An AI-based central air conditioning energy consumption optimization system

Abstract

This invention relates to the field of central air conditioning control system technology, specifically to an AI-based central air conditioning energy consumption optimization system. The system includes a multimodal data acquisition module, an edge fusion processing module, an AI prediction and decision-making module, and an execution feedback module. The multimodal data acquisition module collects environmental physical data, personnel activity data, and space usage status data in real time. The edge fusion processing module cleans, aligns, and extracts features from the collected multimodal data, outputting a regional-level predicted load feature vector. The AI ​​prediction and decision-making module runs an AI algorithm model based on the predicted load feature vector to generate predicted load data and an optimized control strategy set for the central air conditioning system. The execution feedback module converts the optimized control strategy set into control commands and sends them to the actuators of the central air conditioning system. This system has predictive capabilities, enabling precise control of central air conditioning in building zones and reducing central air conditioning energy consumption.

Description

Technical Field

[0001] This invention relates to the field of central air conditioning control system technology, and in particular to an AI-based central air conditioning energy consumption optimization system. Background Technology

[0002] Central air conditioning systems are among the most energy-consuming devices in modern buildings. As buildings become increasingly larger and their functions more complex, optimizing the operation of central air conditioning systems has become a research hotspot in the field of building energy conservation.

[0003] Currently, the control of central air conditioning systems mainly relies on building automation systems (BAS). Traditional control strategies are usually based on segmented control according to a schedule. The system sets the operating mode uniformly according to a preset schedule (such as working hours and non-working hours). This approach cannot adapt to changes in the actual activity status of people, often resulting in the contradiction of continuous cooling in unoccupied areas and insufficient comfort in occupied areas, leading to serious energy waste.

[0004] Therefore, how to construct a central air conditioning energy consumption optimization system with predictive capabilities and the ability to finely control the central air conditioning in building zones in order to reduce central air conditioning energy consumption is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based central air conditioning energy consumption optimization system with predictive capabilities, which can achieve fine-grained control of central air conditioning in building zones and reduce central air conditioning energy consumption.

[0006] To achieve the above objectives, the present invention provides an AI-based central air conditioning energy consumption optimization system, comprising a multimodal data acquisition module, an edge fusion processing module, an AI prediction and decision module, and an execution feedback module, wherein the multimodal data acquisition module, the edge fusion processing module, the AI ​​prediction and decision module, and the execution feedback module are connected in sequence. The multimodal data acquisition module is used to collect environmental physical data, personnel activity data, and space usage status data in real time. The edge fusion processing module is used to clean, align, and extract features from the collected multimodal data, and output a region-level predicted load feature vector. The AI ​​prediction and decision-making module is used to run an AI algorithm model based on the predicted load feature vector to generate predicted load data and an optimized control strategy set for the central air conditioning system. The execution feedback module is used to convert the optimized control strategy set into control commands and send them to the actuators of the central air conditioning system, and to collect actual operating data and feed it back to the AI ​​prediction and decision module for model optimization.

[0007] The AI-based central air conditioning energy consumption optimization system also includes a zoning control module, which is connected to the AI ​​prediction and decision-making module and the execution feedback module, respectively. The zoning control module is used to automatically implement standby mode control for unmanned or vacant areas based on the regional personnel density index and space usage status data in the predicted load feature vector, and to implement pre-cooling or pre-ventilation control for high personnel density areas.

[0008] The multimodal data acquisition module includes an environmental physical acquisition unit, a personnel activity acquisition unit, and a space usage status acquisition unit, which are respectively connected to the edge fusion processing module. The environmental physical acquisition unit is used to access outdoor meteorological data through an API interface and to collect indoor temperature, humidity and CO2 concentration data through wireless sensors deployed indoors. The personnel activity collection unit is used to collect and generate regional personnel heat map data through Wi-Fi probes, workstation occupancy sensors or edge computing cameras deployed in various areas; The space usage status acquisition unit is used to connect to enterprise office collaboration systems, access control systems, or workstation reservation systems via API interfaces to obtain space usage schedule data.

