AI-based energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning
By constructing a deep reinforcement learning agent based on mechanism fusion reinforcement learning, and combining the physical mechanism of the entire central air conditioning system with multi-source data, the problems of poor adaptability of fixed thresholds and insufficient generalization of pure data-driven AI in central air conditioning control are solved, and energy-saving optimization and stable control of the entire system are achieved.
Patent Information
- Application Number
- CN202611399364.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing central air conditioning control technologies suffer from poor adaptability to fixed thresholds, large inertial lag that easily leads to energy waste, insufficient generalization of pure data-driven AI, and easy failure under extreme operating conditions.
A control method based on mechanism fusion reinforcement learning is adopted. By collecting multi-source data in real time, a deep reinforcement learning agent is constructed. Combined with the physical mechanism constraints of the entire central air conditioning system, dynamic control parameters are output to achieve coordinated optimization of the entire system, avoid invalid commands, and carry out advanced regulation by combining weather forecasts and personnel load predictions.
It enables coordinated and optimized operation of all devices across the entire chain, adapts to dynamic changes in the system under different operating conditions, improves energy-saving effect and control reliability and stability, reduces manual dependence and equipment wear and tear, and is suitable for various public building scenarios.
Smart Images

Figure CN123110722A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of AI control technology for energy saving in building central air conditioning, and in particular to an AI energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning. Background Technology
[0002] Central air conditioning systems in public buildings are a core component of building energy consumption, accounting for a significant proportion of the total operating energy consumption of public buildings, making them a key target for energy conservation management. With the increasing sophistication of public building operations and the advancement of dual-carbon management requirements, the core needs for central air conditioning system control have shifted from basic temperature assurance to a multi-objective synergy encompassing thermal comfort standards, reduced operating energy consumption, and less reliance on manual maintenance. The industry is gradually transitioning from traditional manual maintenance to intelligent dynamic management.
[0003] Currently, there are two main technical solutions in the field of central air conditioning control. The first is the traditional control scheme based on fixed thresholds combined with PID regulation. This scheme uses the experience of operation and maintenance personnel to set fixed control thresholds for parameters such as chilled water outlet temperature, water pump operating frequency, and supply air temperature. Stable parameter control is achieved through closed-loop PID regulation. It has the advantages of simple logic, low deployment cost, and strong basic reliability, and is the most widely used control mode in the current central air conditioning systems of public buildings.
[0004] However, central air conditioning systems have the characteristics of large inertia, long lag, and strong coupling of multiple devices. Fixed thresholds cannot adapt to real-time changes in operating conditions such as outdoor weather fluctuations, dynamic changes in indoor occupant load, and equipment performance degradation. This can easily lead to problems such as overcooling, overheating, and high energy consumption in the distribution system. Furthermore, they rely heavily on maintenance personnel to manually adjust parameters periodically, resulting in high labor costs. Energy-saving effects are limited to local components, and it is impossible to achieve system-wide collaborative optimization.
