A multi-AI agent collaborative HVAC control method and system

Abstract

This invention relates to a multi-AI agent collaborative HVAC control method and system, belonging to the field of intelligent manufacturing equipment technology. It includes: a data acquisition module for collecting environmental data, load data, and equipment data from a data center; a multi-AI agent collaboration module for using multiple agents to perform data quality analysis, environmental trend prediction, safety boundary checks, pattern recognition, energy-saving strategy formulation, and end-point optimization based on environmental and equipment data, generating multi-dimensional decision suggestions; a decision coordination module for generating optimal decisions based on the multi-dimensional decision suggestions; an execution control module for generating control commands based on the optimal decisions, wherein the control commands are used to control the operation of HVAC equipment; an effect evaluation module for evaluating the execution effect of the optimal decisions; and a knowledge base management module for storing and retrieving the experience of generating optimal decisions, possessing the advantages of realizing intelligent and autonomous energy-saving control of data center HVAC.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing equipment technology, and in particular to a multi-AI agent collaborative HVAC control method and system. Background Technology

[0002] As a core infrastructure of modern information society, the operational efficiency and energy management of data centers directly affect the reliability and economy of data services. Among these, the Heating, Ventilation, and Air Conditioning (HVAC) system accounts for approximately 40%-50% of data center energy consumption, and optimizing its control strategies plays a crucial role in reducing overall energy consumption and improving system stability. However, traditional HVAC control methods suffer from limitations when facing the complex and ever-changing operating environment of data centers, including simplistic control strategies, insufficient learning capabilities, and difficulties in multi-objective optimization. These limitations make it difficult to meet the demands of modern data centers for intelligent and autonomous energy-saving control.

[0003] Traditional HVAC control methods primarily rely on fixed rules or simple PID control. These methods often fail to adjust control strategies promptly in the face of dynamic factors such as data center load fluctuations and changes in outdoor ambient temperature, leading to low system efficiency. For example, while PID control can stabilize system output to some extent, its parameter adjustments require manual intervention and are ill-suited to nonlinear and time-varying system characteristics. Furthermore, most existing HVAC control systems lack the ability to learn and optimize control strategies from historical operating data. This means the system cannot automatically adjust control parameters based on past operating experience to cope with similar or more complex operating scenarios. Therefore, control effectiveness is often limited to preset rules or empirical values, making it difficult to achieve further energy savings and performance improvements.

[0004] Therefore, there is a need to provide a multi-AI agent collaborative HVAC control method to achieve intelligent, autonomous, and energy-saving control of HVAC in data centers. Summary of the Invention

[0005] This invention provides a multi-AI agent collaborative HVAC control system, comprising: a data acquisition module for acquiring environmental data, load data, and equipment data from a data center; a multi-AI agent collaboration module for using multiple agents to perform data quality analysis, environmental trend prediction, safety boundary checks, pattern recognition, energy-saving strategy formulation, and end-point optimization based on environmental data, load data, and equipment data, generating multi-dimensional decision suggestions; a decision coordination module for generating optimal decisions based on the multi-dimensional decision suggestions; an execution control module for generating control commands based on the optimal decisions, wherein the control commands are used to control the operation of HVAC equipment; an effect evaluation module for evaluating the execution effect of the optimal decisions; and a knowledge base management module for storing and retrieving the experience of generating optimal decisions.

[0006] Furthermore, the multi-AI intelligent agent collaboration module includes: a data analysis intelligent agent, used to perform data quality analysis on the environmental data of the data center and generate optimized environmental data; a predictive analysis intelligent agent, used to predict future environmental data based on optimized environmental data and predict future load data based on load data; a security protection intelligent agent, used to perform security boundary checks and generate risk control strategies based on device data, optimized environmental data, and future environmental data; an efficiency optimization intelligent agent, used to generate energy-saving strategies based on device data, optimized environmental data, future environmental data, load data, and future load data; a self-learning intelligent agent, used to perform pattern recognition and generate historical experience strategies based on optimized environmental data and future environmental data; and an end-point optimization intelligent agent, used to generate end-point optimization strategies based on device data, optimized environmental data, future environmental data, load data, and future load data. The multi-dimensional decision recommendations include risk control strategies, energy-saving strategies, historical experience strategies, and end-point optimization strategies.

[0007] Furthermore, the environmental data of the data center includes outdoor temperature and humidity data and temperature and humidity data from multiple environmental monitoring locations. The data analysis agent performs data quality analysis on the environmental data of the data center and generates optimized environmental data, including: for each environmental monitoring location, determining the associated environmental monitoring locations and the temperature and humidity correlation time window between the environmental monitoring location and any associated environmental monitoring location, as well as the temperature and humidity calibration correlation coefficient, and generating a calibration correlation coefficient matrix; based on the temperature and humidity data from multiple environmental monitoring locations and the temperature and humidity correlation time window between each environmental monitoring location and any associated environmental monitoring location, generating a current correlation coefficient matrix; and based on the calibration correlation coefficient matrix and the current correlation coefficient matrix, performing data quality analysis on the environmental data and generating optimized environmental data.

[0008] Furthermore, the security protection agent performs security boundary checks and generates risk control strategies based on device data, optimized environmental data, and future environmental data. This includes: the security protection agent performs temperature, pressure, and load security boundary checks and generates risk control strategies based on device data, optimized environmental data, and future environmental data.