[0009] The edge fusion processing module includes a data cleaning unit, a feature extraction unit, and a privacy protection unit, which are connected in sequence. The data cleaning unit is used to align timestamps on multimodal data with different sampling frequencies and remove outliers. The feature extraction unit is used to quantify the RSSI signal of the Wi-Fi probe into a people density index, and combine the meeting schedule and people density data to generate a predicted load feature vector for the next 15-30 minutes. The privacy protection unit is used to anonymize personnel data at the edge, outputting only regional density level data.

[0010] The AI ​​prediction and decision-making module includes a load prediction unit and an optimization decision-making unit, which are connected to the load prediction unit. The load forecasting unit is used to use a time-series forecasting model, taking historical indoor environmental data, future weather forecast data and future population density forecast data as input, and output hourly cold and heat load forecasts and predicted CO2 concentration change curves for each zone within the next 30-60 minutes. The optimization decision unit is used to generate an optimized control strategy set based on the prediction data output by the load prediction unit, including the chilled water outlet temperature setpoint, chiller start-up and shutdown plan, water pump frequency, fresh air valve opening degree, and zone standby strategy.

[0011] The load forecasting unit utilizes an LSTM network or a Transformer network as its time-series forecasting model.

[0012] The execution feedback module includes a protocol conversion unit, an instruction issuing unit, and a feedback acquisition unit; the protocol conversion unit, the instruction issuing unit, and the feedback acquisition unit are connected in sequence. The protocol conversion unit is used to convert the optimized control strategy set into control commands that can be recognized by the actuators of the central air conditioning system; The instruction issuing unit is used to issue control instructions to the building automation system or directly control the frequency converter, electric valve and chiller controller. The feedback acquisition unit is used to collect actual energy consumption data and environmental change data in real time, and compare them with the predicted values ​​output by the load prediction unit. The comparison results are then used as feedback data to be input into the AI ​​prediction decision module for online model learning.

[0013] This invention discloses an AI-based central air conditioning energy consumption optimization system. Through a multimodal data acquisition module, it acquires forward-looking data such as meeting schedules and personnel density in real time. Combined with a time-series prediction model in the AI ​​prediction and decision-making module, it can accurately predict changes in cooling and heating loads and CO2 concentrations within the next 30 to 60 minutes. Based on this, the system can pre-cool and pre-ventilate, eliminating temperature fluctuations and energy waste caused by lag in traditional control methods, significantly improving the stability of indoor environmental comfort. This invention integrates three types of multimodal data sources: environmental physical data, personnel activity data, and space usage status data, constructing a comprehensive perception capability for personnel activity and space usage. Compared with existing technologies that rely solely on temperature and humidity sensors, this invention can accurately reflect the dynamic load demand within the building, providing a more accurate data foundation for optimization decisions. This system possesses predictive capabilities, enabling fine-grained control of central air conditioning in building zones, reducing central air conditioning energy consumption. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0016] 1-Multimodal data acquisition module, 2-Edge fusion processing module, 3-AI prediction and decision-making module, 4-Execution feedback module, 5-Zone control module, 11-Environmental physical acquisition unit, 12-Personnel activity acquisition unit, 13-Space usage status acquisition unit, 21-Data cleaning unit, 22-Feature extraction unit, 23-Privacy protection unit, 31-Load prediction unit, 32-Optimization decision-making unit, 41-Protocol conversion unit, 42-Command issuance unit, 43-Feedback acquisition unit. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0018] Please see Figure 1 ,in, Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0019] This invention provides an AI-based central air conditioning energy consumption optimization system, comprising a multimodal data acquisition module 1, an edge fusion processing module 2, an AI prediction and decision-making module 3, an execution feedback module 4, and a zoning control module 5. The multimodal data acquisition module 1 includes an environmental physical acquisition unit 11, a personnel activity acquisition unit 12, and a space usage status acquisition unit 13. The edge fusion processing module 2 includes a data cleaning unit 21, a feature extraction unit 22, and a privacy protection unit 23. The AI ​​prediction and decision-making module 3 includes a load prediction unit 31 and an optimization decision-making unit 32. The execution feedback module 4 includes a protocol conversion unit 41, an instruction issuance unit 42, and a feedback acquisition unit 43. Through the aforementioned scheme, the system possesses predictive capabilities, enabling precise control of central air conditioning in building zones and reducing central air conditioning energy consumption.