[0005] The second type is a purely data-driven AI control solution. This solution trains machine learning or reinforcement learning models based on collected historical operating data, and the models output dynamic control parameters. Under normal operating conditions covered by the training data, it can achieve energy-saving effects better than fixed thresholds. However, this solution relies entirely on data correlation to mine control patterns and does not incorporate the physical operating mechanism of central air conditioning equipment. It has an inherent defect of insufficient generalization ability. Under operating conditions where the training data is not covered, such as extreme weather or partial equipment malfunctions, it is easy to output invalid commands that violate the operating boundaries of the equipment, resulting in control failure, excessive comfort levels, or even affecting the safety of equipment operation, making it difficult to implement stably on a large scale. Summary of the Invention
[0006] The present invention proposes an AI-based energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning, in order to solve the problems mentioned in the prior art, such as poor adaptability of fixed thresholds, large inertial lag that easily leads to energy waste, insufficient generalization of pure data-driven AI, and easy failure under extreme conditions.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning, comprising the following steps: S1. Real-time acquisition of rated mechanism parameters and operating status data of all equipment in the building's central air conditioning system, synchronous acquisition of indoor and outdoor temperature and humidity data, occupancy rate data of people in each area, and weather forecast data; Time alignment, outlier cleaning and standardization preprocessing of the acquired multi-source data. S2. Construct a deep reinforcement learning control framework that embeds the physical mechanism constraints of the entire central air conditioning system. Take the chiller unit's cooling efficiency characteristics, chilled water cooling water heat exchange mechanism, water pump and fan power consumption characteristics, and terminal thermal comfort balance mechanism as prior constraints, and integrate them into the state space construction, action space mapping, and reward function calculation of the deep reinforcement learning agent to avoid outputting control commands that violate the physical operating boundaries of the equipment. S3. Divide the central air conditioning system into full-area control levels, covering the chiller group control layer on the cold source side, the chilled cooling water distribution system control layer, the fresh air air conditioning unit air system control layer, and the terminal fan coil control layer, and clarify the adjustable control variable range and equipment safety operation constraints corresponding to each level; S4. Based on the preprocessed real-time dynamic data, the state input of the reinforcement learning agent is updated continuously. With the requirement of meeting indoor thermal comfort as the constraint and the goal of minimizing the overall energy consumption of the system as the optimization objective, the reinforcement learning agent outputs the dynamic control parameters of each level of equipment hourly to replace the traditional fixed threshold control logic, thereby realizing the dynamic optimization and energy-saving control of the entire link of the central air conditioning system.
[0008] Furthermore, the rated mechanism parameters collected in step S1 include the refrigeration efficiency characteristic curves of the chiller unit under different load rates and different cooling water inlet temperatures, the flow rate and head characteristic curves and the flow rate and power consumption characteristic curves of the water pump, the air volume and air pressure characteristic curves and the air volume and power consumption characteristic curves of the fan, and the heat exchange efficiency characteristic parameters of the terminal heat exchange coil; the collected operating status data include the real-time operating power, operating frequency, fluid medium inlet and outlet temperatures, inlet and outlet pressures, and regulating valve opening data of each device. The time granularity of all data collection is uniformly 1 minute, and outlier cleaning is completed by sliding window filtering.
[0009] Furthermore, the deep reinforcement learning agent constructed in step S2 adopts a deep deterministic policy gradient network architecture. In addition to the preprocessed real-time collected data, its state space also includes the theoretical operating efficiency values of each device, the theoretical minimum energy consumption benchmark value of the system under the current working condition, and the thermal comfort deviation value calculated in real time based on the physical mechanism model as core state variables, thereby compressing the agent's exploration space for meaningless and ineffective actions.
[0010] Furthermore, the action space of the deep reinforcement learning agent corresponds to the adjustable range of control variables at each level. Before all candidate actions are input into the control execution module, they must be verified for compliance by the embedded physical mechanism model. Candidate actions that exceed the safe operation boundary of the equipment or do not conform to the basic physical laws of heat exchange are masked out, and only control actions that conform to the mechanism constraints are retained as the set of output actions, so as to avoid outputting invalid or even harmful control commands under extreme working conditions.
[0011] Furthermore, the reward function in step S2 adopts a multi-objective weighted design, which includes three core parts: system operation energy consumption reward, indoor thermal comfort deviation penalty, and mechanism consistency penalty. The mechanism consistency penalty is calculated based on the deviation between the actual operating parameters of the equipment and the theoretical parameters output by the physical mechanism model. When the deviation exceeds a preset threshold, a penalty is imposed on the agent to guide the agent to output a control strategy that conforms to the physical operating law.
[0012] Furthermore, in step S4, the indoor thermal comfort constraint uses the predicted average voting value as the evaluation index. The comfort zone threshold is dynamically adjusted according to the occupancy rate of each area. In low-load areas where the occupancy rate is lower than the preset threshold, the temperature control range is appropriately relaxed. Under the premise of ensuring the thermal comfort experience of people when they are indoors, the energy-saving potential of non-core areas is fully explored.