[0009] Furthermore, the efficiency optimization agent generates energy-saving strategies based on device data, optimized environmental data, future environmental data, load data, and future load data, including: generating multiple initial strategies; predicting the performance coefficient of each initial strategy based on device data, optimized environmental data, future environmental data, load data, and future load data; and generating an energy-saving strategy based on the performance coefficient of each initial strategy using a particle swarm optimization algorithm.

[0010] Furthermore, the self-learning agent performs pattern recognition based on optimized environmental data and future environmental data to generate historical experience strategies, including: acquiring historical HVAC operation data, wherein the historical HVAC operation data includes environmental data and optimal decisions for multiple historical time periods; determining similar historical time periods based on optimized environmental data, future environmental data, and environmental data for multiple historical time periods; and generating historical experience strategies based on the optimal decisions for similar historical time periods.

[0011] Furthermore, the end-point optimization agent generates an end-point optimization strategy based on device data, optimized environmental data, future environmental data, load data, and future load data. This includes generating the end-point optimization strategy through an end-point optimization model, based on device data, optimized environmental data, future environmental data, load data, and future load data. The end-point optimization strategy includes at least the operating parameters of the air conditioner and the fan wall.

[0012] Furthermore, the decision coordination module generates optimal decisions based on multi-dimensional decision suggestions, including: constructing a state space, wherein the state space includes device data, optimized environmental data, future environmental data, load data, and future load data; constructing an action space, wherein the action space includes the number of devices activated, water supply temperature, pump frequency, and fan speed; constructing value constraints for the action space based on multi-dimensional decision suggestions; designing a multi-objective weighted reward function; and generating optimal decisions based on the value constraints of the action space through a dual-Q network structure and an adaptive ε-greedy strategy.

[0013] Furthermore, the decision coordination module constructs value constraints for the action space based on multi-dimensional decision suggestions, including: constructing multiple candidate value constraints; constructing a fitness function, wherein the fitness function is related to the multi-dimensional decision suggestions; calculating the fitness value of the candidate value constraint for each candidate value constraint; and constructing value constraints for the action space based on the fitness values ​​of the candidate value constraints using a particle swarm optimization algorithm.

[0014] This invention provides a multi-AI agent collaborative HVAC control method, applied to the aforementioned multi-AI agent collaborative HVAC control system, comprising: collecting environmental data and equipment data from a data center; using multiple agents to perform data quality analysis, load and environmental trend prediction, security boundary checks, pattern recognition, energy-saving strategy formulation, and end-point optimization based on environmental data, load data, and equipment data, generating multi-dimensional decision suggestions; generating an optimal decision based on the multi-dimensional decision suggestions; generating control commands based on the optimal decision, wherein the control commands are used to control the operation of HVAC equipment; evaluating the execution effect of the optimal decision; and storing the experience of generating the optimal decision.

[0015] Compared to existing technologies, the multi-AI agent collaborative HVAC control method and system provided in this specification has at least the following advantages:

[0016] 1. The multi-AI agent collaboration module utilizes multiple agents, each with its own specific function. Starting with data quality analysis, it ensures the accuracy and reliability of collected environmental, load, and equipment data, laying a solid foundation for subsequent decision-making. Environmental trend prediction can anticipate environmental changes in advance, while safety boundary checks ensure the system operates within safe limits. Pattern recognition accurately grasps the system's operating mode, and energy-saving strategy formulation and end-point optimization improve system efficiency from different levels. The agents work collaboratively, integrating multi-dimensional information to generate multi-dimensional decision suggestions. The decision coordination module then generates the optimal decision based on these suggestions. This multi-faceted and comprehensive decision-making process avoids the limitations of single-agent decision-making, greatly improving the accuracy and comprehensiveness of decisions, enabling the HVAC control system to make more scientific and reasonable operational decisions based on the actual needs of the data center.

[0017] 2. The energy-saving strategy formulation and end-point optimization functions in the multi-AI intelligent agent collaboration module directly address the system's energy consumption issues. Energy-saving strategy formulation plans the system's operation mode at a holistic level to reduce energy consumption; end-point optimization finely adjusts the operating parameters of HVAC equipment to improve equipment efficiency. The execution control module generates control commands based on optimal decisions, precisely controlling the operation of HVAC equipment and avoiding ineffective operation and energy waste. Through this comprehensive energy-saving measure, the multi-AI intelligent agent collaborative HVAC control system can significantly reduce data center energy consumption, improve energy utilization efficiency, achieve high-efficiency system operation, and save significant operating costs for the data center.

[0018] 3. The knowledge base management module is responsible for systematically storing and easily retrieving the experience of generating optimal decisions. Each process data, key parameters, and final decision content for generating an optimal decision are stored systematically according to specific rules and classification systems. Over time, the knowledge base is continuously enriched and improved. When faced with new decision-making needs, the system can quickly retrieve historical experiences from the knowledge base for similar scenarios, providing reference and guidance for current decisions. Attached Figure Description

[0019] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0020] Figure 1 This is a block diagram of a multi-AI agent collaborative HVAC control system shown in one embodiment of this application;

[0021] Figure 2 This is a structural block diagram of a multi-AI agent collaboration module shown in one embodiment of this application;

[0022] Figure 3 This is a flowchart illustrating a multi-AI agent collaborative HVAC control method in one embodiment of this application. Detailed Implementation

[0023] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0024] Figure 1 This is a block diagram of a multi-AI agent collaborative HVAC control system shown in one embodiment of this application, as follows: Figure 1 As shown, a multi-AI agent collaborative HVAC control system may include a data acquisition module, a multi-AI agent collaboration module, a decision coordination module, an execution control module, an effect evaluation module, and a knowledge base management module.