[0020] In this specific embodiment, the multimodal data acquisition module 1, the edge fusion processing module 2, the AI ​​prediction and decision-making module 3, and the execution feedback module 4 are connected in sequence; The multimodal data acquisition module 1 is used to collect environmental physical data, personnel activity data, and space usage status data in real time; The edge fusion processing module 2 is used to clean, align and extract features from the collected multimodal data, and output a region-level predicted load feature vector. The AI ​​prediction and decision-making module 3 is used to run an AI algorithm model based on the predicted load feature vector to generate predicted load data and an optimized control strategy set for the central air conditioning system. The execution feedback module 4 is used to convert the optimized control strategy set into control commands and send them to the actuators of the central air conditioning system, and to collect actual operating data and feed it back to the AI ​​prediction and decision module 3 for model optimization.

[0021] In this embodiment, the multimodal data acquisition module 1 is configured to collect three types of data sources in real time: First, it periodically sends requests to the meteorological service platform via API to obtain hourly outdoor weather forecast data for the next 24 hours, including temperature, humidity, light intensity, and wind speed. At the same time, it collects indoor temperature, humidity, and CO2 concentration data at a sampling frequency of once per minute using LoRa or NB-IoT wireless temperature and humidity sensors deployed in return air ducts and typical functional areas on each floor. Second, it continuously detects detection request frames sent by mobile devices within the coverage area through Wi-Fi probe devices deployed in the weak current rooms and public areas on each floor, extracts MAC address and signal strength information, and generates a raw dataset containing the number of devices in each area every 5 minutes after deduplication algorithm. It also continuously detects the workstation occupancy status through workstation occupancy sensors and uploads the detection results in real time. Finally, it establishes a connection with the enterprise office collaboration system through RESTful API to periodically obtain meeting schedule data for the next 7 days, including meeting start time, end time, meeting room number, and estimated number of participants. At the same time, it connects with the access control system or workstation reservation system to obtain the area reservation status. All collected data is packaged in a unified data format and transmitted to the edge fusion processing module 2. The edge fusion processing module 2 is deployed in the edge computing gateway of each floor or area. It is configured to first receive multiple data streams from the multimodal data acquisition module 1, and align the data of different sampling frequencies with timestamps according to a unified time sequence alignment strategy. It uses a median filtering algorithm to smooth the instantaneous jump values ​​in the Wi-Fi probe data caused by signal interference, and uses a linear interpolation method to fill in the missing values ​​in the sensor network. Then, it runs a feature extraction algorithm to normalize the cleaned Wi-Fi probe device counts of each area into a personnel density index in the range of 0 to 100. It uses the number of participants in the meeting schedule data as the weight coefficient of the predicted load feature, and combines the current personnel density index with the meeting start event in the next 30 minutes to generate a predicted load feature vector containing the predicted time window, area identification, predicted personnel density change trend and expected CO2 load increment. Finally, the privacy protection unit 23 processes all data involving personal identification, hashes and anonymizes the MAC address list at the edge, discards the original data related to individual identification, and only uploads the density level and feature vector after area-level aggregation to the AI ​​prediction decision module 3 through an encrypted channel.The AI ​​prediction and decision-making module 3 is deployed on a cloud server or a local high-performance server. It is configured to first receive the predicted load feature vectors for each region uploaded by the edge fusion processing module 2, and then concatenate them with the environmental and energy consumption data stored in the historical database to construct a complete model input feature matrix. Next, the load prediction unit 31 loads a pre-trained time-series prediction model, using the indoor temperature and humidity and CO2 concentration sequences of the past hour, the outdoor weather forecast sequence for the next 30 minutes, and the predicted personnel density sequence for the next 30 minutes as model inputs. Through forward calculation, the model outputs the granular data for each zone every 15 minutes for the next 30 to 60 minutes. The system generates hourly cooling load forecasts, heating load forecasts, and CO2 concentration forecast curves. Then, the decision-making unit 32 is optimized by loading model predictive control or reinforcement learning algorithms. The output of the load forecasting unit 31 is used as a constraint, and the optimization objectives are to maximize the overall system energy efficiency ratio and minimize energy consumption. Under the premise of meeting indoor comfort constraints, the system generates an optimized control strategy set by solving the optimization problem. This set includes the setpoint of chilled water outlet temperature, chiller start-stop combination scheme, chilled water pump frequency, cooling water pump frequency, air conditioning unit fan frequency, fresh air valve opening ratio, and zonal standby strategy. The strategy set is then pushed to the execution feedback module 4 in a structured data format.

