A central air conditioner energy-saving optimization system based on load forecasting

Abstract

The present invention discloses an energy-saving optimization system for a central air conditioner based on load prediction, and relates to the technical field of air conditioning energy saving. The present invention introduces a VMDTransformerGRU hybrid deep learning model, and through multi-source data fusion, significantly improves the accuracy of cold load prediction, so that the prediction model can identify the load change trend in advance. This improvement effectively enhances the dynamic response capability of the system, and can perform pre-regulation before load changes to reduce energy waste. By modeling the chiller, cooling tower, and pump as independent intelligent agents, and adopting multi-agent deep reinforcement learning (MADRL), each intelligent agent can autonomously learn and optimize its own control strategy. The intelligent agents cooperate with each other, and through autonomous decision-making, the efficient operation of the overall system is achieved, and operational conflicts between devices are avoided. This method realizes dynamic coordination and joint optimization between various devices, reduces ineffective energy consumption, and significantly improves the overall energy-saving effect and operational stability of the system.

Description

Technical Field

[0001] The invention relates to the technical field of air conditioning energy saving, and in particular to an energy saving optimization system for a central air conditioner based on load prediction. Background Art

[0002] In large shopping malls and buildings, the energy consumption of central air-conditioning systems accounts for a very high proportion. In order to achieve energy conservation, researchers have generally adopted optimization methods based on load forecasting in recent years. Such methods predict future cooling loads and adjust the operating parameters of the air-conditioning system in advance, thereby reducing unnecessary energy consumption.

[0003] The traditional model predictive control (MPC) method is a commonly used method in the current energy-saving optimization of central air-conditioning. By predicting future loads, the operation strategy of air-conditioning equipment is planned in advance, and system parameters are adjusted in real time to achieve energy-saving goals. MPC uses system models and predicted load optimization control strategies to balance energy consumption and comfort based on the dynamic response characteristics of the equipment. Through rolling optimization, the prediction model and optimization results are continuously updated to adapt to changes in load. However, MPC is heavily dependent on the accurate modeling of the air-conditioning system. The central air-conditioning system has complex nonlinear characteristics and there is coupling between devices. It is difficult for traditional MPC to accurately capture these complex relationships, resulting in unsatisfactory control effects. As the scale of the system increases, the scale of the MPC optimization problem also increases sharply, the calculation time is long, and it is difficult to meet real-time requirements. However, in the face of sudden changes in load and unexpected disturbances in large shopping malls and buildings, such as increased customer activities during holidays, MPC relies on static model parameters and is difficult to respond quickly, resulting in a disconnect between the optimization results and actual operations. Traditional solutions mostly use the multi-agent deep reinforcement learning MADRL method for dynamic optimization of central air conditioning. Each agent controls different equipment respectively, and achieves global optimization through mutual learning and strategy adjustment. Although MADRL can adapt to nonlinear systems, it requires a large amount of training data and computing resources. The training process is complex and time-consuming. In a large-scale multivariable environment, the training time is long and it is difficult to deploy quickly. In a multi-agent environment, the strategy interaction between agents is complex, and training instability or poor convergence is prone to occur, resulting in inconsistent or unsatisfactory optimization strategies.

[0004] To address the adaptability problem of MPC, most current solutions introduce adaptive adjustment mechanisms to improve the adaptability of the model through online parameter adjustment and modeling error compensation; or simplify the control model to reduce computational complexity, but this method is only effective under specific conditions and cannot fundamentally solve the problems of model inaccuracy and poor real-time performance; and in terms of MADRL, most of them use experience replay, dynamic exploration and utilization balance and other technologies to improve training efficiency and stability, and use a centralized learning and distributed execution strategy framework to enhance the global perspective, but the improvement effect is still limited by the coordination complexity between intelligent agents and the bottleneck of computing resources, and cannot fully meet the needs of real-time applications. Therefore, there is an urgent need for a central air conditioner energy-saving optimization system based on load forecasting to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a central air conditioner energy-saving optimization system based on load forecasting to solve the problem that the system still has a lag in response to rapid load changes, is difficult to meet the control requirements in a high dynamic environment, and complex algorithms are difficult to execute in real time in large-scale buildings or shopping malls.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In one aspect, the present invention provides a central air conditioner energy-saving optimization system based on load prediction, which comprises:

[0009] The data collection and fusion module is responsible for collecting and integrating multi-source data, including indoor and outdoor temperature and humidity, real-time personnel flow, equipment status, and surrounding weather forecasts;

[0010] The load forecasting module uses a hybrid deep learning model to predict future cooling loads, and the prediction results serve as input to the intelligent control system;

[0011] Intelligent control system, which models chillers, cooling towers and pumps as independent agents, uses the multi-agent deep reinforcement learning (MADRL) framework to independently learn and optimize operation strategies. Each agent continuously adjusts its operation strategy based on load forecasts and equipment status.