[0025] The data acquisition module is used to collect environmental data, load data, and equipment data from the data center.

[0026] The environmental data for the data center includes outdoor temperature and humidity data as well as temperature and humidity data from multiple environmental monitoring locations.

[0027] Specifically, high-precision temperature and humidity sensors are deployed at strategic locations outside the data center to continuously collect outdoor temperature and humidity information in real time. Multiple temperature and humidity sensors are distributed and installed within the data center in different functional areas, near critical equipment, and in areas with high personnel activity to obtain real-time temperature and humidity conditions at these specific locations. Due to factors such as uneven heat distribution from equipment and complex airflow patterns within the data center, temperature and humidity can vary significantly across different locations. By collecting data from multiple monitoring points, a comprehensive understanding of the internal environment of the data center can be obtained, accurately locating potential localized hotspots or areas of abnormal humidity, providing detailed information for fine-tuning the HVAC system. For example, if excessively high temperatures are detected near a server rack, the system can specifically increase the airflow to that area or adjust the air conditioner's exhaust angle to ensure normal equipment operation.

[0028] Load data primarily reflects the operating load of various devices within the data center, including but not limited to power consumption and computing workload information for servers, storage devices, and network equipment. This load data is obtained in real-time by interfacing with the data center's equipment management system or power monitoring system. Load data is crucial for HVAC system control because the operating load of equipment directly affects its heat generation, which in turn influences temperature changes within the data center. For example, when the computing tasks handled by the data center increase and server load rises significantly, the heat generated by the equipment will also increase accordingly. In this case, the HVAC system needs to adjust its cooling capacity promptly based on the load data to prevent equipment performance degradation or damage due to excessively high temperatures.

[0029] Equipment data primarily involves the operating status and parameter information of the HVAC system itself, including the operating mode of air conditioning equipment (cooling, heating, ventilation, etc.), fan speed, compressor status, refrigerant pressure, supply water temperature, return water temperature, etc., as well as data such as the operating frequency and power of auxiliary equipment such as water pumps and fans. By installing corresponding sensors and monitoring devices on various key devices in the HVAC system, real-time operating data is collected and transmitted to the data acquisition module. This equipment data helps system administrators understand the operating status of the HVAC system in a timely manner and detect equipment failures or anomalies.

[0030] The multi-AI agent collaboration module is used to generate multi-dimensional decision recommendations by having multiple agents perform data quality analysis, environmental trend prediction, safety boundary checks, pattern recognition, energy-saving strategy formulation, and end-point optimization based on environmental data, load data, and device data.

[0031] In some embodiments, the multi-AI agent collaboration module may include a data analysis agent, a predictive analysis agent, a security protection agent, an efficiency optimization agent, a self-learning agent, and an end-point optimization agent. Figure 2 This is a structural block diagram of a multi-AI agent collaboration module shown in one embodiment of this application, as follows: Figure 2 As shown:

[0032] A data analytics agent is used to perform data quality analysis on environmental data in a data center and generate optimized environmental data.

[0033] Predictive analytics agents are used to predict future environmental data based on optimized environmental data and to predict future load data based on load data. For example, future environmental data can be predicted based on optimized environmental data using an environmental prediction model, and future load data can be predicted based on load data using a load prediction model. The environmental prediction model and the load prediction model can be long short-term memory network models.

[0034] The security protection agent is used to perform security boundary checks and generate risk control strategies based on device data, optimized environmental data, and future environmental data.

[0035] An efficiency optimization agent is used to generate energy-saving strategies based on device data, optimized environmental data, future environmental data, load data, and future load data.

[0036] A self-learning intelligent agent is used to perform pattern recognition and generate historical experience strategies based on optimized environmental data and future environmental data.

[0037] The edge optimization agent is used to generate edge optimization strategies based on device data, optimized environmental data, future environmental data, load data, and future load data. The multi-dimensional decision recommendations include risk control strategies, energy-saving strategies, historical experience strategies, and edge optimization strategies.

[0038] In some embodiments, the data analysis agent performs data quality analysis on environmental data from a data center and generates optimized environmental data, including:

[0039] For each environmental monitoring location, determine the associated environmental monitoring locations and the temperature and humidity correlation time window and temperature and humidity calibration correlation coefficient between the environmental monitoring location and any associated environmental monitoring location;

[0040] Generate a calibration correlation coefficient matrix, where each row vector of the calibration correlation coefficient matrix corresponds to an environmental monitoring location, and each element of the row vector represents the temperature and humidity calibration correlation coefficient between the environmental monitoring location and other environmental monitoring locations. The calibration correlation coefficient matrix has N-1 columns. If an environmental monitoring location is not associated with other environmental monitoring locations, the value of the element is 0, and N is the total number of environmental monitoring locations.

[0041] Based on temperature and humidity data from multiple environmental monitoring locations and the time window of temperature and humidity correlation between each environmental monitoring location and any associated environmental monitoring location, a current correlation coefficient matrix is ​​generated.