[0013] Furthermore, in step S4, when updating the status input, hourly weather forecast data for the next 2 to 4 hours and data on changes in the indoor occupancy rate of people in each region based on historical patterns are simultaneously incorporated. By using the embedded physical mechanism model to predict the trend of system cooling load changes, the operating parameters of each device are adjusted in advance to offset the control deviation caused by the large inertia and long lag characteristics of the central air conditioning system.
[0014] Furthermore, the deep reinforcement learning agent adopts a deployment mode that combines offline pre-training with online fine-tuning. During the offline training phase, historical operating data and full-condition simulation data generated by the mechanism simulation model are integrated to complete the initial policy convergence of the agent. After going online, the agent is fine-tuned with a small learning rate using real-time collected operating data at fixed time intervals. During the online fine-tuning process, the physical mechanism constraint module is kept in effect to avoid the control policy from drifting randomly.
[0015] Furthermore, the dynamic control parameters of each device output by the reinforcement learning agent need to be processed by the instruction smoothing module before being sent out for execution. The instruction smoothing module sets a threshold for the adjustment range of control parameters at adjacent times, limiting the single adjustment range of device operating frequency, valve opening, and temperature setpoint, so as to avoid the problem of frequent device start-stop and drastic parameter fluctuations, and ensure the service life of the device while achieving energy-saving control.
[0016] Furthermore, the control method continuously performs self-verification of control effect during operation, compares the actual operating energy consumption of the system with the benchmark energy consumption calculated based on the mechanism model under the same operating conditions in real time, and calculates the dynamic energy saving rate. When it is detected that the indoor thermal comfort parameters continuously exceed the constraint range or the equipment operating parameters touch the safety boundary, the backup control mechanism is automatically triggered to switch the system to the preset safe operation mode, ensuring the reliability and stability of the central air conditioning system control.
[0017] Compared with existing technologies, the beneficial effects of this invention are: This invention constructs a comprehensive hierarchical control architecture covering the central air conditioning system's cold source, distribution, air system, and terminals. Combined with a rolling dynamic optimization mechanism using reinforcement learning agents, it incorporates weather forecasts and personnel load prediction data for proactive regulation. This allows it to adapt to dynamic changes in the system under different operating conditions, achieving coordinated and optimized operation of all devices across the entire chain. It eliminates the reliance on manual threshold settings and periodic maintenance adjustments, effectively solving the shortcomings of existing fixed threshold control, such as large inertia, long lag, poor adaptability, fragmented energy-saving effects, and high dependence on manual intervention. It fully taps the energy-saving potential of the entire system while ensuring indoor thermal comfort standards are met.
[0018] This invention uses the physical operating mechanism of the entire central air conditioning system as a priori constraint, deeply integrating it into the entire process of state space construction, action mapping selection, and reward function calculation of a reinforcement learning agent. This imposes physical boundary constraints on the agent's exploration space, preventing the agent from outputting invalid commands that violate the operating rules of the equipment. At the same time, it combines full-condition mechanism simulation data for offline pre-training, which greatly improves the agent's adaptability to unseen conditions, especially extreme conditions. This effectively solves the shortcomings of existing pure data-driven AI control, such as insufficient generalization ability and easy failure under extreme conditions, and significantly improves the reliability and stability of AI control.