[0012] Model predictive control MPC optimization module, which uses MPC to perform global optimization control on the energy-saving optimization system. MPC outputs a coordinated control strategy based on load forecast and current operation strategy, provides operation references and constraints for each intelligent agent, and feeds back the learning process of the intelligent agent to guide its strategy optimization direction;

[0013] The digital twin simulation module uses the digital twin of the air conditioning system to simulate the actual air conditioning system in real time. The digital twin provides real-time feedback on the operation of the air conditioning system and generates optimization suggestions.

[0014] The adaptive control module dynamically adjusts the operating parameters of the air-conditioning system based on the feedback and optimization suggestions of the digital twin.

[0015] On the other hand, the present invention provides a central air conditioner energy-saving optimization method based on load prediction, comprising:

[0016] Step S1, multi-source data fusion and load forecasting,

[0017] Collect and integrate indoor and outdoor temperature and humidity, real-time personnel flow, equipment status, and surrounding weather forecasts, and use the hybrid deep learning model VMDTransformerGRU (combination of variational mode decomposition, Transformer, and GRU) to predict future cooling loads.

[0018] Step S2: The agent independently learns the operation strategy.

[0019] According to the load forecast data in step S1, the chiller, cooling tower and pump are modeled as independent agents. As controllable devices in the system, each agent adopts the multi-agent deep reinforcement learning MADRL framework to autonomously learn the operation strategy based on the load forecast data and the current equipment status. The agent optimizes its control strategy in continuous attempts to achieve a balance between energy consumption and comfort.

[0020] Step S3, collaborative optimization control,

[0021] Combining the load forecasting results of step S1 and the agent learning strategy of step S2, the model predictive control MPC is introduced for global optimization;

[0022] Step S4, digital twin simulation and dynamic feedback,

[0023] Based on the prediction, learning and optimization process from steps S1 to S3, a digital twin of the air-conditioning system is established. This virtual model simulates the state changes of the actual air-conditioning system in real time, providing a high-precision simulation environment for strategy verification and adjustment.

[0024] The digital twin provides real-time feedback on the operation of the air-conditioning system and generates optimization suggestions to guide the control of the air-conditioning system. The feedback includes the energy consumption performance of the equipment and whether the temperature and humidity meet the standards.

[0025] Step S5, adaptive control of the air conditioning system,

[0026] According to the feedback and optimization suggestions of the digital twin in step S4, the air-conditioning system dynamically adjusts the operating parameters of the chiller, cooling tower and pump to perform continuous adaptive control.

[0027] Furthermore, the multi-source data fusion and load forecasting method in step S1 is:

[0028] Perform data collection and preprocessing. Data sources include indoor temperature. and humidity , outdoor temperature and humidity , real-time personnel flow , device status , including the operating status of chillers, cooling towers and pumps, weather forecast data , including future temperature, humidity, and wind speed;

[0029] All collected data are normalized and timestamped to form a multi-dimensional input sequence. ,in, Represents the fused multi-dimensional input sequence;

[0030] For multidimensional input sequences Perform variational mode decomposition to decompose the original signal into several intrinsic mode functions IMFs. ,in, Indicates eigenmode functions, representing characteristic modes, represents the number of decomposed modes, represents the residual noise term;

[0031] The decomposed intrinsic mode function Input into the Transformer model to extract the temporal dependency features. ,in, Represents the deep time series features extracted by Transformer, including load characteristics at different time scales;

[0032] The features extracted by Transformer Input into the GRU gated recurrent unit network for load forecasting. ,in, Indicates the future The cooling load value predicted at each moment.

[0033] Furthermore, in step S2, the agent independently learns the operation strategy:

[0034] Each device is modeled as an independent agent. Each agent learns load forecast data and device status and performs control optimization autonomously. The agent is defined as: Chiller Agent , Cooling Tower Agent and pump agents ;

[0035] definition Represents the state vector, which is composed of load forecast data, equipment status and environmental variables. The state of each agent is the data fusion result of step S1.

[0036] ,

[0037] ,

[0038] ,

[0039] in, Indicates the current time The load forecast value, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the pump inlet and outlet water temperature, Indicates the current operating power of the chiller, cooling tower and pump. Indicates indoor temperature and humidity. Indicates outdoor temperature and humidity. Indicates real-time personnel flow, Represents weather forecast data;

[0040] definition Indicates the actions taken by the agent, including starting and stopping the device and adjusting the speed, making decisions based on the current state.

[0041] ,

[0042] ,

[0043] ,in, Indicates the control parameters of the chiller, including compressor power and frequency adjustment, Represents the control parameters of the cooling tower, including fan speed and water flow control, Indicates the control parameters of the pump, including the speed and flow rate regulation of the pump.

