Air conditioner control terminal system based on data acquisition analysis

Abstract

The application relates to the technical field of air conditioner system control, and particularly discloses an air conditioner control terminal system based on data collection and analysis, which aims to solve the problems of control inaccuracy, low energy efficiency and poor comfort caused by insufficient data collection and analysis. The system comprises a multi-source heterogeneous data collection module, a data fusion and feature extraction module, a dynamic double-model collaborative construction module, a forward-looking strategy generation module and a self-adaptive execution and feedback module. Through full-dimensional data collection and fusion, the dynamic models of equipment energy efficiency and user comfort are constructed and updated in parallel, forward-looking multi-objective rolling optimization and closed-loop feedback correction are carried out based on the double models, and the long-term optimal balance of energy efficiency and comfort of the air conditioner system in a dynamic environment is realized.

Description

Technical Field

[0001] This invention belongs to the field of air conditioning system control technology, specifically relating to an air conditioning control terminal system based on data acquisition and analysis. Background Technology

[0002] In the field of building environment and energy management, intelligent control of air conditioning systems is a key technological direction for achieving energy conservation and emission reduction, improving user comfort, and optimizing equipment operating efficiency. This field integrates sensors, controllers, and communication networks to monitor indoor and outdoor environmental parameters and adjust air conditioning equipment according to preset strategies or algorithms to achieve precise management of the building's thermal environment.

[0003] Among them, air conditioning control terminal systems based on data acquisition and analysis are currently a research and application hotspot. These systems aim to achieve a dynamic balance between energy efficiency and comfort in air conditioning systems by continuously collecting multi-source data related to air conditioning operation and using data analysis techniques to generate optimized control commands.

[0004] In existing technologies, such systems typically rely on the collection of basic environmental parameters such as temperature and humidity, and their control logic is mostly based on simple threshold judgments or preset fixed operating modes. However, existing systems have significant limitations in the comprehensiveness of data collection and the depth of analysis: on the one hand, the collected data dimensions are limited, failing to effectively integrate key dynamic information that directly affects air conditioning load and energy efficiency, such as personnel distribution density, equipment operating current, and the opening and closing status of doors and windows, leading to inaccurate judgments of actual indoor heat demand and equipment operating status; on the other hand, their data analysis methods are relatively simple, lacking deep integration and intelligent mining of historical operating data and real-time collected data, making it difficult to build accurate equipment energy efficiency models and user comfort models, and thus unable to generate personalized and forward-looking control strategies that adapt to complex and changing actual operating conditions. This directly leads to problems such as energy waste, uneven regional heating and cooling, and delayed response in the actual operation of air conditioning systems, hindering further improvements in intelligent management. Summary of the Invention

[0005] The purpose of this invention is to provide an air conditioning control terminal system based on data acquisition and analysis, so as to solve the problems of inaccurate control, low energy efficiency and poor comfort of air conditioning systems caused by the single dimension of data acquisition and insufficient depth of analysis in the prior art.

[0006] The technical solution of this invention is an air conditioning control terminal system based on data acquisition and analysis. This system includes a multi-source heterogeneous data acquisition module, a data fusion and feature extraction module, a dynamic dual-model collaborative construction module, a forward-looking strategy generation module, and an adaptive execution and feedback module. These modules work together to form a complete technical system from data perception to intelligent decision-making and then to closed-loop optimization.

[0007] The multi-source heterogeneous data acquisition module is used to collect real-time dynamic data related to the operation of the air conditioning system and the building environment across all dimensions. This module integrates an environmental parameter sensing unit, a personnel activity sensing unit, an equipment status monitoring unit, and a building structure status sensing unit. The environmental parameter sensing unit is deployed in various indoor zones and on the building facade to collect environmental data including air temperature, relative humidity, carbon dioxide concentration, and light intensity. The personnel activity sensing unit uses a fusion sensing technology combining distributed infrared thermal imaging sensors and millimeter-wave radar sensors to acquire real-time information on the number of people, their distribution density, and the intensity of static or dynamic activity in each zone. The equipment status monitoring unit uses non-invasive current and vibration sensors to collect real-time operating current, power factor, and vibration spectrum data of core equipment such as air conditioning terminal fan coil units, compressors, and circulating water pumps. The building structure status sensing unit uses magnetic induction sensors installed on main doors and windows and a distributed temperature sensor array deployed on the building envelope surface to collect data on the opening and closing status of doors and windows and the temperature difference between the inner and outer surfaces of the building envelope. All acquisition units synchronously upload data to the system's central processing unit via wired or wireless communication networks at a preset sampling frequency.