[0022] The execution feedback module 4 is configured to first convert the optimized control strategy set issued by the AI ​​prediction and decision module 3 into a control command format recognizable by the central air conditioning actuator through the protocol conversion unit 41. Then, the command issuing unit 42 sends the command to the DDC controller of the building automation system or directly to the frequency converter, electric regulating valve, and chiller group control controller according to the BACnet, Modbus TCP / IP, or OPC UA industrial protocol. At the same time, the feedback acquisition unit 43 reads the actual data of smart meters, heating and cooling meters, and sensors in real time, including the real-time power of each chiller, water pump operating frequency, chilled water supply and return water temperature, actual temperature and humidity of each area, and CO2 concentration. These actual operating data are compared with the predicted values ​​output by the load prediction unit 31 item by item to calculate the prediction error index. Finally, the prediction error and actual operating data are encapsulated into a feedback data packet and sent back to the AI ​​prediction and decision module 3 through the API interface to trigger the online learning process of the model. The model performs incremental training periodically or full retraining once a week based on the feedback data to achieve continuous self-optimization of the control strategy.

[0023] Secondly, the partition control module 5 is connected to the AI ​​prediction and decision-making module 3 and the execution feedback module 4 respectively; The zoning control module 5 is used to automatically execute standby mode control for unmanned or vacant areas based on the regional personnel density index and space usage status data in the predicted load feature vector, and to execute pre-cooling or pre-ventilation control for high personnel density areas.

[0024] In this embodiment, the zoning control module 5 is configured to first obtain real-time personnel density index and space usage status data for each area from the edge fusion processing module 2, and obtain the predicted load feature vector for each area from the AI ​​prediction and decision-making module 3; then, it makes decisions based on a preset zoning control strategy library. For areas where the personnel density index is below the threshold and the space usage status is "idle", a standby mode command is generated, closing the fan coil solenoid valve of the corresponding area, adjusting the VAV opening of the variable air volume terminal to the minimum maintenance value, and closing the fresh air valve of the corresponding branch of the air conditioning unit; for areas where the space usage status indicates that a meeting will be held in the next 15 to 30 minutes... For areas where the estimated number of attendees exceeds a threshold, a pre-cooling mode command is generated. 10 to 20 minutes before the meeting begins, the cooling supply to the air conditioning terminals in that area is increased by 20% to 30%, and the fresh air valve opening is preset to 80% to replace indoor air in advance. For areas where the personnel density index remains high, an enhanced cooling command is generated, appropriately reducing the supply air temperature setpoint and increasing the fan speed. After all zone control commands are generated, they are sent to the corresponding terminal devices through the execution feedback module 4. At the same time, the control mode and execution status of each zone are fed back to the AI ​​prediction and decision module 3 as a reference for subsequent optimization.

[0025] Meanwhile, the environmental physical acquisition unit 11, the personnel activity acquisition unit 12, and the space usage status acquisition unit 13 are respectively connected to the edge fusion processing module 2; The environmental physical acquisition unit 11 is used to access outdoor meteorological data through an API interface and to collect indoor temperature, humidity and CO2 concentration data through wireless sensors deployed indoors. The personnel activity collection unit 12 is used to collect and generate regional personnel heat map data through Wi-Fi probes, workstation occupancy sensors or edge computing cameras deployed in various areas; The space usage status acquisition unit 13 is used to connect to the enterprise office collaboration system, access control system or workstation reservation system through the API interface to obtain space usage schedule data.