[0042] Based on the calibrated correlation coefficient matrix and the current correlation coefficient matrix, data quality analysis is performed on the environmental data, and optimized environmental data is generated.

[0043] Specifically, the correlation coefficient for temperature and humidity calibration can be determined according to the following procedure:

[0044] S11. Obtain temperature and humidity values ​​at multiple historical time points from multiple environmental monitoring locations in the data center;

[0045] S12. For any two environmental monitoring locations, calculate the initial correlation coefficient between the two environmental monitoring locations based on the temperature and humidity values ​​at multiple historical time points using the correlation coefficient (e.g., Pearson correlation coefficient) calculation formula.

[0046] S13. For any two environmental monitoring locations, if the initial correlation coefficient between the two environmental monitoring locations is greater than the correlation coefficient threshold (e.g., 0.5), then the two environmental monitoring locations are related environmental monitoring locations.

[0047] S14. Construct a fitness function, wherein the fitness function is related to the length of the temperature and humidity associated time window, the window time interval, and the temperature and humidity correlation coefficient between the environmental monitoring location corresponding to the length of the temperature and humidity associated time window and the associated environmental monitoring location. The window time interval represents the difference in the start time of two adjacent temperature and humidity associated time windows.

[0048] S15. For each environmental monitoring location and each associated environmental monitoring location, multiple particles are randomly generated. Each particle includes a window time interval and a temperature and humidity associated time window. The window time interval or temperature and humidity associated time window of any two particles may differ. For each particle, based on the particle's window time interval and temperature and humidity associated time window, the temperature and humidity values ​​of multiple environmental monitoring locations at multiple historical time points are processed to generate the temperature and humidity values ​​for each temperature and humidity associated time window corresponding to that environmental monitoring location and associated environmental monitoring locations. That is, the average temperature and humidity values ​​of the historical time points within the temperature and humidity associated time window are calculated to obtain the temperature and humidity values ​​of the temperature and humidity associated time window. The correlation coefficient (e.g., Pearson correlation coefficient) is calculated based on the temperature and humidity values ​​of each temperature and humidity correlation time window corresponding to the environmental monitoring location and associated environmental monitoring locations. The correlation coefficient of temperature and humidity between the environmental monitoring location and associated environmental monitoring locations corresponding to the particle is calculated, and then the fitness function value of the particle is calculated. Through the particle swarm optimization algorithm, the length and window time interval of the optimal temperature and humidity correlation time window corresponding to the environmental monitoring location and associated environmental monitoring locations are determined based on the fitness function value of each particle. The temperature and humidity correlation coefficient of the environmental monitoring location and associated environmental monitoring locations corresponding to the length and window time interval of the optimal temperature and humidity correlation time window is used as the temperature and humidity calibration correlation coefficient.

[0049] For example, the fitness function can be:

[0050]

[0051] in, For the fitness function, , and As weight, , and Greater than 0, ,For example, It is 0.5. It is 0.3. It is 0.2. This represents the temperature and humidity correlation coefficient between the normalized particle-corresponding environmental monitoring location and the associated environmental monitoring location. The normalized particles include the length of the temperature and humidity-related time window. This represents the window time interval included in the normalized particles.

[0052] Understandably, the above formula, by weighting and combining three factors, allows the fitness function to comprehensively evaluate the merits of each particle. The particle swarm optimization algorithm continuously adjusts the position and velocity of each particle based on its fitness function value to find the parameter combination that maximizes the result—that is, the optimal length and time interval of the temperature and humidity correlation time window. These determined parameters effectively eliminate the lag in temperature and humidity changes between the two correlated environmental monitoring locations and avoid excessive data processing caused by unreasonable parameter settings, thus providing more accurate and efficient data support for subsequent temperature and humidity correlation analysis and control.

[0053] In the process of filtering related locations, after obtaining historical temperature and humidity data from multiple environmental monitoring locations in S11, the initial correlation coefficient is calculated in S12, and related locations are filtered out in S13 using a threshold. This process accurately locates the pairs of locations with temperature and humidity correlation, providing an effective basis for subsequent analysis, avoiding unnecessary calculations for irrelevant locations, and improving analysis efficiency.

[0054] Regarding parameter optimization, the fitness function constructed in S14 comprehensively considers the length of the temperature and humidity correlation time window, the window time interval, and the corresponding temperature and humidity correlation coefficient, thus comprehensively and reasonably evaluating the advantages and disadvantages of different parameter combinations. In S15, particles containing different window time intervals and temperature and humidity correlation time windows are randomly generated. Using the particle swarm optimization algorithm, the optimal parameters are found based on the fitness function value of each particle. This process can automatically and efficiently determine the optimal temperature and humidity correlation time window length and window time interval, avoiding the subjectivity and blindness of manually setting parameters.

[0055] Calculating the average of historical data within the temperature and humidity correlation time window can effectively eliminate short-term fluctuations and noise interference, making the calculated temperature and humidity correlation coefficient more reflective of the true correlation between the temperature and humidity of the two related environmental monitoring locations.

[0056] The temperature and humidity data of each environmental monitoring location and associated environmental monitoring locations can be processed based on the length and time interval of the optimal temperature and humidity correlation time window corresponding to each environmental monitoring location and associated environmental monitoring locations to generate the current correlation coefficient matrix.