[0008] The data fusion and feature extraction module performs spatiotemporal alignment, noise reduction and cleaning, and deep feature extraction on the raw data uploaded by the multi-source heterogeneous data acquisition module. This module first establishes a unified spatiotemporal reference, synchronizing the timestamps of all acquired data and mapping them to a pre-defined 3D spatial grid in the Building Information Model (BIM) to achieve precise alignment of the data in both time and space. Further, the module employs a wavelet transform-based filtering algorithm to reduce noise in the raw sensor data and uses box plot analysis to identify and remove outlier data points. After data preprocessing, the module performs deep feature extraction, specifically including: extracting first- and second-order derivative features such as temperature change rate, humidity gradient, and carbon dioxide accumulation rate from time-series environmental data; extracting time-varying curves of zonal population density and activity heat load intensity indices from personnel activity data; extracting features characterizing equipment health and operating efficiency such as current harmonic distortion rate, instantaneous power efficiency, and vibration dominant frequency offset from equipment status data; and extracting dynamic estimates of thermal resistance of the building envelope and estimated air infiltration volume from doors and windows from building structure status data. All extracted high-dimensional feature vectors are packaged into a unified feature data package for use by downstream modules.

[0009] The dynamic dual-model collaborative construction module is the core analysis engine of the system, used to construct and continuously update dynamic models of equipment energy efficiency and user comfort based on historical data and real-time feature data. This module includes an energy efficiency model construction unit and a comfort model construction unit. The energy efficiency model construction unit adopts a physically constrained neural network architecture. The input layer of this network receives real-time feature data packets from the data fusion and feature extraction module, including equipment operating current, power factor, vibration spectrum characteristics, indoor and outdoor temperature difference, occupant heat load intensity index, and dynamic estimates of building envelope thermal resistance. The hidden layer of the network incorporates the thermodynamic equations of the air conditioning refrigeration cycle as a physical constraint layer, ensuring that the model predictions conform to basic physical laws. This unit uses an online learning algorithm, utilizing real-time collected equipment power consumption data as a supervision signal, to dynamically adjust network weights, thereby outputting the predicted instantaneous energy efficiency ratio of the overall air conditioning system under the current operating conditions, and calculating the theoretical optimal operating parameter range for each terminal device.

[0010] The comfort model construction unit employs a personalized transfer learning framework. This unit first pre-trains a general thermal comfort prediction model based on historical environmental data and anonymized occupant distribution data for each area within the building. This model's core outputs are the predicted average voting index and the predicted percentage of dissatisfaction. Further, when the system detects a stable presence of people in a specific zone, this unit initiates a personalized adaptation process. By continuously collecting real-time temperature, humidity, wind speed, and micro-motion information from the occupant activity sensing unit within that zone, incremental learning techniques are used to fine-tune the basic model, thereby generating a dynamic comfort model specific to the current occupants in that zone. This model can output personalized temperature setting ranges and wind speed preference parameters that meet the comfort requirements of 95% of the occupants in the current area.

[0011] The forward-looking strategy generation module is used to perform multi-objective optimization calculations based on the output of the dynamic dual-model collaborative construction module, generating a forward-looking sequence of air conditioning system control strategies. This module incorporates a multi-objective optimization solver whose objective function simultaneously minimizes the predicted total energy consumption of the system over the next two hours and maximizes the overall thermal comfort satisfaction of each zone. Optimization variables include the set temperature, airflow velocity, and airflow angle of the air conditioning terminals in each zone, as well as the outlet water temperature and pump frequency of the central cooling source equipment. The constraints of the optimization process are derived from the output of the dynamic dual-model collaborative construction module; that is, the equipment operating parameters must be within the theoretically optimal operating parameter range calculated by the energy efficiency model, and the environmental parameters must be within the personalized comfort range output by the comfort model. Furthermore, this module also introduces door and window status data from the multi-source heterogeneous data acquisition module as hard constraints. When it is detected that a door or window in a certain zone has been continuously open for more than 5 minutes, the optimization solver will temporarily block the cooling capacity allocation for that zone in subsequent strategies until the door or window is closed. This module generates a sequence of optimal control strategies for eight time steps over the next two hours at 15-minute intervals, but only issues and executes the strategy instructions for the first time step.

[0012] The adaptive execution and feedback module executes control commands issued by the forward-looking strategy generation module and collects actual system response data to form a closed-loop optimization. This module includes a command distribution unit, an execution status monitoring unit, and a model calibration unit. The command distribution unit converts received setting parameters into specific equipment control commands and sends them to the corresponding air conditioning terminal controllers and central chiller station controllers via industrial bus protocols or IoT protocols. The execution status monitoring unit collects real-time operating data and environmental change data after the equipment executes the commands, forming an execution feedback data packet. The model calibration unit compares the execution feedback data packet with the prediction data used by the forward-looking strategy generation module when generating the strategy, calculating the prediction deviation. If key indicators such as zone temperature deviation consistently exceed 0.5 degrees Celsius or equipment power consumption deviation consistently exceed 10%, the model calibration unit triggers the model recalibration process of the dynamic dual-model collaborative construction module, injecting new data with high weights to drive rapid adaptive updates to the model, thereby ensuring the system maintains decision-making accuracy when facing sudden changes in building usage patterns or equipment performance degradation.