[0026] In this embodiment, the environmental physics acquisition unit 11 is configured to first send a request to the API interface provided by the meteorological data service provider via the HTTP protocol. The request parameters include the latitude and longitude information of the building's geographical location. After receiving the returned JSON format data, it parses out the hourly outdoor dry-bulb temperature, relative humidity, total horizontal irradiance, and wind speed data for the next 24 hours and caches the data in local memory. At the same time, through the wireless sensor gateways deployed on each floor, it sends read commands to the temperature and humidity sensors distributed in the return air vents, typical office areas, and meeting rooms in a polling manner. After the sensors respond, they return the indoor temperature, relative humidity, and CO2 concentration values ​​at the current moment. The gateway binds these values ​​with the sensor ID and the acquisition timestamp and reports them to the data acquisition center via the MQTT protocol to form a time-series environmental physics database. The personnel activity acquisition unit 12 is configured to continuously capture Probe Request frames periodically broadcast by surrounding smart terminal devices in passive listening mode through Wi-Fi probe devices deployed in corridors, lobbies, and open office areas on each floor. The source MAC address and received signal strength indication value are extracted from the frames. A time window-based MAC address deduplication algorithm is used to count the same MAC address that appears multiple times within the same time window only once. The number of devices counted is mapped to the corresponding spatial grid according to the location of the probes in each area, and a heat map of personnel density covering the entire floor is generated every 5 minutes. At the same time, millimeter-wave radar sensors installed under the workstations or on the ceiling use frequency-modulated continuous wave technology to detect micro-displacement and determine whether the workstation is occupied. The occupancy status is aggregated to the area acquisition unit via RS485 bus. For public areas with edge computing cameras, the AI ​​chip built into the camera runs a people counting algorithm. After completing video stream decoding, human detection and counting locally, only the desensitized people statistics and timestamps are uploaded. The original video stream is not transmitted. All personnel activity data is packaged and sent to the edge fusion processing module 2 after generation.The space usage status acquisition unit 13 is configured to first obtain an access token for the enterprise office collaboration system via OAuth2.0 authentication, then initiate a query request to the calendar API of Microsoft Exchange, DingTalk, or WeChat Work to obtain a list of meeting room reservations for the current day and the next few days. It then parses the meeting room ID, start time, end time, organizer, and estimated number of attendees from each reservation record to construct a meeting room timeline status table. Simultaneously, it connects to the access control system database via the SDK interface to obtain personnel entry and exit records for each area. By counting the number of people leaving the area without swiping their access cards at the current moment, it estimates the real-time personnel inventory in the area. For buildings with a workstation reservation system, it obtains the current reservation status and reservation time period for each workstation area via API. All the above space usage status data are then associated and merged according to the area ID to form a space status time series data containing changes in area occupancy status, occupancy time period, and expected number of personnel, which is then output to the edge fusion processing module 2.

[0027] In addition, the data cleaning unit 21, the feature extraction unit 22, and the privacy protection unit 23 are connected in sequence; The data cleaning unit 21 is used to align the timestamps of multimodal data with different sampling frequencies and remove outliers; The feature extraction unit 22 is used to quantify the RSSI signal of the Wi-Fi probe into a people density index, and combine the meeting schedule and people density data to generate a predicted load feature vector for the next 15-30 minutes. The privacy protection unit 23 is used to anonymize personnel data at the edge and output only the regional density level data.

[0028] In this embodiment, the data cleaning unit 21 is configured to first align the temperature and humidity data reported by the environmental physical acquisition unit 11, the personnel density data reported by the personnel activity acquisition unit 12, and the schedule data reported by the space usage status acquisition unit 13 with timestamps according to a unified time base. For temperature and humidity data with a sampling frequency of one minute and personnel density data with a sampling frequency of every 5 minutes, the nearest neighbor interpolation method is used to downsample the high-frequency data to the same time granularity as the low-frequency data. Then, anomaly detection is performed on the raw signal strength data reported by the Wi-Fi probe. For abrupt values ​​of zero or more than three times the standard deviation of the historical mean at multiple consecutive time points, the sliding window mean based on the historical data of the same period is used for replacement. For temporary data loss caused by communication failure in the sensor network, a linear interpolation algorithm is used to fill in the missing data based on the effective data points before and after, ensuring that the data stream output to subsequent units is continuous and complete. The feature extraction unit 22 is configured to first divide the count value of Wi-Fi probe devices in each cleaned area by the baseline coverage capacity of the probes deployed in that area to obtain a normalized personnel density index. The index ranges from 0 to 100, where 0 indicates no personnel activity in the area and 100 indicates that the area has reached the maximum designed personnel density. Then, it parses the meeting schedule data provided by the space usage status acquisition unit 13, extracts the meeting records that will start in the next 30 minutes, calculates the expected personnel increase of the meeting event based on the meeting room area and number of seats, and superimposes the increase onto the current personnel density index to form a predicted personnel density curve. Next, it uses the slope of the indoor CO2 concentration change in the past 15 minutes as a reference feature for ventilation demand. Combined with the predicted personnel density curve, it generates a predicted load feature vector containing area identifier, timestamp, predicted personnel density, expected CO2 concentration change direction, and load urgency. This feature vector is encapsulated in JSON format and then sent to the privacy protection unit 23. The privacy protection unit 23 is configured to first receive the predicted load feature vector and the original Wi-Fi probe data output by the feature extraction unit 22. For the MAC address list in the original Wi-Fi probe data, the SHA-256 hash algorithm is used to encrypt each MAC address one-way, and the encrypted hash value is bound to the corresponding signal strength value. The hashing process is irreversible and does not retain the original MAC address mapping relationship, thus blocking the traceability of personal identity information from the source. For the number of people data uploaded by the edge computing camera, the privacy protection unit 23 only verifies the data integrity and then transmits it directly without any additional processing. All data after hashing is aggregated into regional statistical information before leaving the edge gateway. Finally, only the regional population density index and the predicted load feature vector are used as output content and sent to the AI ​​prediction decision module 3 of the cloud or local server through the TLS encrypted transmission channel.