[0057] Calculate the cosine similarity between the calibrated correlation coefficient matrix and the current correlation coefficient matrix. If the cosine similarity is less than the cosine similarity threshold (e.g., 0.5), it is determined that there is an anomaly in the environmental data.

[0058] The temperature and humidity data of multiple environmental monitoring locations are corrected by a data correction model to generate corrected temperature and humidity data of multiple environmental monitoring locations. The optimized environmental data includes the corrected temperature and humidity data of multiple environmental monitoring locations. The data correction model can be a long short-term memory network model.

[0059] By identifying the associated locations, temperature and humidity correlation time windows, and calibrated correlation coefficients for each environmental monitoring location, and generating a calibration correlation coefficient matrix, the temperature and humidity correlation benchmark between locations was clarified, constructing a standardized data correlation framework and providing a precise reference for subsequent analysis. Secondly, based on real-time temperature and humidity data and correlation time windows, a current correlation coefficient matrix was generated, dynamically capturing the temperature and humidity correlation characteristics in the actual environment, enabling the analysis to reflect real environmental changes. By calculating the cosine similarity between the calibration matrix and the current matrix and setting a threshold to determine anomalies, automated and quantitative assessment of environmental data quality was achieved, enabling timely identification of data deviations or abnormal fluctuations, avoiding the subjectivity and lag of manual judgment. When anomalies are detected, a Long Short-Term Memory (LSTM) network model is used to correct the data. Leveraging its advantage in processing time-series data, it can effectively capture the long-term dependencies of temperature and humidity changes. The corrected data better reflects the real environmental state, improving the accuracy and reliability of the data.

[0060] In some embodiments, the security guardian agent performs security boundary checks based on device data, optimized environmental data, and future environmental data, and generates risk control strategies, including:

[0061] The safety protection intelligent agent performs temperature, pressure, and load safety boundary checks based on device data, optimized environmental data, and future environmental data, and generates risk control strategies.

[0062] Specifically, the safety protection agent performs safety boundary checks on temperature, pressure, and load. Based on the equipment's design parameters, historical operating data, and industry standards, it determines the safe range for the equipment in terms of temperature, pressure, and load. Then, it compares and analyzes the actual equipment data, optimized environmental data, and future environmental data against the safe range. For example, if the current equipment temperature is close to or exceeds the safe upper limit, and the future ambient temperature is trending upwards, while optimized environmental data shows poor heat dissipation, the agent will determine that the equipment has a risk of overheating. Based on these check and analysis results, the safety protection agent generates risk control strategies. For HVAC equipment, the risk control strategy specifies constraints on its operating parameters, such as limiting the equipment's maximum operating temperature, setting reasonable pressure fluctuation ranges, and specifying the maximum load capacity.

[0063] In some embodiments, the efficiency optimization agent generates energy-saving strategies based on device data, optimized environmental data, future environmental data, load data, and future load data, including:

[0064] Generate multiple initial strategies, which may include the operating parameters of HVAC equipment;

[0065] For each initial strategy, the performance coefficient of the initial strategy is predicted based on equipment data, optimized environmental data, future environmental data, load data, and future load data. For example, a performance prediction model can be used to predict the performance coefficient of the initial strategy based on equipment data, optimized environmental data, future environmental data, load data, and future load data. This performance prediction model employs a multi-layer neural network structure, including an input layer, hidden layers, and an output layer. The input layer receives equipment data, optimized environmental data, future environmental data, load data, and future load data. After entering the input layer, the data is passed to the hidden layer. The hidden layer contains multiple neurons that perform deep analysis and feature extraction on the input data. For example, convolutional operations are used to mine spatial features in the data, and a recurrent structure is used to process the temporally sequential load data and future load data to capture their changing patterns. After feature extraction and analysis are completed in the hidden layer, the output layer transforms the processing results into the performance coefficient of the initial strategy. This performance coefficient characterizes the ratio of the effective output (cooling or heating capacity) of the HVAC system to electrical energy.

[0066] Using the particle swarm optimization algorithm, energy-saving strategies are generated based on the performance coefficients of each initial strategy, where the energy-saving strategy can be the strategy with the highest performance coefficient.

[0067] In some embodiments, the self-learning agent performs pattern recognition based on optimized environmental data and future environmental data to generate historical experience strategies, including:

[0068] Acquire historical HVAC operation data, which includes environmental data and optimal decisions for multiple historical time periods. The optimal decision is the best operating parameter setting determined by a series of optimization algorithms or human experience within the corresponding historical time period, based on the environmental conditions and system operation objectives (such as ensuring indoor comfort while minimizing energy consumption).

[0069] Based on optimized environmental data, future environmental data, and environmental data from multiple historical time periods, similar historical time periods are identified.

[0070] Based on the optimal decisions made during similar historical time periods, historical experience strategies are generated.

[0071] Specifically, the self-learning agent determines which historical time periods are most similar to the current and future situations by calculating the similarity between optimized environmental data, future environmental data, and environmental data from multiple historical time periods. Similarity can be calculated using various methods, such as Euclidean distance and cosine similarity, by comparing the degree of difference between different data vectors to determine the degree of similarity.