[0044] Furthermore, in step S2, the agent independently learns the operation strategy mode and further includes:

[0045] definition represents the reward function, ,

[0046] in, represents the total energy consumption of the system,

[0047] Indicates the temperature and humidity comfort score, reflecting the deviation between the indoor environment and the set value. represents the operating cost of each device, represents the weight parameter;

[0048] The agent uses a deep reinforcement learning algorithm to optimize the strategy, and the strategy update method is:

[0049] ,in represents the optimal strategy, represents the discount factor; through mutual learning and collaborative control of multiple agents, each agent autonomously adjusts its operation strategy.

[0050] Furthermore, the global optimization method in step S3 is:

[0051] MPC adjusts system operation according to load forecast results and outputs coordinated control strategies as reference and constraints for the operation of the agent. The coordinated control strategies of MPC feed back into the learning process of the MADRL agent, and the agent adjusts its strategy optimization direction according to the guidance of MPC.

[0052] Furthermore, the collaborative optimization control method in step S3 is:

[0053] Construct the optimization objective function, ,in, represents the control input vector, which contains the control parameters of the chiller, cooling tower and pump,

[0054] Represents the total energy consumption of the system, including the chiller , Cooling tower and pump The power consumption,

[0055] Indicates comfort deviation, reflecting indoor temperature and humidity With the set value The deviation Represents the equipment operating costs, including state conversion losses of chillers, cooling towers and pumps, represents the weight parameter;

[0056] Define constraints,

[0057] ,

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] in, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the maximum temperature allowed by the device. Indicates the maximum power limitations of chillers, cooling towers and pumps.

[0063] Furthermore, the collaborative optimization control method in step S3 also includes:

[0064] Output coordinated control strategy, perform rolling optimization on future predicted load changes, and output optimal control strategy , including the control input of each device, ,in, Represents the control signal of the chiller, i.e. the compressor frequency, represents the control signal of the cooling tower, i.e. the fan speed, Indicates the control signal of the pump, i.e. the speed of the pump;

[0065] MPC output strategy As a reference and constraint condition for the MADRL agent, the agent fine-tunes its own control strategy based on the MPC strategy during the learning process. The agent strategy adjustment:

[0066] ,

[0067] in, represents the optimal strategy, represents an immediate reward, measuring energy efficiency and comfort, represents the guiding constraint of MPC, represents the discount factor;

[0068] Through the macro-control of MPC and the micro-adjustment of the intelligent body, the system achieves global and local coordinated optimization, and ultimately reaches the optimal state of balance between energy saving and comfort.

[0069] Furthermore, the digital twin simulation and dynamic feedback method in step S4 is:

[0070] Construct a digital twin model. The digital twin consists of multiple sub-models, including thermodynamic model, equipment performance model and environmental interaction model. The system state equation is: ,in, Represents the current state vector of the system, including the state parameters of each device, represents the input state vector, which is composed of the load forecast data obtained in step S1, represents the control input vector, provided by the MPC output control strategy of step S3, Represents environmental parameters, including indoor and outdoor temperature and humidity, and real-time data on personnel flow;

[0071] Conduct thermodynamic and equipment performance modeling. The thermodynamic model represents the modeling of the cooling load and heat transfer of the air-conditioning system, simulating the dynamic changes of indoor temperature and humidity. ,in, Indicates the current time The cooling load, represents the air mass flow rate, is the specific heat capacity of air, represents the rate of change of indoor temperature, Indicates additional heat dissipation;

[0072] The equipment performance model represents the performance modeling of each equipment, including chillers, cooling towers and pumps, and simulates the energy consumption and efficiency of the equipment under different conditions.

[0073] ,

[0074] ,

[0075] ,in, represents the energy consumption of chillers, cooling towers and pumps, It represents the efficiency coefficient of the equipment. Indicates the thermal power of each device, Indicates the energy efficiency ratio of the chiller.

[0076] Furthermore, the digital twin simulation and dynamic feedback method in step S4 also includes:

[0077] Input real-time data and predicted data into the digital twin, run dynamic simulation calculations, and output the system state feedback. The state feedback equation is:

[0078] ,in, Represents the feedback output of the system, including equipment energy consumption, temperature and humidity deviations. The feedback information is compared with the MPC strategy and actual system response in real time to generate optimization suggestions;

[0079] Based on the deviation between the simulation output and the actual system operation, the digital twin generates optimization suggestions to guide equipment regulation and optimization suggestions:

[0080] ,in, represents the adjusted control input, function Combine simulation feedback with system goals to optimize equipment regulation.

[0081] The beneficial effects of the present invention are:

[0082] This invention introduces the VMDTransformerGRU hybrid deep learning model, which significantly improves the accuracy of cooling load prediction by fusing multi-source data (indoor and outdoor temperature and humidity, personnel flow, equipment status and weather forecast), enabling the prediction model to identify load change trends in advance. This improvement effectively enhances the dynamic response capability of the system, enables pre-regulation before load changes, and reduces energy waste.