[0019] This invention, through the design of smooth control command processing, self-verification of operating effects, and a fallback mechanism, can avoid losses caused by drastic fluctuations in equipment parameters. In the event of abnormal operating conditions, it automatically switches to a safe mode to ensure stable system operation. At the same time, it adopts a deployment mode that combines offline pre-training with online fine-tuning, reducing the difficulty of implementation and deployment. It can be adapted to central air conditioning systems of public buildings of different brands and service years, and can be quickly deployed without large-scale modification of existing hardware. It can be widely used in various public building scenarios such as office buildings, commercial complexes, hospitals, and transportation hubs, and has strong replicability and industry promotion value. Attached Figure Description
[0020] Figure 1 This is the overall flowchart of the AI energy-saving control of building central air conditioning based on mechanism fusion reinforcement learning proposed in this invention; Figure 2 This is a flowchart of the multi-source data acquisition and standardized preprocessing process of this invention; Figure 3 This is a flowchart illustrating the construction of the mechanism-integrated reinforcement learning control framework of this invention. Figure 4 This is a flowchart of the system status rolling update and dynamic optimization control of the present invention; Figure 5 This is a flowchart illustrating the closed-loop process of model deployment fine-tuning and instruction secure execution in this invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] Reference Figures 1 to 5 This invention discloses an AI-based energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning, comprising the following execution flow: First, a multi-source data acquisition module is deployed to connect with existing equipment monitoring points of the building's central air conditioning system, building automation system, indoor environmental sensors, personnel counting devices, and meteorological data interfaces. It collects the rated mechanism parameters and operating status data of the entire central air conditioning system in real time, and simultaneously collects multi-dimensional data such as indoor and outdoor temperature and humidity, occupancy rate of personnel in each area, and future weather forecasts. The collected data first enters the data preprocessing module, which unifies data from different sources and collection frequencies to the same time dimension by timestamp alignment. It uses the 3σ criterion combined with sliding window filtering to identify and remove outliers and missing values in the data collection. Data with different dimensions is standardized by z-score processing to solve the problem of heterogeneous multi-source data and poor data quality affecting control accuracy. Subsequently, a mechanism model library was built to sort out the physical operation laws of chillers, water pumps, fans and terminal heat exchange coils, forming a set of quantifiable and calculable mechanism models. The mechanism models were then embedded as prior constraints into a deep reinforcement learning control framework to break the inherent limitation of pure data-driven models lacking physical common sense. Then, the entire control hierarchy is divided according to the physical links of the central air conditioning system, and the adjustable control variables and safety constraint boundaries of each level from the cold source to the terminal are clearly defined, solving the problems of traditional control that can only optimize locally and the fragmentation of energy saving. Finally, in the control logic execution stage, state variables are continuously input into the reinforcement learning agent based on real-time updated dynamic data. The agent takes meeting the standard requirements for indoor thermal comfort as a hard constraint and minimizing the overall energy consumption of the system as the optimization goal, and outputs the dynamic control parameters of each level of equipment step by step. It directly connects to the building automation system to issue execution, replacing the traditional fixed threshold control logic. Through continuous rolling optimization, the entire system achieves dynamic optimization, solving the problems of fixed thresholds being unable to adapt to dynamic operating conditions and control deviations caused by large inertial lag.
[0023] This invention also discloses the specific parameters and preprocessing rules for the data acquisition process. The collected rated mechanism parameters cover the inherent physical characteristics of the core equipment across the entire chain, including the cooling efficiency characteristic curves of the chiller unit under different load rates and different cooling water inlet temperatures, the flow rate and head characteristic curves and the flow rate and power consumption characteristic curves of the water pump, the air volume and air pressure characteristic curves and the air volume and power consumption characteristic curves of the fan, and the heat exchange efficiency characteristic parameters of the terminal heat exchange coil. The above parameters are extracted from the equipment manufacturer's manual and third-party performance test reports, forming the basis for the calculation of the equipment mechanism model. The collected operational status data includes real-time operating power, operating frequency, inlet and outlet temperatures and pressures of the fluid medium, and valve opening data for each device. The time granularity of all data acquisition is uniformly set to 1 minute. Outlier cleaning is completed through a sliding window filter with a length of 5 time steps. Data within the window that deviates from the mean by more than 3 times the standard deviation is replaced with the window mean to avoid erroneous output of control commands due to data acquisition errors, thus providing a reliable data foundation for subsequent mechanism calculations and agent decision-making.
