 is characterized in that: The preset adjustment interval in step S4 is defined as: The adjustment range of each air conditioner is [Temp seti-2 , Temp seti+2 ], where Temp seti is the current set temperature of the i-th air conditioner, i=1,2,...,N.

7. The energy-saving method for air-cooled air-conditioning units based on the integration of RAG and MPC algorithms according to claim 1 is characterized in that: The log information in step S6 includes: The historical related data and their similarity found in step S4; The control strategy generation method used; The current return air temperature setting value of each air conditioning device; The new return air temperature setpoint is generated.

Images