[0083] The present invention models the chiller, cooling tower and pump as independent agents and adopts multi-agent deep reinforcement learning (MADRL), so that each agent can autonomously learn and optimize its own control strategy. The agents work together to achieve efficient operation of the overall system through autonomous decision-making, avoiding operational conflicts between devices. This approach achieves dynamic coordination and joint optimization between devices, reduces ineffective energy consumption, and significantly improves the overall energy-saving effect and operational stability of the system.

[0084] The present invention introduces model predictive control (MPC) for global optimization. MPC calculates the optimal strategy for multiple future moments based on load forecast data and outputs a coordinated control strategy as a reference and constraint for the operation of the intelligent agent. The intelligent agent accepts the guidance of MPC during the local optimization process, so that the global macro-control is combined with the micro-adjustment of the intelligent agent, achieving global and local consistency optimization. Through the feedback mechanism of MPC, the intelligent agent can make autonomous adjustments under the guidance of the global goal, significantly improving the comprehensive performance of energy-saving efficiency and comfort.

[0085] The present invention, by establishing a digital twin, integrates thermodynamic models, equipment performance models and environmental interaction models to perform high-precision simulation of the actual system. The digital twin simulates the operating status of the air-conditioning system in real time, generates optimization suggestions and guides equipment control to ensure that the system can continuously adjust operating parameters. By comparing actual operation and simulation feedback, the digital twin can promptly detect deviations in operation and provide optimization adjustments, ensuring that the system is continuously optimized under various load and environmental conditions and always maintained in the optimal state.

[0086] In this invention, the optimization suggestions generated by the digital twin guide the adaptive control of the system, so that the operating parameters of the chiller, cooling tower and pump can be continuously adjusted according to real-time feedback. The system dynamically responds to load changes, achieves efficient operation while ensuring comfortable indoor temperature and humidity, reduces ineffective energy consumption, and achieves a dual improvement in energy saving and comfort. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0088] Figure 1 It is a structural schematic diagram of a central air conditioner energy-saving optimization system based on load prediction of the present invention;

[0089] Figure 2 It is a schematic flow chart of the energy-saving optimization method for a central air conditioner based on load forecasting of the present invention. DETAILED DESCRIPTION

[0090] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0091] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0092] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0093] Example 1, reference Figure 1 This embodiment provides a central air conditioner energy-saving optimization system based on load forecasting, including:

[0094] The data collection and fusion module is responsible for collecting and integrating multi-source data, including indoor and outdoor temperature and humidity, real-time personnel flow, equipment status, and surrounding weather forecasts;

[0095] The load forecasting module uses a hybrid deep learning model to predict future cooling loads, and the prediction results serve as input to the intelligent control system;

[0096] Intelligent control system, which models chillers, cooling towers and pumps as independent agents, uses the multi-agent deep reinforcement learning (MADRL) framework to independently learn and optimize operation strategies. Each agent continuously adjusts its operation strategy based on load forecasts and equipment status.

[0097] Model predictive control MPC optimization module, which uses MPC to perform global optimization control on the energy-saving optimization system. MPC outputs a coordinated control strategy based on load forecast and current operation strategy, provides operation references and constraints for each intelligent agent, and feeds back the learning process of the intelligent agent to guide its strategy optimization direction;

[0098] The digital twin simulation module uses the digital twin of the air conditioning system to simulate the actual air conditioning system in real time. The digital twin provides real-time feedback on the operation of the air conditioning system and generates optimization suggestions.

[0099] The adaptive control module dynamically adjusts the operating parameters of the air-conditioning system based on the feedback and optimization suggestions of the digital twin.

[0100] Example 2, reference Figure 2 This embodiment provides a central air conditioner energy-saving optimization method based on load forecasting, comprising the following steps:

[0101] Step S1, multi-source data fusion and load forecasting,

[0102] Collect and integrate indoor and outdoor temperature and humidity, real-time personnel flow, equipment status, and surrounding weather forecasts, and use the hybrid deep learning model VMDTransformerGRU (combination of variational mode decomposition, Transformer, and GRU) to predict future cooling loads.

[0103] The multi-source data fusion and load forecasting method in step S1 is:

[0104] Perform data collection and preprocessing. Data sources include indoor temperature. and humidity , outdoor temperature and humidity , real-time personnel flow , device status , including the operating status of chillers, cooling towers and pumps, weather forecast data , including future temperature, humidity, and wind speed;

[0105] All collected data are normalized and timestamped to form a multi-dimensional input sequence. ,in, Represents the fused multi-dimensional input sequence;

[0106] For multidimensional input sequences Perform variational mode decomposition to decompose the original signal into several intrinsic mode functions IMFs. ,in, Indicates eigenmode functions, representing characteristic modes, represents the number of decomposed modes, represents the residual noise term;

[0107] The decomposed intrinsic mode function Input into the Transformer model to extract the temporal dependency features. ,in, Represents the deep time series features extracted by Transformer, including load characteristics at different time scales;

[0108] The features extracted by Transformer Input into the GRU gated recurrent unit network for load forecasting. ,in, Indicates the future The cooling load value predicted at each moment;

[0109] Specifically, through the multi-source data fusion and hybrid deep learning model VMDTransformerGRU in step S1, the future cooling load is predicted. The indoor and outdoor temperature and humidity, personnel flow, equipment status and weather forecast data are integrated, and the variational mode decomposition is used to eliminate noise interference. The multi-scale time series features are extracted through the Transformer and GRU models. The deep learning architecture significantly improves the accuracy of load forecasting, enables the air-conditioning system to adjust its operation strategy in advance, reduces the response lag to sudden load changes, and improves the system's response speed and adaptability.