[0013] As one embodiment of the present invention, the data fusion process of the infrared thermal imaging sensor and the millimeter-wave radar sensor in the personnel activity sensing unit is as follows: First, the original point cloud data collected by the two types of sensors are aligned with the coordinate system and the timestamp; then, a spatial clustering algorithm based on density clustering is used to identify heat source clusters from the infrared point cloud and moving target and stationary target clusters from the radar point cloud; finally, the Hungarian matching algorithm is used to associate and match heat source clusters and target clusters in the same spatiotemporal range, eliminate interference caused by non-human heat sources, and comprehensively judge the activity status of the target, outputting a matrix of personnel number and activity intensity accurate to the partition grid.

[0014] As one embodiment of the present invention, the energy efficiency model construction unit adopts a physical constraint-based neural network. The physical constraint layer is implemented as follows: during the forward propagation of the neural network, some intermediate variables output by the hidden layer are forcibly substituted into a set of preset refrigerant pressure-enthalpy relationship and heat exchanger efficiency equation for calculation. The calculation results are used as constraint loss terms and network prediction loss terms to participate in back propagation and weight update, thereby embedding the constraints of the first and second laws of thermodynamics into the data-driven model.

[0015] In one embodiment of the present invention, the personalized adaptation process of the comfort model construction unit adopts a federated learning framework. Terminal devices within each partition fine-tune their local comfort model copy using locally collected real-time data, encrypting and uploading only the updated gradients of the model weights to the central processing unit for secure aggregation, generating a global model update, and then distributing it to each partition. This process protects individual user privacy data while achieving the sharing and evolution of building-wide comfort knowledge.

[0016] In one embodiment of the present invention, the multi-objective optimization solver in the prospective strategy generation module employs a non-dominated sorting genetic algorithm with an elite strategy. This algorithm treats energy efficiency and comfort objectives as two independent fitness functions, iteratively evolves a control strategy population through selection, crossover, and mutation operations, and finally selects the solution closest to the preset weight preference from the non-dominated solution set as the optimal strategy sequence output, ensuring that satisfactory solutions can still be found efficiently under complex constraints.

[0017] In one embodiment of the present invention, the system also integrates a digital twin simulation verification submodule. Before issuing the control strategy sequence to the real equipment, this submodule first performs rapid simulation and deduction in a high-fidelity digital twin constructed based on the building information model and the actual parameters of the equipment, predicting the environmental and energy consumption results after the strategy is executed. The instruction is only allowed to be issued when the simulation results meet the safety threshold and expected benefits; otherwise, the strategy will be returned to the forward-looking strategy generation module for re-optimization.

[0018] In summary, this application includes at least one of the following beneficial technical effects: First, this invention fundamentally changes the traditional system's siloed data structure by constructing a comprehensive data acquisition system covering environment, personnel, equipment, and building structure, and by deeply integrating and extracting its features. The system can accurately perceive various dynamic factors affecting air conditioning load, including zone-specific heat load from personnel activities, real-time energy efficiency status of equipment, and changes in the thermal performance of the building envelope. This lays a solid data foundation for subsequent accurate modeling and optimization decisions, fundamentally solving the problem of inaccurate control caused by incomplete information.

[0019] Second, this invention innovatively proposes a dynamic dual-model collaborative construction mechanism, deeply integrating data-driven machine learning with domain physics knowledge. The dynamic energy efficiency model for equipment, by introducing physical constraints, ensures that the prediction results conform to thermodynamic laws, improving its generalization ability and reliability under unexperienced operating conditions. The dynamic user comfort model, through personalized transfer learning and federated learning frameworks, achieves precise adaptation from general comfort standards to real-time preferences of individuals or groups. The parallel updates and mutual constraints of the two models together constitute a digital mirror that can accurately characterize the real-time state and performance boundaries of the system, making control decisions based on evidence and bounded by rules.

[0020] Third, this invention replaces traditional real-time reactive control with forward-looking rolling optimization. Based on dual-model prediction, the system solves for the optimal control sequence for future time periods using a multi-objective optimization approach. This not only considers current comfort and energy saving but also proactively smooths out equipment operating load, avoiding energy waste and equipment damage caused by frequent start-ups or sudden setpoint changes. Simultaneously, the introduction of digital twin simulation verification and closed-loop feedback correction mechanisms ensures the safety and robustness of the optimization strategy, enabling the system to continuously self-optimize and adapt to environmental changes. Ultimately, this achieves a long-term optimal balance between energy efficiency and comfort in dynamically changing environments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the air conditioning control terminal system based on data acquisition and analysis proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic dual-model collaborative construction module in this invention; Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow of multi-source heterogeneous data acquisition and fusion processing in this invention; Figure 4 This is a logical flow diagram of the closed loop of forward-looking strategy generation and adaptive execution in this invention. Detailed Implementation

[0022] The overall technical architecture of the air conditioning control terminal system based on data acquisition and analysis proposed in this invention is shown in the attached figure. Figure 1 As shown in the figure, the system consists of five core functional units: a multi-source heterogeneous data acquisition module, a data fusion and feature extraction module, a dynamic dual-model collaborative construction module, a forward-looking strategy generation module, and an adaptive execution and feedback module. These units communicate via a high-bandwidth, low-latency data bus, forming a complete closed-loop intelligent control system from perception, modeling, decision-making to execution and feedback. The following will be discussed in conjunction with the attached diagram. Figure 1 To be continued Figure 4 The specific implementation methods of each module are described in detail layer by layer.