[0024] This invention also discloses the basic architecture and state space construction rules of the deep reinforcement learning agent. The deep reinforcement learning agent adopted uses a deep deterministic policy gradient network architecture, which is suitable for handling control problems of continuous action space such as central air conditioning. It can output continuously adjustable equipment control parameters and avoid the problem of insufficient control accuracy caused by discrete actions. In addition to the preprocessed real-time collected data, the state space of the intelligent agent also includes the theoretical operating efficiency values of each device, the theoretical minimum energy consumption benchmark value of the system under the current operating conditions, and the thermal comfort deviation value calculated in real time based on the physical mechanism model as core state variables. By introducing the benchmark value obtained from the mechanism calculation, the intelligent agent is guided to quickly explore the optimal control direction, which greatly reduces the exploration space of the intelligent agent for meaningless and ineffective actions, and improves the training convergence speed and control stability of the intelligent agent after going online.
[0025] This invention also discloses the construction and mechanism verification rules for the action space of the intelligent agent. The action space of the intelligent agent corresponds one-to-one with the physical adjustable range of control variables at each level, covering all automatically adjustable device parameters. Before inputting all candidate actions into the control execution module, they must be verified for compliance by the embedded physical mechanism model. Candidate actions that exceed the safe operation boundary of the device or do not conform to the basic physical laws of heat exchange are masked and blocked, that is, the selection probability of such actions is directly set to zero. Only control actions that conform to the mechanism constraints are retained as the set of output actions. From the action output level, the intelligent agent outputs invalid or even harmful control commands under extreme conditions, and solves the problem that pure data-driven models are prone to random output in scenarios where training data is not covered.
[0026] This invention also discloses a multi-objective reward function design rule for reinforcement learning agents. The reward function adopts a multi-objective weighted design and includes three core parts: a system operation energy consumption reward term, an indoor thermal comfort deviation penalty term, and a mechanism consistency penalty term. The specific calculation form of the reward function is as follows: ;in, for The instantaneous reward value obtained by the agent at any given moment. , , These are the weight coefficients for the three optimization objectives. for The total operating power of the central air conditioning system at all times. for The baseline operating power based on fixed threshold control under the same operating conditions at the same time. for Average predicted average thermal comfort vote value at any given time. This represents the total number of core devices within the system. for Time of the first The actual operating parameter values of the equipment for The time calculated based on the physical mechanism model is the first Theoretical operating parameter values of the equipment For the first The rated maximum value of the corresponding operating parameters for each device.
[0027] During training and operation, the above reward function not only guides the agent to pursue lower system energy consumption and better indoor thermal comfort, but also punishes situations where the actual operating parameters of the equipment deviate too much from the predicted values of the mechanism through the mechanism consistency penalty term, thereby guiding the agent to output control strategies that conform to the laws of physical operation and avoiding control behaviors that violate common sense physics.
[0028] This invention also discloses a dynamic adjustment rule for indoor thermal comfort constraints. The thermal comfort constraints use the predicted average voting value as the evaluation index, changing the traditional control mode that uses a fixed temperature threshold. The comfort zone threshold is dynamically adjusted according to the occupancy rate of each area. In low-load areas where the occupancy rate is lower than the preset threshold, the temperature control range is appropriately relaxed. In core areas where people stay for a long time, strict comfort zone control is maintained. While ensuring the thermal comfort experience of people when they are indoors, the energy-saving potential of non-core areas such as corridors, elevator halls, and non-office areas is fully explored to avoid unnecessary energy waste.
[0029] This invention also discloses a system-wide proactive rolling optimization mechanism. When the intelligent agent updates the state input step by step, it simultaneously incorporates hourly weather forecast data for the next 2 to 4 hours and data on changes in the indoor occupancy rate of people in each region based on historical population patterns. By using an embedded physical mechanism model to predict the trend of system cooling load changes, it makes small pre-adjustments to the operating parameters of each device in advance, thereby offsetting the control deviation caused by the large inertia and long lag characteristics of the central air conditioning system. This avoids the temperature fluctuations and energy waste caused by traditional feedback control that only starts adjusting after the load changes, and improves the stability and comfort level of the control process.