[0110] Step S2: The agent independently learns the operation strategy.

[0111] According to the load forecast data in step S1, the chiller, cooling tower and pump are modeled as independent agents. As controllable devices in the system, each agent adopts the multi-agent deep reinforcement learning MADRL framework to autonomously learn the operation strategy based on the load forecast data and the current equipment status. The agent optimizes its control strategy in continuous attempts to achieve a balance between energy consumption and comfort.

[0112] In step S2, the agent independently learns the operation strategy:

[0113] Each device is modeled as an independent agent. Each agent learns load forecast data and device status and performs control optimization autonomously. The agent is defined as: Chiller Agent , Cooling Tower Agent and pump agents ;

[0114] definition Represents the state vector, which is composed of load forecast data, equipment status and environmental variables. The state of each agent is the data fusion result of step S1.

[0115] ,

[0116] ,

[0117] ,

[0118] in, Indicates the current time The load forecast value, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the pump inlet and outlet water temperature, Indicates the current operating power of the chiller, cooling tower and pump. Indicates indoor temperature and humidity. Indicates outdoor temperature and humidity. Indicates real-time personnel flow, Represents weather forecast data;

[0119] definition Indicates the actions taken by the agent, including starting and stopping the device and adjusting the speed, making decisions based on the current state.

[0120] ,

[0121] ,

[0122] ,in, Indicates the control parameters of the chiller, including compressor power and frequency adjustment, Represents the control parameters of the cooling tower, including fan speed and water flow control, Indicates the control parameters of the pump, including the speed and flow rate regulation of the pump;

[0123] In step S2, the agent independently learns the operation strategy and also includes:

[0124] definition represents the reward function, ,

[0125] in, represents the total energy consumption of the system,

[0126] Indicates the temperature and humidity comfort score, reflecting the deviation between the indoor environment and the set value. represents the operating cost of each device, represents the weight parameter;

[0127] The agent uses a deep reinforcement learning algorithm to optimize the strategy, and the strategy update method is:

[0128] ,in represents the optimal strategy, represents the discount factor; through mutual learning and collaborative control of multiple agents, each agent autonomously adjusts its operation strategy;

[0129] Specifically, in step S2, the equipment is modeled as an independent intelligent agent, and the multi-agent deep reinforcement learning (MADRL) framework is adopted to autonomously learn load forecast data and current equipment status and optimize its own operation strategy. Through independent learning and collaborative control, the intelligent agent can achieve a dynamic balance between energy consumption and comfort under different operating conditions, avoiding the inefficiency caused by the independent operation of each device in the traditional control method, and significantly improving the coordination and energy-saving effect of the overall system.

[0130] Step S3, collaborative optimization control,

[0131] Combining the load forecasting results of step S1 and the agent learning strategy of step S2, the model predictive control MPC is introduced for global optimization;

[0132] The global optimization method in step S3 is:

[0133] MPC adjusts system operation according to load forecast results and outputs coordinated control strategies as reference and constraints for the operation of the agent. The coordinated control strategies of MPC feed back into the learning process of the MADRL agent, and the agent adjusts its strategy optimization direction according to the guidance of MPC.

[0134] The collaborative optimization control method in step S3 is:

[0135] Construct the optimization objective function, ,in, represents the control input vector, which contains the control parameters of the chiller, cooling tower and pump,

[0136] Represents the total energy consumption of the system, including the chiller , Cooling tower and pump The power consumption,

[0137] Indicates comfort deviation, reflecting indoor temperature and humidity With the set value The deviation Represents the equipment operating costs, including state conversion losses of chillers, cooling towers and pumps, represents the weight parameter;

[0138] Define constraints,

[0139] ,

[0140] ,

[0141] ,

[0142] ,

[0143] ,

[0144] in, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the maximum temperature allowed by the device. Indicates the maximum power limits of chillers, cooling towers, and pumps;

[0145] The collaborative optimization control method in step S3 also includes:

[0146] Output coordinated control strategy, perform rolling optimization on future predicted load changes, and output optimal control strategy , including the control input of each device, ,in, Represents the control signal of the chiller, i.e. the compressor frequency, represents the control signal of the cooling tower, i.e. the fan speed, Indicates the control signal of the pump, i.e. the speed of the pump;

[0147] MPC output strategy As a reference and constraint condition for the MADRL agent, the agent fine-tunes its own control strategy based on the MPC strategy during the learning process. The agent strategy adjustment:

[0148] ,

[0149] in, represents the optimal strategy, represents an immediate reward, measuring energy efficiency and comfort, represents the guiding constraint of MPC, represents the discount factor;

[0150] Through the macro-control of MPC and the micro-adjustment of the intelligent body, the system realizes the coordinated optimization of the global and local aspects, and finally reaches the optimal state of balance between energy saving and comfort.