[0023] The multi-source heterogeneous data acquisition module serves as the system's sensing front end, deployed in various functional zones and key equipment nodes within the building to acquire real-time dynamic information affecting the operating status of the air conditioning system and the quality of the indoor thermal environment.

[0024] This module comprises four sub-units: an environmental parameter sensing unit, a personnel activity sensing unit, an equipment status monitoring unit, and a building structure status sensing unit.

[0025] The environmental parameter sensing unit is equipped with no less than three composite air temperature and humidity sensors, one carbon dioxide concentration sensor, and one light intensity sensor in each air-conditioned service zone. Two outdoor weather stations are also installed on the exterior wall facade to collect basic environmental variables such as indoor and outdoor air temperature, relative humidity, carbon dioxide volume fraction, and solar irradiance.

[0026] All sensors employ industrial-grade precision, with a sampling frequency set to once every 10 seconds, and are uploaded to the central processing unit via the building's local area network using a timestamp synchronization mechanism.

[0027] The personnel activity sensing unit adopts a heterogeneous fusion sensing architecture combining infrared thermal imaging sensors and millimeter-wave radar sensors. The infrared thermal imaging sensors are deployed at the center of the ceiling in each zone, covering the entire area with a field of view, and outputting thermal image point clouds with a resolution of 64×64.

[0028] The millimeter-wave radar sensor is installed high on the wall, operates at a frequency of 77 GHz, can penetrate non-metallic obstacles, and outputs four-dimensional point cloud data including target distance, speed, angle, and reflection intensity.

[0029] The data streams from both types of sensors are first aligned with coordinate system 1 and timestamps at the local edge computing node, and then fed into a clustering and matching engine. This engine first performs a density-based clustering algorithm on the infrared point cloud to identify several heat source clusters; at the same time, it performs a static and dynamic target separation algorithm on the radar point cloud to distinguish between stationary and moving human bodies.

[0030] Finally, the Hungarian matching algorithm is used to optimally associate heat source clusters and radar target clusters within the same spatial grid, eliminating false detections caused by non-human heat sources such as radiators and electronic devices, and outputting the number of people, distribution center coordinates, and activity intensity level (divided into three levels: sedentary, light activity, and moderate activity) in each partition grid to the nearest whole number.

[0031] The output of this process is encapsulated into a personnel activity matrix, updated every 30 seconds, for use by subsequent modules.

[0032] The equipment status monitoring unit monitors the operating status of key equipment in the air conditioning system through non-intrusive current transformers and triaxial vibration acceleration sensors.

[0033] The current transformer is clamped on the power supply line of the fan coil unit, compressor and circulating water pump, with a sampling frequency of 5 kHz, and is used to acquire the instantaneous current waveform, active power, reactive power and power factor of the equipment in real time; the vibration sensor is attached near the bearing seat of the equipment housing, with a sampling frequency of 10 kHz, and is used to collect the three-dimensional vibration signal of the equipment during operation.

[0034] After the original current and vibration signals are converted from analog to digital locally, frequency domain features are extracted by fast Fourier transform, including total harmonic distortion of current, fundamental phase offset, vibration dominant frequency and its amplitude, and high-frequency energy ratio.

[0035] These features are encapsulated into a device health status vector and uploaded to the central processing unit once per minute.

[0036] The building structure status sensing unit consists of magnetic induction switches for doors and windows and a distributed temperature sensor array for the building envelope. The magnetic induction switches are installed at the frame-sash joints of all operable exterior windows and main interior doors to report the opening and closing status of doors and windows in real time; the distributed temperature sensors are arranged at 0.5-meter intervals along the inner and outer surfaces of the exterior walls to form a two-dimensional temperature field grid for measuring the temperature difference between the inner and outer surfaces of the building envelope.

[0037] This unit uploads the Boolean value sequence of door and window status and the surface temperature field data of the building envelope every 5 minutes. The data streams of all four types of sub-units are encapsulated through the building IoT protocol stack to ensure that the timestamp accuracy error is less than 50 milliseconds, and it has breakpoint resume and data integrity verification mechanisms.

[0038] The data fusion and feature extraction module receives raw data streams from the multi-source heterogeneous data acquisition module and performs a three-stage processing flow of spatiotemporal alignment, noise reduction and cleaning, and deep feature extraction. This module first establishes a unified spatiotemporal reference framework, using a 3D spatial grid from the Building Information Model (BIM) as its base map, with a grid resolution of 1 meter × 1 meter × 1 meter, and using Coordinated Universal Time (UTC) as the time base. All input data are mapped to corresponding spatiotemporal grid nodes based on their physical location labels and acquisition timestamps, completing the spatiotemporal alignment.

[0039] For missing or delayed data points, a spatiotemporal interpolation algorithm based on Kalman filtering is used to fill them in, ensuring that each spatiotemporal grid has complete data records in each sampling period.