[0029] Furthermore, the optimization decision-making unit 32 and the load forecasting unit 31 are connected; The load prediction unit 31 is used to use a time-series prediction model, taking historical indoor environmental data, future weather forecast data and future population density prediction data as input, and output hourly cold and heat load prediction values ​​and predicted CO2 concentration change curves for each zone within the next 30-60 minutes. The optimization decision unit 32 is used to generate an optimized control strategy set based on the prediction data output by the load prediction unit 31, including the chilled water outlet temperature setpoint, chiller start-up and shutdown plan, water pump frequency, fresh air valve opening degree and zone standby strategy.

[0030] In this embodiment, the load prediction unit 31 integrates a time-series prediction model. It is configured to first read environmental physical data sequences and population density data sequences for each region from a historical database over the past hour, obtain outdoor weather forecast data for the next 30 minutes from a meteorological API, and obtain predicted population density sequences for each region for the next 30 minutes from the edge fusion processing module 2. These data are then standardized and concatenated into a multi-dimensional feature tensor. A pre-trained LSTM or Transformer network model is then loaded, and the feature tensor is input into the model for forward computation. The model's multi-layered neurons sequentially encode and decode the temporal dependencies, ultimately outputting the prediction results for the next 30 to 60 minutes through a fully connected layer. The predicted cooling load is output in kilowatts, and the predicted CO2 concentration is output in ppm. Each prediction time step is set to 15 minutes, and the output results include confidence intervals for both the upper and lower bounds of the predicted values, which are then used by the optimization decision unit 32 to formulate strategies. The optimization decision unit 32 integrates a model predictive control algorithm. It is configured to first acquire the hourly cooling load forecasts and CO2 concentration forecasts for each zone within the next hour, output by the load forecast unit 31. Simultaneously, it reads the real-time operating status parameters of the current central air conditioning system, including chiller load rate, chilled water supply and return temperatures, cooling water inlet and outlet temperatures, pump operating frequencies, and valve openings. Then, it constructs an optimization problem with the objective function of minimizing total system energy consumption, constrained by indoor temperature and humidity ranges and CO2 concentration limits. The model predictive control algorithm is used to solve this problem in the prediction time domain, and the solver calculates the optimal control sequence for the next hour. If the system is configured in reinforcement learning mode, the decision unit loads a trained deep Q-network or a near-end strategy optimization model, using the current system state as input. The model directly outputs the optimal action value, with the action space including the chilled water outlet temperature setpoint, pump frequency adjustment, and fresh air valve opening adjustment. Finally, the generated control strategy set is sorted by execution priority and output to the execution feedback module 4.

[0031] The load prediction unit 31 utilizes a time-series prediction model that is either an LSTM network or a Transformer network.

[0032] Finally, the protocol conversion unit 41, the instruction issuing unit 42, and the feedback acquisition unit 43 are connected in sequence; The protocol conversion unit 41 is used to convert the optimized control strategy set into control commands that can be recognized by the actuators of the central air conditioning system; The instruction issuing unit 42 is used to issue control instructions to the building automation system or directly control the frequency converter, electric valve and chiller controller. The feedback acquisition unit 43 is used to collect actual energy consumption data and environmental change data in real time, and compare them with the predicted values ​​output by the load prediction unit 31. The comparison results are used as feedback data and input to the AI ​​prediction decision module 3 for online model learning.