[0151] Specifically, step S3 globally optimizes future loads by introducing model predictive control (MPC) and outputs a coordinated control strategy. MPC continuously optimizes the control parameters of chillers, cooling towers and pumps based on load forecast data and equipment status to achieve energy efficiency improvement at the macro level. The macro-control strategy serves as a guide and constraint for the learning of the intelligent agent, so that the micro-adjustments of each intelligent agent can be carried out within the framework of global optimization. Through the feedback of MPC, the intelligent agent can not only achieve local optimal control, but also adjust the optimization direction under the guidance of the global goal, further improving the energy-saving effect and operational stability of the overall system.

[0152] Step S4, digital twin simulation and dynamic feedback,

[0153] Based on the prediction, learning and optimization process from steps S1 to S3, a digital twin of the air-conditioning system is established. This virtual model simulates the state changes of the actual air-conditioning system in real time, providing a high-precision simulation environment for strategy verification and adjustment.

[0154] The digital twin provides real-time feedback on the operation of the air-conditioning system and generates optimization suggestions to guide the control of the air-conditioning system. The feedback includes the energy consumption performance of the equipment and whether the temperature and humidity meet the standards.

[0155] The digital twin simulation and dynamic feedback method in step S4 is:

[0156] Construct a digital twin model. The digital twin consists of multiple sub-models, including thermodynamic model, equipment performance model and environmental interaction model. The system state equation is: ,in, Represents the current state vector of the system, including the state parameters of each device, represents the input state vector, which is composed of the load forecast data obtained in step S1, represents the control input vector, provided by the MPC output control strategy of step S3, Represents environmental parameters, including indoor and outdoor temperature and humidity, and real-time data on personnel flow;

[0157] Conduct thermodynamic and equipment performance modeling. The thermodynamic model represents the modeling of the cooling load and heat transfer of the air-conditioning system, simulating the dynamic changes of indoor temperature and humidity. ,in, Indicates the current time The cooling load, represents the air mass flow rate, is the specific heat capacity of air, represents the rate of change of indoor temperature, Indicates additional heat dissipation;

[0158] The equipment performance model represents the performance modeling of each equipment, including chillers, cooling towers and pumps, and simulates the energy consumption and efficiency of the equipment under different conditions.

[0159] ,

[0160] ,

[0161] ,in, represents the energy consumption of chillers, cooling towers and pumps, It represents the efficiency coefficient of the equipment. Indicates the thermal power of each device, Indicates the energy efficiency ratio of the chiller;

[0162] The digital twin simulation and dynamic feedback method in step S4 also includes:

[0163] Input real-time data and predicted data into the digital twin, run dynamic simulation calculations, and output the system state feedback. The state feedback equation is:

[0164] ,in, Represents the feedback output of the system, including equipment energy consumption, temperature and humidity deviations. The feedback information is compared with the MPC strategy and actual system response in real time to generate optimization suggestions;

[0165] Based on the deviation between the simulation output and the actual system operation, the digital twin generates optimization suggestions to guide equipment regulation and optimization suggestions:

[0166] ,in, represents the adjusted control input, function Combine simulation feedback and system goals to optimize equipment control;

[0167] Specifically, in step S4, a digital twin model of the air-conditioning system is established. Through the thermodynamic model, equipment performance model and environmental interaction model, the state changes of the air-conditioning system are simulated in real time. The digital twin can compare the difference between the prediction and the actual operation, generate real-time optimization suggestions, and guide the adaptive regulation of the system.