[0030] This invention also discloses the deployment and online update rules of the intelligent agent. The deep reinforcement learning intelligent agent adopts a deployment mode that combines offline pre-training with online fine-tuning. In the offline training stage, in addition to using historical project operation data, full-condition simulation data covering normal working conditions, extreme high temperature working conditions, extreme low temperature working conditions, and some equipment failure working conditions are generated through mechanism simulation models. After fusing the two types of data, the initial strategy of the intelligent agent is converged, ensuring that the intelligent agent has basic control capabilities under all working conditions before going online. After going online, the agent policy network is fine-tuned with a small learning rate using the actual operating data collected on the same day as a fixed time period. During the online fine-tuning process, the physical mechanism constraint module remains in effect, and the actions output by the policy network after fine-tuning are continuously verified by mechanism to avoid irregular drift of the control policy during online updates and ensure long-term stability.
[0031] This invention also discloses a smoothing processing rule for control commands. The dynamic control parameters of each device output by the reinforcement learning agent need to be processed by the command smoothing module before being sent to the building automation system for execution. The command smoothing module sets reasonable thresholds for the adjustment range of control parameters at adjacent times, respectively limiting the single adjustment range of equipment operating frequency, valve opening, and temperature setpoint, avoiding frequent start-stop and shutdown of equipment and drastic parameter fluctuations, reducing mechanical wear and electrical shock of equipment, ensuring equipment lifespan while achieving energy-saving control, reducing the negative impact of AI control on hardware equipment, and improving the engineering feasibility of the solution.
[0032] This invention also discloses self-verification and safety baseline rules for the control process. During the operation of the control method, the control effect is continuously self-verified, and the actual operating energy consumption of the system is compared with the baseline energy consumption calculated based on the mechanism model under the same operating conditions in real time. The dynamic energy saving rate is statistically analyzed. At the same time, the status of indoor thermal comfort parameters and equipment operating parameters is continuously monitored. When abnormal situations such as indoor thermal comfort parameters continuously exceeding the constraint range, equipment operating parameters touching the safety boundary, or large-scale interruption of data acquisition are detected, the safety baseline control mechanism is automatically triggered to switch the system to a preset, long-term verified safe operating mode.
[0033] This safe operation mode adopts the industry's mature fixed threshold PID control logic, which operates completely independently of the AI control module. It synchronously pushes abnormal alarm information to the operation and maintenance end, records the full-link operation data for 2 hours before and after the abnormality is triggered and stores it in the model iteration sample library. After the abnormality is investigated and the system status returns to the normal operating range, it automatically switches back to the AI energy-saving control mode. This mechanism avoids the control risks caused by AI module failure and data abnormality, and ensures that the central air conditioning system can stably meet the basic cooling and heating needs under any operating conditions.
[0034] Reference Figure 1 This diagram illustrates the overall business architecture of energy-saving control in a central air conditioning system. First, the system comprehensively collects multi-source data, including mechanistic parameters, operating status, and external environmental data from all devices in the central air conditioning system chain, and performs refined preprocessing. Then, it clarifies the overall control levels and corresponding adjustable boundaries for the chiller units, pumps, fans, and terminals within the system. Based on this, a reinforcement learning control framework that deeply integrates prior constraints related to physical mechanisms is constructed. Finally, the intelligent agent utilizes the continuously updated environmental state characteristics to deduce the optimal control parameters that minimize the overall energy consumption of the system hourly, while meeting indoor thermal comfort requirements. These parameters are then smoothly processed and sent to the devices for execution, forming a closed-loop, dynamic optimization energy-saving control system.
[0035] Reference Figure 2 This diagram details the data preparation process for the reinforcement learning model's front-end. The system collects real-time data on rated mechanistic parameters, including chiller efficiency curves and power consumption characteristics of pumps and fans, while simultaneously gathering dynamic status data such as equipment operating power, medium temperature and pressure, and valve opening. Regarding the external environment, indoor and outdoor temperature and humidity, regional occupancy rates, and weather forecasts are integrated concurrently. To ensure data quality, the time granularity of all collected data is strictly standardized. Subsequently, the system utilizes sliding window filtering to remove outliers and performs time alignment and standardization operations, transforming massive amounts of chaotic, multi-source, heterogeneous data into high-quality state features with a unified format, laying the data foundation for accurate decision-making by the intelligent agent.