[0040] After completing the spatiotemporal alignment, the module starts the noise reduction and cleaning subroutine. For continuous variables such as environmental parameters and equipment current, discrete wavelet transform is used for multi-scale decomposition, retaining low-frequency approximation coefficients and removing high-frequency noise detail coefficients before reconstruction to achieve smooth filtering. For discrete variables such as the number of people and the status of doors and windows, the sliding window mid-range filtering method is used to suppress instantaneous jumps.

[0041] Outlier detection employs an improved boxplot method, which calculates the interquartile range for each variable's historical sliding window (1440 sampling points, corresponding to 24 hours) and sets an outlier threshold as the upper quartile plus 2.5 times the interquartile range or the lower quartile minus 2.5 times the interquartile range. Data points exceeding this range are marked as outliers and either removed or replaced with nearby valid values.

[0042] Deep feature extraction is the core function of this module. For time-series environmental parameter data, the first derivative (i.e., the rate of temperature change) and the second derivative (i.e., the temperature acceleration) of temperature with respect to time are calculated. The humidity gradient is defined as the spatial rate of change of absolute air humidity along the vertical direction, and the carbon dioxide accumulation rate is estimated by the ratio of the concentration increment to the volume of the partition per unit time.

[0043] For personnel activity data, extract the slope of the curve showing the change in personnel density over time in different zones, and the activity heat load intensity index (this index comprehensively considers the number of people, activity level, and duration; the calculation formula is: ). Where N is the number of people, The weighting coefficient for the i-th person is 1.0 for sedentary, 1.2 for light activity, and 1.5 for moderate activity. (The percentage of the activity duration to the sampling period); for equipment status data, calculate the current harmonic distortion rate (THD), instantaneous power efficiency (defined as the ratio of active power to apparent power), and the offset of the dominant vibration frequency relative to the frequency corresponding to the rated speed of the equipment; for building structure data, use the temperature difference between the inner and outer surfaces of the building envelope and the known heat flow direction to estimate the dynamic thermal resistance by simplifying the heat conduction equation, and combine the door and window opening time and wind speed data to estimate the infiltration air volume.

[0044] All of the above features are normalized and encapsulated into high-dimensional feature vectors with a dimension of no less than 128. These vectors are then stored in the feature database in a unified data packet format for downstream modules to access as needed.

[0045] The dynamic dual-model collaborative construction module is the core analysis engine of the system, and its internal structure is shown in the attached figure. Figure 2 As shown.

[0046] This module runs two independent but interconnected sub-models in parallel: a dynamic model of device energy efficiency and a dynamic model of user comfort. The energy efficiency model building unit employs a deep neural network architecture with embedded physical constraints.

[0047] The input layer of the network receives real-time feature vectors from the feature database, which include 32-dimensional input variables such as equipment operating current, power factor, vibration spectrum characteristics, indoor and outdoor temperature difference, personnel heat load intensity index, and dynamic estimation of thermal resistance of building envelope.

[0048] The network consists of three fully connected hidden layers, each containing 256 neurons, with modified linear units as the activation function. The key innovation lies in introducing a physically constrained layer after the second hidden layer. This layer does not participate in conventional weight learning but instead forces calculations by substituting some intermediate output variables into a pre-defined thermodynamic equation for a refrigeration cycle.

[0049] Specifically, the compressor power consumption predicted by the network must satisfy the following relationship with the evaporator / condenser heat exchange:

[0050] in, For system energy efficiency ratio, To absorb heat from the evaporator, For compressor power consumption, and These are the evaporation temperature and condensation temperature (estimated from sensor data), respectively. This inequality originates from the second law of thermodynamics, constituting a rigid physical boundary. During network training, in addition to the conventional mean squared error loss, a penalty term for violating this boundary is added, together forming the total loss function. The network weights are jointly optimized through the backpropagation algorithm. The model receives the latest measured power consumption values ​​of the devices online every 15 minutes as a supervision signal, and uses stochastic gradient descent for fine-tuning, thereby continuously outputting the instantaneous energy efficiency ratio prediction value under the current operating conditions, and back-calculating the theoretical optimal operating parameter range of each terminal device (such as fan coil unit speed and water valve opening).

[0051] The comfort model building unit employs a personalized transfer learning framework. This unit first uses three months of historical data accumulated during the initial operation of the building to train a general thermal comfort prediction model. The model's core outputs are PMV (Predicted Average Voting Value) and PPD (Predicted Percentage of Dissatisfaction), while the input variables include the zone's average temperature, relative humidity, average wind speed, average radiant temperature, and thermal resistance of clothing.

[0052] When the system continuously detects the presence of people in a specific partition for more than 30 minutes, it triggers a personalized adaptation process. This process uses real-time collected micro-environmental data (temperature fluctuation standard deviation, local wind speed vector, frequency of personnel micro-movements, etc.) and anonymized activity patterns to incrementally fine-tune the last fully connected layer of the base model.