[0168] Step S5, adaptive control of the air conditioning system,

[0169] According to the feedback and optimization suggestions of the digital twin in step S4, the air-conditioning system dynamically adjusts the operating parameters of the chiller, cooling tower and pump to perform continuous adaptive control.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A central air conditioner energy-saving optimization system based on load forecasting, characterized in that: include, The data collection and fusion module is responsible for collecting and integrating multi-source data, including indoor and outdoor temperature and humidity, real-time personnel flow, equipment status, and surrounding weather forecasts; The load forecasting module uses a hybrid deep learning model to predict future cooling loads, and the prediction results serve as input to the intelligent control system; Intelligent control system, which models chillers, cooling towers and pumps as independent agents, uses the multi-agent deep reinforcement learning (MADRL) framework to independently learn and optimize operation strategies. Each agent continuously adjusts its operation strategy based on load forecasts and equipment status. Model predictive control MPC optimization module, which uses MPC to perform global optimization control on the energy-saving optimization system. MPC outputs a coordinated control strategy based on load forecast and current operation strategy, provides operation references and constraints for each intelligent agent, and feeds back the learning process of the intelligent agent to guide its strategy optimization direction; The digital twin simulation module uses the digital twin of the air conditioning system to simulate the actual air conditioning system in real time. The digital twin provides real-time feedback on the operation of the air conditioning system and generates optimization suggestions. Adaptive control module, dynamically adjusts the operating parameters of the air conditioning system based on the feedback and optimization suggestions of the digital twin; The agent learns and runs independently, and the learning and running strategy includes: definition represents the reward function, , in, represents the total energy consumption of the system, Indicates the temperature and humidity comfort score, reflecting the deviation between the indoor environment and the set value. Indicates the operating cost of each device, represents the weight parameter; The agent uses a deep reinforcement learning algorithm to optimize the strategy, and the strategy update method is: ,in represents the optimal strategy, represents the discount factor, represents the state vector, represents the action taken by the agent; Collaborative optimization control is carried out in the model predictive control MPC optimization module. The collaborative optimization control method is: Construct the optimization objective function, ,in, represents the control input vector, which contains the control parameters of the chiller, cooling tower and pump, Represents the total energy consumption of the system, including the chiller , Cooling tower and pump The power consumption, Indicates comfort deviation, reflecting indoor temperature and humidity With the set value The deviation Represents the equipment operating costs, including state conversion losses of chillers, cooling towers and pumps, represents the weight parameter; Define constraints, ， ， ， ， ， in, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the maximum temperature allowed by the device. Indicates the maximum power limitations of chillers, cooling towers and pumps.

2. A central air conditioner energy-saving optimization method based on load forecasting, characterized in that: Based on the central air conditioner energy-saving optimization system based on load prediction described in claim 1, The following steps are included: Step S1, multi-source data fusion and load forecasting, Collect and integrate indoor and outdoor temperature and humidity, real-time personnel flow, equipment status and surrounding weather forecast, and use the hybrid deep learning model VMDTransformerGRU to predict future cooling load. Step S2: The agent independently learns the operation strategy. According to the load forecast data in step S1, the chiller, cooling tower and pump are modeled as independent agents as controllable devices in the system. Each agent adopts the multi-agent deep reinforcement learning MADRL framework to autonomously learn the operation strategy based on the load forecast data and the current equipment status; Step S3, collaborative optimization control, Combining the load forecasting results of step S1 and the agent learning strategy of step S2, the model predictive control MPC is introduced for global optimization; Step S4, digital twin simulation and dynamic feedback, Based on the prediction, learning and optimization process from step S1 to S3, a digital twin of the air-conditioning system is established to simulate the state changes of the actual air-conditioning system in real time. The digital twin provides real-time feedback on the operation of the air-conditioning system and generates optimization suggestions to guide the control of the air-conditioning system. The feedback includes the energy consumption performance of the equipment and whether the temperature and humidity meet the standards. Step S5, adaptive control of the air conditioning system, According to the feedback and optimization suggestions of the digital twin in step S4, the air-conditioning system dynamically adjusts the operating parameters of the chiller, cooling tower and pump to perform continuous adaptive control.

3. The energy-saving optimization method for central air conditioners based on load forecasting according to claim 2 is characterized in that: The multi-source data fusion and load forecasting method in step S1 is: Perform data collection and preprocessing. Data sources include indoor temperature. and humidity , outdoor temperature and humidity , real-time personnel flow , device status , including the operating status of chillers, cooling towers and pumps, weather forecast data , including future temperature, humidity, and wind speed; All collected data are normalized and timestamped to form a multi-dimensional input sequence. ,in, Represents the fused multi-dimensional input sequence; For multidimensional input sequences Perform variational mode decomposition to decompose the original signal into several intrinsic mode functions IMFs. ,in, Indicates eigenmode functions, representing characteristic modes, represents the number of decomposed modes, represents the residual noise term; The decomposed intrinsic mode function Input into the Transformer model to extract the temporal dependency features. ,in, Represents the deep time series features extracted by Transformer, including load characteristics at different time scales; The features extracted by Transformer Input into the GRU gated recurrent unit network for load forecasting. ,in, Indicates the future The cooling load value predicted at each moment.

4. The energy-saving optimization method for central air conditioners based on load forecasting according to claim 3 is characterized in that: In step S2, the agent independently learns the operation strategy: Each device is modeled as an independent agent. Each agent learns load forecast data and device status and performs control optimization autonomously. The agent is defined as: Chiller Agent , Cooling Tower Agent and pump agents ; definition Represents the state vector, which is composed of load forecast data, equipment status and environmental variables. The state of each agent is the data fusion result of step S1. , , , in, Indicates the current time The load forecast value, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the pump inlet and outlet water temperature, Indicates the current operating power of the chiller, cooling tower and pump. Indicates indoor temperature and humidity. Indicates outdoor temperature and humidity. Indicates real-time personnel flow, Represents weather forecast data; definition Indicates the actions taken by the agent, including starting and stopping the device and adjusting the speed, making decisions based on the current state. ， ， ,in, Indicates the control parameters of the chiller, including compressor power and frequency adjustment, Indicates the control parameters of the cooling tower, including fan speed and water flow control, Indicates the control parameters of the pump, including the speed and flow rate regulation of the pump.