[0036] Reference Figure 3 This diagram illustrates how deep reinforcement learning models are deeply coupled with physical mechanisms. The core model employs a deep deterministic policy gradient network architecture. Its state space not only includes real-time collected data but also innovatively incorporates theoretical operational efficiency and energy consumption benchmarks from the mechanism model, effectively compressing the ineffective exploration space. In the action generation stage, all candidate actions must undergo compliance verification by the physical mechanism model before output, with actions that exceed limits or violate heat exchange laws being masked and filtered. Simultaneously, the reward function adopts a multi-objective weighted design, using mechanism consistency deviation as one of the core penalty terms. This guides the reinforcement learning agent to learn and output safe control strategies that absolutely conform to physical operating laws in complex environments.
[0037] Reference Figure 4This diagram illustrates the control logic of the system in dealing with the large inertia and load fluctuations of the central air conditioning system. When continuously updating the agent's status input, the system proactively incorporates weather forecasts and predicted changes in occupancy rates for the next few hours. Utilizing a physical mechanism model, the system can predict future cooling load trends and adjust equipment parameters in advance to effectively offset the long lag effect of the air conditioning system. Regarding constraint management, the system uses predicted average voting values to assess thermal comfort and flexibly adjusts comfort zone thresholds based on dynamic occupancy rates, reducing temperature control requirements in low-load areas. Ultimately, the agent optimizes for minimizing overall system energy consumption, outputting dynamic control commands for each device hourly.
[0038] Reference Figure 5 This diagram illustrates the complete lifecycle management of the algorithm, from model training and deployment to security protection. In the early stages, the agent utilizes historical operational data and mechanistic simulation data for offline pre-training, achieving rapid convergence of the initial strategy. After deployment, the model uses real-time transmitted operational data for online fine-tuning at fixed intervals with small learning rates, ensuring continuous adaptation to environmental changes. To protect hardware, the control parameters output by the agent undergo instruction smoothing, strictly limiting the single adjustment range of device frequency and valve opening. During continuous system operation, the background performs real-time self-verification of performance. Once thermal comfort is severely exceeded or equipment operation touches safety boundaries, the system immediately triggers a safety protection mechanism, forcibly switching to a preset safety mode.
[0039] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for AI-based energy-saving control of building central air conditioning based on mechanism fusion reinforcement learning, characterized in that, Includes the following steps: S1. Real-time acquisition of rated mechanism parameters and operating status data of all equipment in the building's central air conditioning system, synchronous acquisition of indoor and outdoor temperature and humidity data, occupancy rate data of people in each area, and weather forecast data; Time alignment, outlier cleaning and standardization preprocessing of the acquired multi-source data. S2. Construct a deep reinforcement learning control framework that embeds the physical mechanism constraints of the entire central air conditioning system. Take the chiller unit's cooling efficiency characteristics, chilled water cooling water heat exchange mechanism, water pump and fan power consumption characteristics, and terminal thermal comfort balance mechanism as prior constraints, and integrate them into the state space construction, action space mapping, and reward function calculation of the deep reinforcement learning agent. S3. Divide the central air conditioning system into full-area control levels, covering the chiller group control layer on the cold source side, the chilled cooling water distribution system control layer, the fresh air air conditioning unit air system control layer, and the terminal fan coil control layer, and clarify the adjustable control variable range and equipment safety operation constraints corresponding to each level; S4. Based on the preprocessed real-time dynamic data, the state input of the reinforcement learning agent is updated continuously. With the requirement of meeting indoor thermal comfort as the constraint and the goal of minimizing the overall energy consumption of the system as the optimization objective, the reinforcement learning agent outputs the dynamic control parameters of each level of equipment hourly to replace the traditional fixed threshold control logic, thereby realizing the dynamic optimization and energy-saving control of the entire link of the central air conditioning system.
2. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning as described in claim 1, characterized in that, In step S1, the rated mechanism parameters of the entire chain of the building's central air conditioning system are collected in real time, including the cooling efficiency characteristic curves of the chiller unit under different load rates and different cooling water inlet temperatures, the flow rate and head characteristic curves and the flow rate and power consumption characteristic curves of the water pump, the air volume and air pressure characteristic curves and the air volume and power consumption characteristic curves of the fan, and the heat exchange efficiency characteristic parameters of the terminal heat exchange coils; the operating status data includes the real-time operating power, operating frequency, fluid medium inlet and outlet temperatures, inlet and outlet pressures, and regulating valve opening data of each device. The time granularity of all data collection is uniformly 1 minute, and outlier cleaning is completed by sliding window filtering.
3. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning as described in claim 1, characterized in that, The deep reinforcement learning agent constructed in step S2 adopts a deep deterministic policy gradient network architecture. In addition to the preprocessed real-time collected data, its state space also includes the theoretical operating efficiency values of each device, the theoretical minimum energy consumption benchmark value of the system under the current working condition, and the thermal comfort deviation value calculated in real time based on the physical mechanism model as core state variables, thereby compressing the agent's exploration space for meaningless and ineffective actions.
4. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning as described in claim 3, characterized in that, The action space of the deep reinforcement learning agent corresponds to the adjustable range of control variables at each level. Before all candidate actions are input into the control execution module, they must be verified for compliance by the embedded physical mechanism model. Candidate actions that exceed the safe operation boundary of the equipment or do not conform to the basic physical laws of heat exchange are masked out, and only control actions that conform to the mechanism constraints are retained as the set of output actions.
5. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning according to claim 1, characterized in that, The reward function in step S2 adopts a multi-objective weighted design, which includes three core parts: system operation energy consumption reward, indoor thermal comfort deviation penalty, and mechanism consistency penalty. The mechanism consistency penalty is calculated based on the deviation between the actual operating parameters of the equipment and the theoretical parameters output by the physical mechanism model. When the deviation exceeds a preset threshold, a penalty is imposed on the agent to guide the agent to output a control strategy that conforms to the physical operating law.
6. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning according to claim 1, characterized in that, In step S4, the indoor thermal comfort constraint uses the predicted average voting value as the evaluation index. The comfort zone threshold is dynamically adjusted according to the occupancy rate of each area. In low-load areas where the occupancy rate is lower than the preset threshold, the temperature control range is reduced.
7. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning according to claim 1, characterized in that, In step S4, when the status input is updated on a rolling basis, hourly weather forecast data for the next 2 to 4 hours and data on changes in the indoor occupancy rate of people in each region based on historical patterns are simultaneously incorporated. By using the embedded physical mechanism model to predict the trend of system cooling load changes, the operating parameters of each device are adjusted in advance to offset the control deviation caused by the large inertia and long lag characteristics of the central air conditioning system.
8. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning according to claim 7, characterized in that, The deep reinforcement learning agent adopts a deployment mode that combines offline pre-training with online fine-tuning. During the offline training phase, the agent's initial policy convergence is completed by integrating historical running data and full-condition simulation data generated by the mechanism simulation model. After going online, the agent is fine-tuned with a small learning rate using real-time collected running data at fixed time intervals. During the online fine-tuning process, the physical mechanism constraint module remains effective.
9. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning according to claim 1, characterized in that, The dynamic control parameters of each device output by the reinforcement learning agent need to be processed by the instruction smoothing module before being sent out for execution. The instruction smoothing module sets a threshold for the adjustment range of control parameters at adjacent times, limiting the single adjustment range of device operating frequency, valve opening, and temperature setpoint.
10. The AI-powered energy-saving control method for building central air conditioning based on mechanism fusion reinforcement learning according to claim 1, characterized in that, During operation, the control method continuously performs self-verification of control effect, compares the actual operating energy consumption of the system with the benchmark energy consumption calculated based on the mechanism model under the same working conditions in real time, and calculates the dynamic energy saving rate. When it is detected that the indoor thermal comfort parameters continuously exceed the constraint range or the equipment operating parameters touch the safety boundary, the backup control mechanism is automatically triggered to switch the system to the preset safe operation mode.