[0072] A self-learning agent can generate historical experience strategies by weighted summation of the optimal decisions for similar historical time periods. The weights of the optimal decisions for similar historical time periods can be determined based on the similarity between environmental data.

[0073] In some embodiments, the end-point optimization agent generates an end-point optimization strategy based on device data, optimized environmental data, future environmental data, load data, and future load data, including:

[0074] The terminal optimization model generates terminal optimization strategies based on equipment data, optimized environmental data, future environmental data, load data, and future load data. The terminal optimization strategies include at least the operating parameters of air conditioners and fan walls.

[0075] Specifically, the end-point optimization model adopts a layered, progressive architecture, encompassing a data fusion layer, an analysis and prediction layer, and a strategy decision-making layer. The data fusion layer, as the foundation of the model, is responsible for integrating equipment data, optimized environmental data, future environmental data, load data, and future load data. The analysis and prediction layer uses machine learning algorithms to deeply mine the fused data, predicting future operating trends of the equipment and environmental changes. The strategy decision-making layer, based on the analysis and prediction results and combined with preset optimization objectives (such as minimum energy consumption and optimal comfort), generates end-point optimization strategies. First, the data fusion layer receives various types of data and performs cleaning, alignment, and fusion, constructing a unified data format for subsequent use. Next, the analysis and prediction layer uses time series analysis, neural networks, and other algorithms to predict the future operating status of the equipment (such as the probability of air conditioner failures and changes in fan wall efficiency). Then, based on the prediction results, the strategy decision-making layer searches for the optimal solution using optimization algorithms (such as genetic algorithms and particle swarm optimization) while meeting equipment safety constraints and environmental comfort requirements. The end-point optimization strategy focuses on adjusting the operating parameters of the air conditioner and fan wall, aiming to achieve efficient and energy-saving system operation. For air conditioning, the terminal optimization strategy includes dynamic temperature setting. Based on future ambient temperature forecasts and changes in indoor load, the set temperature is appropriately increased or decreased while ensuring occupant comfort, thus reducing cooling / heating energy consumption. Simultaneously, fan speed is adjusted: when the room is crowded and the load is high, the fan speed is increased to quickly regulate temperature and humidity; when the load is low, the fan speed is decreased to save energy. For wall-mounted fans, the terminal optimization strategy optimizes fan speed and airflow. Based on future air quality forecasts and regional ventilation needs, fan operation is precisely controlled to avoid excessive ventilation and energy waste, ensuring indoor air circulation meets health standards. Through these parameter adjustments, the terminal optimization strategy effectively balances energy consumption and comfort, improving overall system performance.

[0076] The decision coordination module is used to generate optimal decisions based on multi-dimensional decision recommendations.

[0077] Specifically, it includes:

[0078] Construct a state space, which includes device data, optimized environmental data, future environmental data, load data, and future load data;

[0079] Construct the action space, which includes the number of devices activated, water supply temperature, pump frequency, and fan speed;

[0080] Based on multi-dimensional decision-making suggestions, value constraints are constructed for the action space;

[0081] The design incorporates a multi-objective weighted reward function, specifically addressing multiple objectives such as energy saving, safety, stability, and economy. The energy saving objective is achieved by minimizing PUE (Power Usage Effectiveness) and maximizing COP (Coefficient of Performance). In the reward function, a decrease in PUE and an increase in COP will yield corresponding positive rewards. The safety objective requires parameters such as temperature, pressure, and load to be within safe ranges; if all parameters are within safe ranges, a safety reward will be given. The stability objective aims to reduce system fluctuations and avoid frequent start-ups and shutdowns; a stability reward will be obtained when the system is running stably. The economy objective aims to minimize operating costs; a reduction in operating costs will yield an economy reward. By setting appropriate weights for different objectives, a multi-objective weighted reward function is formed. During decision-making, each objective is comprehensively considered, weighing the pros and cons, and selecting the action that yields the maximum total reward, thereby achieving balanced optimization of multiple objectives.

[0082] By employing a dual-Q network structure and an adaptive ε-greedy strategy, the algorithm generates optimal decisions based on the value constraints of the action space. Specifically, the dual-Q network structure consists of two Q-networks: one for selecting actions and the other for evaluating the value of actions. Through interactive updates, the estimated Q-values ​​are made closer to the true values. The adaptive ε-greedy strategy balances exploration and exploitation. In the early stages of training, when the ε-value is large, the algorithm randomly selects actions to explore and discover better strategies. As training progresses, the ε-value gradually decreases, and the algorithm utilizes learned knowledge to select the optimal action. Within the value constraints of the action space, the algorithm calculates the Q-value of each action based on the current state using the dual-Q network structure, and combines this with the adaptive ε-greedy strategy to select actions. Through continuous learning and optimization, the algorithm ultimately generates optimal decisions that achieve multi-objective optimization.

[0083] In some embodiments, the decision coordination module constructs value constraints for the action space based on multi-dimensional decision suggestions, including:

[0084] Multiple candidate value constraints are constructed, where the candidate solutions cover different combinations of value ranges for various key parameters in the action space. For example, for the temperature setpoint, multiple different value range candidates are generated, such as 20-22℃ and 21-23℃; for the fan speed, different candidates are generated, such as 800-1000 rpm and 900-1100 rpm.