5. The energy-saving optimization method for central air conditioners based on load forecasting according to claim 4 is characterized in that: In step S2, the agent independently learns the operation strategy and also includes: definition represents the reward function, , in, represents the total energy consumption of the system, Indicates the temperature and humidity comfort score, reflecting the deviation between the indoor environment and the set value. Indicates the operating cost of each device, represents the weight parameter; The agent uses a deep reinforcement learning algorithm to optimize the strategy, and the strategy update method is: ,in represents the optimal strategy, Represents the discount factor.

6. A method for optimizing energy saving of a central air conditioner based on load forecasting according to claim 5, characterized in that: The global optimization method in step S3 is: MPC adjusts system operation according to load forecast results and outputs coordinated control strategies as reference and constraints for the operation of the agent. The coordinated control strategies of MPC feed back into the learning process of the MADRL agent, and the agent adjusts its strategy optimization direction according to the guidance of MPC.

7. A central air conditioner energy saving optimization method based on load forecasting according to claim 6, characterized in that: The collaborative optimization control method in step S3 is: Construct the optimization objective function, ,in, represents the control input vector, which contains the control parameters of the chiller, cooling tower and pump, Represents the total energy consumption of the system, including the chiller , Cooling tower and pump The power consumption, Indicates comfort deviation, reflecting indoor temperature and humidity With the set value The deviation Represents the equipment operating costs, including state conversion losses of chillers, cooling towers and pumps, represents the weight parameter; Define constraints, ， ， ， ， ， in, Indicates the inlet and outlet water temperature of the chiller. Indicates the inlet and outlet water temperature of the cooling tower. Indicates the maximum temperature allowed by the device. Indicates the maximum power limitations of chillers, cooling towers and pumps.

8. The energy-saving optimization method for central air conditioners based on load forecasting according to claim 7 is characterized in that: The collaborative optimization control method in step S3 also includes: Output coordinated control strategy, perform rolling optimization on future predicted load changes, and output optimal control strategy , including the control input of each device, ,in, Represents the control signal of the chiller, i.e. the compressor frequency, represents the control signal of the cooling tower, i.e. the fan speed, Indicates the control signal of the pump, that is, the speed of the pump; MPC output strategy As a reference and constraint condition for the MADRL agent, the agent fine-tunes its own control strategy based on the MPC strategy during the learning process. The agent strategy adjustment: ， in, represents the optimal strategy, represents an immediate reward, measuring energy efficiency and comfort, represents the guiding constraint of MPC, Represents the discount factor.

9. The energy-saving optimization method for a central air conditioner based on load forecasting according to claim 8, characterized in that: The digital twin simulation and dynamic feedback method in step S4 is: Construct a digital twin model. The digital twin consists of multiple sub-models, including thermodynamic model, equipment performance model and environmental interaction model. The system state equation is: ,in, Represents the current state vector of the system, including the state parameters of each device, represents the input state vector, which is composed of the load forecast data obtained in step S1, represents the control input vector, provided by the MPC output control strategy of step S3, Represents environmental parameters, including indoor and outdoor temperature and humidity, and real-time data on personnel flow; Conduct thermodynamic and equipment performance modeling. The thermodynamic model represents the modeling of the cooling load and heat transfer of the air-conditioning system, simulating the dynamic changes of indoor temperature and humidity. ,in, Indicates the current time The cooling load, represents the air mass flow rate, is the specific heat capacity of air, represents the rate of change of indoor temperature, Indicates additional heat dissipation; The equipment performance model represents the performance modeling of each equipment, including chillers, cooling towers and pumps, and simulates the energy consumption and efficiency of the equipment under different conditions. ， ， ,in, represents the energy consumption of chillers, cooling towers and pumps, It represents the efficiency coefficient of the equipment. Indicates the thermal power of each device, Indicates the energy efficiency ratio of the chiller.

10. The energy-saving optimization method for central air conditioners based on load forecasting according to claim 9, characterized in that: The digital twin simulation and dynamic feedback method in step S4 also includes: Input real-time data and predicted data into the digital twin, run dynamic simulation calculations, and output the system state feedback. The state feedback equation is: ,in, Represents the feedback output of the system, including equipment energy consumption, temperature and humidity deviations. The feedback information is compared with the MPC strategy and actual system response in real time to generate optimization suggestions; Based on the deviation between the simulation output and the actual system operation, the digital twin generates optimization suggestions to guide equipment regulation and optimization suggestions: ,in, Represents the adjusted control input.

Images