[0053] The fine-tuning process employs a federated learning mechanism: each edge computing node in each partition holds a local copy of the model, encrypting and uploading only the gradient updates of the model weights to the central server; the central server aggregates the gradient updates from all partitions, generates a global model increment, and distributes it to each partition for local model synchronization. This mechanism ensures that individual privacy data does not leave the local machine, while simultaneously enabling the collaborative evolution of group comfort knowledge. The final output dynamic comfort model can provide the current partition's population with personalized temperature setting ranges (typically 1.5 degrees Celsius wide) and wind speed preference parameters (such as silent mode, natural wind mode, etc.) that meet the comfort requirements of 95% of the population.

[0054] The forward-looking strategy generation module performs multi-objective optimization in the rolling time domain based on the output of the dynamic dual model. Its logical flow is shown in the attached figure. Figure 4 As shown. This module has a built-in non-dominated sorting genetic algorithm solver with an elitist strategy. The optimization objective function is defined as: minimizing the predicted total energy consumption of the system over the next 2 hours. And maximizing the weighted average thermal comfort satisfaction of each zone The weighted sum. Optimization variables include the set temperature at the end of each partition. Air supply speed Air supply angle and the chilled water outlet temperature of the central cooling station With water pump frequency .

[0055] The constraints are strictly derived from the outputs of the dual models: all equipment operating parameters must be within the theoretically optimal range calculated by the energy efficiency model; and all environmental parameters for each zone must fall within the personalized comfort range output by the comfort model. Furthermore, the module reads the door and window status data from the building structure status sensing unit in real time. If a zone's doors and windows remain open for more than 5 minutes, the cooling load demand for that zone is forcibly reset to zero during that period, and its set temperature adjustment permission is disabled in the optimization variables.

[0056] The optimization process uses a 15-minute time step to generate control strategy sequences for the next 8 time steps (a total of 2 hours). The population size is set to 200, and the maximum number of iterations is 100. After non-dominated sorting and crowding calculation in each generation, elite individuals are retained for the next generation.

[0057] Finally, from the converged Pareto front, based on a preset energy efficiency-comfort weight preference (e.g., 7:3), the non-dominated solution closest to this preference is selected as the optimal policy sequence. It is worth noting that this module only issues control instructions for the first time step in the sequence (i.e., the next 0 to 15 minutes) to the execution layer; the remaining 7 steps are only used for internal state prediction and policy coherence assurance.

[0058] The adaptive execution and feedback module is responsible for instruction execution and closed-loop correction. The instruction distribution unit receives control parameters from the forward-looking strategy generation module, converts them into equipment control instructions conforming to the BACnet MS / TP or Modbus TCP protocol, and distributes them to each air conditioning terminal controller and the central chiller station PLC via the building control bus.

[0059] After the command is issued, the execution status monitoring unit continuously collects the actual response data of the equipment, including the actual set temperature achieved, fan speed, water valve opening, equipment power consumption, etc., and combines this with the actual temperature and humidity of the zone fed back by the environmental parameter sensing unit to form an execution feedback data packet. The time granularity of this data packet is 1 minute, and it contains a sequence of deviations between the expected value of the command and the actual value achieved.

[0060] The model calibration unit is crucial for closed-loop optimization. It compares the execution feedback data packets with the prediction data used during strategy generation, calculating the prediction deviation of key indicators. If the absolute value of the temperature deviation in a certain zone exceeds 0.5 degrees Celsius for three consecutive sampling periods, or the power consumption deviation of a certain device exceeds 10% for three consecutive periods, the model is considered inaccurate. In this case, the calibration unit sends a high-priority recalibration request to the dynamic dual-model collaborative construction module and injects the high-deviation data from the most recent hour into the model training set with a 10-fold weight, triggering rapid online retraining of the model. This mechanism enables the system to quickly restore decision accuracy in scenarios such as sudden changes in building usage patterns (e.g., large conferences), equipment performance degradation (e.g., heat exchanger fouling), or drastic changes in external climate.

[0061] To further enhance system security and robustness, this embodiment also integrates a digital twin simulation verification submodule. After the forward-looking strategy generation module outputs the control strategy sequence and before the instruction distribution unit executes it, this submodule first performs rapid simulation and deduction in a high-fidelity digital twin. This digital twin is built based on a building information model and accurately reproduces the building geometry, material thermal properties, equipment performance curves, and personnel activity patterns.

[0062] The simulation is performed in 1-minute increments, predicting the evolution of temperature, humidity, and carbon dioxide concentration in each zone, as well as the total system energy consumption, after the execution of the strategy sequence. If the simulation results show that the temperature in any zone will exceed the safety threshold (e.g., below 18 degrees Celsius or above 28 degrees Celsius), or the energy saving rate is lower than the preset benchmark (e.g., 5%), the strategy is deemed unqualified, and the system returns to the forward-looking strategy generation module for re-optimization. Commands are only allowed to be issued when all simulation results meet the standards. This mechanism effectively avoids control risks caused by model prediction biases.

[0063] In summary, this embodiment achieves precise, efficient, and comfortable operation of the air conditioning system in complex dynamic environments by constructing a five-layer technical architecture encompassing all-dimensional perception, deep fusion, dual-model collaboration, forward-looking optimization, and closed-loop adaptation. The system not only solves the energy efficiency problems caused by data silos and lag in traditional control, but also improves control precision to a level that meets individual comfort needs through physical information fusion modeling and federated personalized learning, while ensuring long-term operational reliability and safety.