[0085] A fitness function is constructed, which is related to multi-dimensional decision recommendations. Specifically, the risk control strategy specifies constraints on equipment operating parameters, such as maximum operating temperature, reasonable pressure fluctuation range, and maximum load capacity. The fitness function assesses the risk of candidate value constraints based on these constraints. If a parameter in a candidate value constraint exceeds the set range, a significant penalty is imposed. For example, when the temperature setpoint exceeds the equipment's maximum operating temperature, a fixed negative score is awarded for every 1°C exceeding the limit; similarly, if the pressure fluctuation exceeds the reasonable range, points are deducted according to the degree of exceedance. This reduces the fitness value of candidate value constraints with high risks. Combined with the system energy consumption model, the performance coefficient under each candidate value constraint is calculated. The closer the maximum performance coefficient under a candidate value constraint is to the performance coefficient of the energy-saving strategy, the larger the fitness function value, measuring the similarity between the candidate value constraint and the optimal decision over similar historical time periods. The higher the similarity, the more likely the candidate value constraint is to bring good running results, and the fitness function will give it a higher score; conversely, if it deviates greatly from historical successful cases, its fitness value will be reduced. At the same time, the similarity between the candidate value constraint and the final optimization strategy is measured. The higher the similarity, the higher the fitness function will give it a higher score; conversely, if it deviates greatly from the final optimization strategy, its fitness value will be reduced.

[0086] For each candidate's value constraint, calculate the fitness value of the candidate's value constraint;

[0087] By using the particle swarm optimization algorithm, value constraints for the action space are constructed based on the fitness values ​​of the candidate value constraints.

[0088] The execution control module is used to generate control commands based on optimal decisions, and these control commands are used to control the operation of HVAC equipment.

[0089] Specifically, the execution control module analyzes the parameters in the optimal decision to clarify the operating state that the HVAC equipment needs to achieve, such as the air conditioner's temperature setpoint, fan speed, and air volume. Then, based on these specific parameters and the control protocol and interface specifications of the HVAC equipment, the abstract optimal decision is transformed into a series of concrete, executable control commands.

[0090] Once control commands reach the HVAC equipment, the equipment will make precise adjustments according to the commands. For example, the air conditioner will adjust the indoor temperature to the set value according to the command, and the fan wall will operate at the specified speed and air volume, thereby enabling the entire HVAC system to operate efficiently and stably according to optimal decisions.

[0091] The effectiveness evaluation module is used to evaluate the effectiveness of the optimal decision.

[0092] Specifically, after the optimal decision is made and the HVAC equipment is put into operation, the effect evaluation module comprehensively collects actual operating data, such as environmental parameters like indoor temperature, humidity, and air quality, as well as status information like equipment energy consumption and operating time. This actual data is then carefully compared with the expected goals before the decision was implemented, such as determining whether the indoor temperature has accurately reached the set value and whether energy consumption is within a reasonable reduction range.

[0093] The knowledge base management module is used to store and retrieve the experience of generating optimal decisions.

[0094] Specifically, the knowledge base management module stores the process data, key parameters, and final decision content of each optimal decision-making process in an orderly manner according to specific rules and classification systems. When faced with new decision-making needs, this module can quickly retrieve historical best-practice decision-making experiences in similar scenarios. By analyzing the success factors and potential problems in these experiences, it provides reference and guidance for current decisions, avoiding repeating mistakes and improving the accuracy and efficiency of decision-making. Moreover, as the system continues to operate and decision-making experience accumulates, the knowledge base will be continuously updated and enriched, becoming an important support for improving the intelligent decision-making level of the entire HVAC system.

[0095] Figure 3 This is a flowchart illustrating a multi-AI agent collaborative HVAC control method in one embodiment of this application, as shown below. Figure 3 As shown, a multi-AI agent collaborative HVAC control method may include the following steps:

[0096] Collect environmental data, load data, and equipment data from the data center;

[0097] By using multiple intelligent agents based on environmental data, load data, and device data, the system performs data quality analysis, load and environmental trend prediction, security boundary checks, pattern recognition, energy-saving strategy formulation, and end-point optimization to generate multi-dimensional decision recommendations.

[0098] Generate optimal decisions based on multi-dimensional decision recommendations;

[0099] Based on the optimal decision, control commands are generated, which are used to control the operation of HVAC equipment;

[0100] Evaluate the effectiveness of implementing the optimal decision;

[0101] Store the experience of generating optimal decisions.

[0102] For a more detailed description of a multi-AI agent collaborative HVAC control method, please refer to the relevant description of a multi-AI agent collaborative HVAC control system, which will not be repeated here.