[0033] In this embodiment, the protocol conversion unit 41 is configured to first parse the optimized control strategy set output by the AI ​​prediction and decision module 3. This strategy set is described in JSON format and includes device identifier, control parameter name, and target value fields. Then, based on the configuration information of the building automation system, it determines the communication protocol type corresponding to each device. For devices using the BACnet protocol, the control parameters are mapped to BACnet object attributes and encapsulated into BACnet write attribute request messages through the BACnet protocol stack. For devices using the Modbus TCP / IP protocol, the control parameters are converted into register addresses and write values ​​according to the Modbus address mapping table and encapsulated into Modbus protocol frames. For devices using the OPC UA protocol, a secure channel is established through the OPC UA client, and the control parameters are written to the corresponding node variables. After the conversion is completed, each protocol message is grouped by device and sent to the instruction issuing unit 42. The instruction issuing unit 42 is configured to first receive the multi-protocol control message generated by the protocol conversion unit 41, and establish a TCP connection or maintain an existing long connection channel according to the network address of the target device in the message. For the DDC controller, it sends a message containing the chilled water outlet temperature setpoint and fan frequency command to the controller via the BACnet / IP protocol, and records the sending timestamp and execution status after waiting for the controller to return a confirmation response. For the frequency converter, it sends the frequency adjustment command to the frequency converter via the Modbus TCP / IP protocol, reads the current frequency value returned by the frequency converter for closed-loop verification, and confirms that the command was executed successfully. For the chiller group control controller, it writes chiller start / stop combination and load distribution commands via the OPC UA protocol, and reads the operating status returned by the chiller for confirmation. After all commands are issued, an execution log is generated, recording the issuance time, target device, command content and execution result of each command, for use by the feedback acquisition unit 43. The feedback acquisition unit 43 is configured to first read real-time active power data of each chiller, water pump, fan and cooling tower from the smart meter at regular intervals via Modbus TCP / IP protocol, and read the instantaneous flow rate and supply and return water temperature of the chilled water system from the heating and cooling water meter to calculate the real-time cooling capacity; at the same time, it reads the real-time values ​​of indoor temperature, humidity and CO2 concentration reported by the DDC controller of each area from the building automation system via BACnet protocol, with the sampling frequency set to once every 5 minutes; then, it compares the collected actual operating data with the predicted load value and predicted CO2 concentration value at the same time point output by the load prediction unit 31 item by item, and calculates the absolute error and relative error percentage; then, it packages the error data, actual operating data and corresponding timestamps into a feedback data packet, and uploads it to the model training service of the AI ​​prediction decision module 3 via HTTPS protocol to trigger the incremental learning process of the model; after the model training service collects a certain amount of feedback data, it updates the model parameters using a mini-batch gradient descent algorithm so that the model continuously adapts to changes in building load characteristics.

[0034] This invention discloses an AI-based central air conditioning energy consumption optimization system. Through a multimodal data acquisition module 1, it acquires real-time forward-looking data such as meeting schedules and personnel density. Combined with a time-series prediction model in the AI ​​prediction and decision-making module 3, it can accurately predict changes in cooling and heating loads and CO2 concentrations within the next 30 to 60 minutes. Based on this, the system can pre-cool and pre-ventilate, eliminating temperature fluctuations and energy waste caused by lag in traditional control methods, significantly improving the stability of indoor environmental comfort. This invention integrates three types of multimodal data sources: environmental physical data, personnel activity data, and space usage status data, constructing a comprehensive perception capability for personnel activity and space usage. Compared to existing technologies that rely solely on temperature and humidity sensors, this invention can accurately reflect the dynamic load demand within the building, providing a more accurate data foundation for optimization decisions. This system possesses predictive capabilities, enabling fine-grained control of central air conditioning in building zones, thereby reducing central air conditioning energy consumption.