[0064] Example 2 Building upon Example 1, this example further optimizes the data fusion algorithm for the personnel activity perception unit to address the perception challenges in high-density crowd scenarios. In large open-plan office areas or multi-functional halls, when the personnel density exceeds 0.3 people per square meter, traditional point cloud clustering and matching algorithms are prone to target adhesion and underestimation of counts. Therefore, this example introduces a deep learning-based semantic segmentation auxiliary mechanism.

[0065] Specifically, the raw thermal images from the infrared thermal imaging sensor are first fed into a lightweight convolutional neural network for semantic segmentation. This network, pre-trained on a dataset containing thousands of labeled thermal images, can accurately distinguish between human heat spots, device heat sources, and background thermal radiation. The segmented binary mask is used to guide subsequent density clustering, extracting heat source clusters only within the human mask region, significantly reducing interference from non-human heat sources. Simultaneously, the millimeter-wave radar point cloud is input into a graph neural network. This network constructs the point cloud as a graph structure, with nodes representing radar echo points and edge weights representing Euclidean distance and velocity similarity between points. Graph convolution operations enhance the internal consistency of the target, thereby more accurately separating adjacent individuals in high-density scenes.

[0066] In the data matching phase, this embodiment upgrades the Hungarian algorithm to a confidence-weighted multi-hypothesis tracking framework. Each heat source cluster and radar target cluster is assigned a confidence score, which comprehensively considers the cluster's size, shape regularity, motion continuity, and consistency with historical trajectories. The matching process no longer pursues a rigid one-to-one association but allows one heat source cluster to correspond to multiple radar targets (suitable for scenarios with many people closely gathered), and Kalman filtering is used to smooth the trajectory of the matching results. The final output personnel activity matrix not only includes the number and intensity of activities but also adds personnel gathering pattern labels (such as linear queues, circular meetings, and distributed offices), providing a more refined behavioral context for the comfort model.

[0067] Furthermore, this embodiment enhances the online learning mechanism of the dynamic dual-model collaborative construction module. A concept drift detection mechanism is introduced into the energy efficiency model construction unit. This mechanism continuously monitors the statistical characteristics of the model's prediction residuals, using cumulative sum control charts to detect significant changes in the residual mean or variance. Once concept drift is detected (indicating a substantial change in equipment performance or building thermal characteristics), the system immediately freezes the current model and initiates a batch retraining process with a larger historical window (7 days), rather than relying solely on incremental fine-tuning. This mechanism significantly improves the model's adaptability to long-term changing scenarios such as equipment aging or building renovations.

[0068] The comfort model building unit adds the function of identifying and adapting to special groups. By analyzing the spectral characteristics of people's micro-movement patterns, the system can initially determine whether there are special groups such as the elderly, children, or patients (whose micro-movement frequency is usually lower than that of ordinary people). Once identified, the comfort model will automatically switch to the corresponding special comfort template, which has a narrower temperature tolerance range (e.g., ±0.8 degrees Celsius) and a lower wind speed limit (e.g., 0.2 m / s), thereby providing a more humane environmental service.

[0069] In the forward-looking strategy generation module, this embodiment introduces electricity price signals as an external economic constraint. The system accesses real-time time-of-use electricity price data from the local power grid and adds an electricity cost term to the multi-objective optimization function. The optimization solver balances energy efficiency, comfort, and economy. For example, during peak electricity price periods, if there is sufficient comfort margin, the temperature setting is appropriately relaxed to reduce peak load; during off-peak electricity price periods, cooling is stored in advance to reserve cooling capacity. This function enables the system not only to optimize technical indicators but also to directly generate economic benefits.

[0070] The adaptive execution and feedback modules have also been enhanced. The model calibration unit can now distinguish the source of deviation: if the deviation mainly occurs on the equipment execution side (such as valve jamming), it sends an equipment maintenance warning to the building operation and maintenance platform; if the deviation originates from inaccurate environmental predictions (such as unforeseen strong sunlight), it only triggers fine-tuning of model parameters. This refined fault diagnosis capability upgrades the system from a simple control optimization tool to an intelligent decision-making hub for building energy management.

[0071] Through the above enhancements, this embodiment, while retaining all the advantages of embodiment 1, further improves the system's perception accuracy, long-term adaptability, humanistic care level, and economic value in complex scenarios, making it more suitable for places with extremely high requirements for air conditioning system performance, such as large commercial complexes, hospitals, and data centers.