[0103] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A multi-AI agent collaborative HVAC control system, characterized in that, The method comprises the following steps: a data acquisition module is used to acquire environmental data, load data and equipment data of the data center; a multi-AI agent cooperation module is used to perform data quality analysis, environmental trend prediction, security boundary check, pattern recognition, energy-saving strategy formulation and end optimization based on the environmental data, load data and equipment data through multiple agents, and generate multi-dimensional decision suggestions; a decision coordination module is used to generate an optimal decision based on the multi-dimensional decision suggestions; an execution control module is used to generate a control instruction based on the optimal decision, wherein the control instruction is used to control the operation of the heating and ventilation equipment; an effect evaluation module is used to evaluate the execution effect of the optimal decision; a knowledge base management module is used to store and retrieve the generation experience of the optimal decision; the multi-AI agent cooperation module comprises: a data analysis agent used to perform data quality analysis on the environmental data of the data center and generate optimized environmental data; a prediction analysis agent used to predict future environmental data based on the optimized environmental data and predict future load data based on the load data; a security guardian agent used to perform security boundary check based on the equipment data, the optimized environmental data and the future environmental data and generate a risk control strategy; an efficiency optimization agent used to generate an energy-saving strategy based on the equipment data, the optimized environmental data, the future environmental data, the load data and the future load data; a self-learning agent used to perform pattern recognition based on the optimized environmental data and the future environmental data and generate a historical experience strategy; an end optimization agent used to generate an end optimization strategy based on the equipment data, the optimized environmental data, the future environmental data, the load data and the future load data, wherein the multi-dimensional decision suggestions comprise the risk control strategy, the energy-saving strategy, the historical experience strategy and the end optimization strategy; the self-learning agent performs pattern recognition based on the optimized environmental data and the future environmental data and generates a historical experience strategy, comprising: obtaining historical heating operation data, wherein the historical heating operation data comprises environmental data of multiple historical time periods and optimal decisions; calculating the similarity between the optimized environmental data, the future environmental data and the environmental data of the multiple historical time periods to determine similar historical time periods; generating a historical experience strategy based on the optimal decisions of the similar historical time periods.

2. The multi-AI agent cooperative HVAC control system of claim 1, wherein, The environmental data of the data center comprises outdoor temperature and humidity data and temperature and humidity data of multiple environmental monitoring positions; the data analysis agent performs data quality analysis on the environmental data of the data center and generates optimized environmental data, comprising: for each environmental monitoring position, determining associated environmental monitoring positions of the environmental monitoring position, a temperature and humidity correlation time window and a temperature and humidity calibration correlation coefficient between the environmental monitoring position and any one of the associated environmental monitoring positions; generating a calibration correlation coefficient matrix; generating a current correlation coefficient matrix based on the temperature and humidity data of the multiple environmental monitoring positions and the temperature and humidity correlation time window between each environmental monitoring position and any one of the associated environmental monitoring positions. Based on the calibration correlation coefficient matrix and the current correlation coefficient matrix, data quality analysis is performed on the environment data, and optimized environment data is generated.

3. The multi-AI agent cooperative HVAC control system of claim 1, wherein, The security daemon intelligent agent performs security boundary checks based on the device data, the optimized environment data, and the future environment data, and generates a risk control strategy, including: The security daemon intelligent agent performs temperature, pressure, and load security boundary checks based on the device data, the optimized environment data, and the future environment data, and generates a risk control strategy.

4. The multi-AI agent cooperative HVAC control system of claim 1, wherein, The efficiency optimization intelligent agent generates an energy-saving strategy based on the device data, the optimized environment data, the future environment data, the load data, and the future load data, including: A plurality of initial strategies are generated. For each initial strategy, the performance coefficient of the initial strategy is predicted based on the device data, the optimized environment data, the future environment data, the load data, and the future load data. An energy-saving strategy is generated through a particle swarm optimization algorithm based on the performance coefficient of each initial strategy.

5. The multi-AI agent cooperative HVAC control system of claim 1, wherein, The end optimization intelligent agent generates an end optimization strategy based on the device data, the optimized environment data, the future environment data, the load data, and the future load data, including: An end optimization strategy is generated through an end optimization model based on the device data, the optimized environment data, the future environment data, the load data, and the future load data, wherein the end optimization strategy at least includes the operating parameters of the air conditioner and the fan wall.

6. The multi-AI agent cooperative HVAC control system of any one of claims 1-5, wherein, The decision coordination module generates an optimal decision based on multi-dimensional decision suggestions, including: A state space is constructed, wherein the state space includes device data, optimized environment data, future environment data, load data, and future load data; An action space is constructed, wherein the action space includes the number of device openings, water supply temperature, pump frequency, and fan speed; The value constraints of the action space are constructed based on multi-dimensional decision suggestions; A multi-objective weighted reward function is designed; An optimal decision is generated through a double Q network structure and an adaptive epsilon-greedy strategy based on the value constraints of the action space.

7. The multi-AI agent cooperative HVAC control system of claim 6, wherein, The decision coordination module constructs the value constraints of the action space based on multi-dimensional decision suggestions, including: A plurality of candidate value constraints are constructed; An adaptive function is constructed, wherein the adaptive function is related to multi-dimensional decision suggestions; For each candidate value constraint, the adaptive value of the candidate value constraint is calculated; The value constraints of the action space are constructed through a particle swarm optimization algorithm based on the adaptive values of the candidate value constraints. 8.A method for multi-AI agent cooperative HVAC control, characterized in that, The multi-AI intelligent agent cooperation HVAC control system of claim 1, comprising: Collecting environment data and device data of a data center; Performing data quality analysis, load and environment trend prediction, security boundary check, pattern recognition, energy-saving strategy formulation, and end optimization based on the environment data, load data, and device data through a plurality of intelligent agents to generate multi-dimensional decision suggestions; Generating an optimal decision based on multi-dimensional decision suggestions; Generating a control instruction based on the optimal decision, wherein the control instruction is used to control the operation of the heating and ventilation equipment; Evaluating the execution effect of the optimal decision; Storing and retrieving the generation experience of the optimal decision.

Images