[0035] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. An AI-based central air conditioning energy consumption optimization system, characterized in that, It includes a multimodal data acquisition module, an edge fusion processing module, an AI prediction and decision-making module, and an execution feedback module, wherein the multimodal data acquisition module, the edge fusion processing module, the AI ​​prediction and decision-making module, and the execution feedback module are connected in sequence; The multimodal data acquisition module is used to collect environmental physical data, personnel activity data, and space usage status data in real time. The edge fusion processing module is used to clean, align, and extract features from the collected multimodal data, and output a region-level predicted load feature vector. The AI ​​prediction and decision-making module is used to run an AI algorithm model based on the predicted load feature vector to generate predicted load data and an optimized control strategy set for the central air conditioning system. The execution feedback module is used to convert the optimized control strategy set into control commands and send them to the actuators of the central air conditioning system, and to collect actual operating data and feed it back to the AI ​​prediction and decision module for model optimization.

2. The AI-based central air conditioning energy consumption optimization system as described in claim 1, characterized in that, The AI-based central air conditioning energy consumption optimization system also includes a zoning control module, which is connected to the AI ​​prediction and decision-making module and the execution feedback module respectively. The zoning control module is used to automatically implement standby mode control for unmanned or vacant areas based on the regional personnel density index and space usage status data in the predicted load feature vector, and to implement pre-cooling or pre-ventilation control for high personnel density areas.

3. The AI-based central air conditioning energy consumption optimization system as described in claim 2, characterized in that, The multimodal data acquisition module includes an environmental physical acquisition unit, a personnel activity acquisition unit, and a space usage status acquisition unit, which are respectively connected to the edge fusion processing module. The environmental physical acquisition unit is used to access outdoor meteorological data through an API interface and to collect indoor temperature, humidity and CO2 concentration data through wireless sensors deployed indoors. The personnel activity collection unit is used to collect and generate regional personnel heat map data through Wi-Fi probes, workstation occupancy sensors or edge computing cameras deployed in various areas; The space usage status acquisition unit is used to connect to enterprise office collaboration systems, access control systems, or workstation reservation systems via API interfaces to obtain space usage schedule data.

4. The AI-based central air conditioning energy consumption optimization system as described in claim 3, characterized in that, The edge fusion processing module includes a data cleaning unit, a feature extraction unit, and a privacy protection unit, which are connected in sequence. The data cleaning unit is used to align timestamps on multimodal data with different sampling frequencies and remove outliers. The feature extraction unit is used to quantify the RSSI signal of the Wi-Fi probe into a people density index, and combine the meeting schedule and people density data to generate a predicted load feature vector for the next 15-30 minutes. The privacy protection unit is used to anonymize personnel data at the edge, outputting only region-level density level data.

5. The AI-based central air conditioning energy consumption optimization system as described in claim 4, characterized in that, The AI ​​prediction and decision-making module includes a load prediction unit and an optimization decision-making unit, and the optimization decision-making unit and the load prediction unit are connected. The load forecasting unit is used to use a time-series forecasting model, taking historical indoor environmental data, future weather forecast data and future population density forecast data as input, and output hourly cold and heat load forecasts and predicted CO2 concentration change curves for each zone within the next 30-60 minutes. The optimization decision unit is used to generate an optimized control strategy set based on the prediction data output by the load prediction unit, including the chilled water outlet temperature setpoint, chiller start-up and shutdown plan, water pump frequency, fresh air valve opening degree, and zone standby strategy.

6. The AI-based central air conditioning energy consumption optimization system as described in claim 5, characterized in that, The load forecasting unit utilizes a time-series forecasting model either an LSTM network or a Transformer network.

7. The AI-based central air conditioning energy consumption optimization system as described in claim 6, characterized in that, The execution feedback module includes a protocol conversion unit, an instruction issuing unit, and a feedback acquisition unit; the protocol conversion unit, the instruction issuing unit, and the feedback acquisition unit are connected in sequence. The protocol conversion unit is used to convert the optimized control strategy set into control commands that can be recognized by the actuators of the central air conditioning system; The instruction issuing unit is used to issue control instructions to the building automation system or directly control the frequency converter, electric valve and chiller controller. The feedback acquisition unit is used to collect actual energy consumption data and environmental change data in real time, and compare them with the predicted values ​​output by the load prediction unit. The comparison results are then used as feedback data to be input into the AI ​​prediction decision module for online model learning.

Images