[0072] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0073] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An air conditioning control terminal system based on data acquisition and analysis, characterized in that, include: The multi-source heterogeneous data acquisition module is used to collect dynamic data related to the operation of the air conditioning system and the building environment in real time. The multi-source heterogeneous data acquisition module includes an environmental parameter sensing unit, a personnel activity sensing unit, an equipment status monitoring unit, and a building structure status sensing unit. The data fusion and feature extraction module is used to perform spatiotemporal alignment, noise reduction and cleaning, and deep feature extraction on the raw data uploaded by the multi-source heterogeneous data acquisition module, and to encapsulate the extracted high-dimensional feature vectors into a feature data package in a unified format. A dynamic dual-model collaborative construction module is used to construct and continuously update a dynamic model of device energy efficiency and a dynamic model of user comfort in parallel based on historical data and real-time feature data. The dynamic dual-model collaborative construction module includes an energy efficiency model construction unit and a comfort model construction unit. The forward-looking strategy generation module is used to perform multi-objective optimization calculations based on the output of the dynamic dual-model collaborative construction module to generate a forward-looking air conditioning system control strategy sequence. The adaptive execution and feedback module is used to execute the control commands issued by the forward-looking strategy generation module and collect the actual response data of the system to form a closed-loop optimization.

2. The air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The personnel activity sensing unit adopts a fusion sensing technology that combines distributed infrared thermal imaging sensors with millimeter-wave radar sensors; The data fusion process of the personnel activity sensing unit is as follows: First, the original point cloud data collected by the two types of sensors are aligned with the coordinate system and the timestamp; then, a spatial clustering algorithm based on density clustering is used to identify heat source clusters from the infrared point cloud and moving and stationary target clusters from the radar point cloud; finally, the Hungarian matching algorithm is used to associate and match heat source clusters and target clusters within the same spatiotemporal range, eliminate interference caused by non-human heat sources, and comprehensively judge the activity status of the target to output a matrix of personnel number and activity intensity accurate to the partition grid.

3. The air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The energy efficiency model building unit adopts a neural network architecture based on physical constraints; The physical constraint layer of the neural network based on physical constraints is implemented as follows: During the forward propagation of the neural network, some intermediate variables output by the hidden layer are forcibly substituted into a set of preset refrigerant pressure-enthalpy relationship and heat exchanger efficiency equation for calculation. The calculation results are used as constraint loss terms and network prediction loss terms to participate in backpropagation and weight update, thereby embedding the constraints of the first and second laws of thermodynamics into the data-driven model.

4. The air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The comfort model building unit adopts a personalized transfer learning framework; The personalized adaptation process of the personalized transfer learning framework adopts a federated learning framework. The terminal devices in each partition use locally collected real-time data to fine-tune the local comfort model copy. Only the updated gradient of the model weights is encrypted and uploaded to the central processing unit for secure aggregation, generating a global model update, which is then distributed to each partition.

5. An air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The forward-looking strategy generation module has a built-in multi-objective optimization solver; The multi-objective optimization solver employs a non-dominated sorting genetic algorithm with an elite strategy. This algorithm treats energy efficiency and comfort as two independent fitness functions, iteratively evolves a control strategy population through selection, crossover, and mutation operations, and finally selects the solution closest to the preset weight preference from the non-dominated solution set as the optimal strategy sequence output.

6. The air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The forward-looking strategy generation module also incorporates door and window status data from the multi-source heterogeneous data acquisition module as a hard constraint; When it is detected that the doors and windows of a certain zone have been open for more than 5 minutes, the multi-objective optimization solver will temporarily block the allocation of cooling capacity to that zone in subsequent strategies until the doors and windows are closed.

7. The air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The adaptive execution and feedback module includes an instruction distribution unit, an execution status monitoring unit, and a model correction unit; The model calibration unit is used to compare the execution feedback data packet with the prediction data on which the prospective strategy generation module is based when generating the strategy, and calculate the prediction deviation. If key indicators such as partition temperature deviation continue to exceed 0.5 degrees Celsius or device power consumption deviation continue to exceed 10%, the model calibration unit will trigger the model recalibration process of the dynamic dual-model collaborative construction module.

8. The air conditioning control terminal system based on data acquisition and analysis according to claim 1, characterized in that, The system also integrates a digital twin simulation verification submodule; The digital twin simulation verification submodule is used to perform rapid simulation and deduction in a high-fidelity digital twin constructed based on the building information model and the actual parameters of the equipment before the control strategy sequence is sent to the real equipment, and to predict the environmental and energy consumption results after the strategy is executed; the instruction is only allowed to be sent when the simulation result meets the safety threshold and expected benefits, otherwise the strategy will be returned to the forward-looking strategy generation module for re-optimization.

9. An air conditioning control terminal system based on data acquisition and analysis according to claim 2, characterized in that, The activity intensity level output by the personnel activity sensing unit is divided into three levels: sitting, light activity, and moderate activity. The data fusion and feature extraction module extracts the activity heat load intensity index from personnel activity data. Its calculation formula is the sum of the products of the activity level weights and the activity duration ratios of all personnel in the partition.

10. An air conditioning control terminal system based on data acquisition and analysis according to claim 3, characterized in that, The energy efficiency model building unit uses an online learning algorithm to dynamically adjust the network weights by using real-time collected device power consumption data as a supervision signal, thereby outputting the predicted instantaneous energy efficiency ratio of the air conditioning system under the current operating conditions, and calculating the theoretical optimal operating parameter range of each terminal device.

Images