Air Conditioning Control Method and System Based on Physical Modeling and Big Data Prediction

Abstract

This application provides an air conditioning control method and system based on physical modeling combined with big data prediction. The method includes: triggering a prediction task through a timed task scheduler; acquiring environmental temperature and humidity parameters and fusing them with production load, weather forecast, and equipment health data to extract feature vectors containing thermodynamically derived features; inputting the feature vectors into an air conditioning start-up prediction model trained based on an improved random forest to obtain the prediction duration for each air conditioning unit; calculating the target time based on the prediction duration and automatically issuing start-up commands. This application, through the deep integration of physical modeling and big data prediction, achieves accurate prediction and automatic control of air conditioning start-up time, significantly reducing the energy consumption of redundant operation of the air conditioning system while ensuring that the production environment meets standards.

Description

Technical Field

[0001] This document relates to the field of air conditioning control technology, and in particular to an air conditioning control method and system based on physical modeling combined with big data prediction. Background Technology

[0002] As the global manufacturing industry accelerates its transformation towards digitalization, intelligentization, and green, low-carbon practices, industrial air conditioning systems have evolved from basic environmental control equipment into critical infrastructure that ensures production processes and influences core product quality and production costs. In many industries, such as precision manufacturing, biopharmaceuticals, food processing, and high-end electronics assembly, production processes place extremely stringent and stable requirements on the temperature and humidity parameters of the workshop environment. To ensure that the workshop environmental parameters are precisely stable within the process window when production starts each day, the air conditioning system must be turned on and finely adjusted in advance.

[0003] Currently, the timing of industrial air conditioning system startups relies heavily on the personal experience of on-site operators or engineers. This experience-based, estimation-based startup model has revealed multiple challenges in long-term operation: To avoid production stoppages due to substandard environmental conditions, operators often tend to set excessively long safety buffer times, resulting in a large amount of unnecessary redundant operation and significant energy consumption; when facing atypical weather conditions such as late spring cold snaps or lingering summer heat waves, or when production plans or equipment status change abnormally, experience-based decisions may fail, leading to delayed startups that directly affect material properties, process precision, and even the performance of the final product, potentially causing batch quality issues and order delays; the stable and efficient operation of the system is deeply tied to the experience of a few senior employees, which not only poses a single point of failure risk but also makes it difficult to solidify, pass on, and optimize valuable operational knowledge.

[0004] Therefore, there is an urgent need for a technical solution that can deeply integrate physical mechanisms with real-time data, adaptively learn system characteristics and environmental changes, so as to achieve intelligent, accurate and robust prediction of air conditioner start-up and shutdown times. Summary of the Invention

[0005] According to embodiments of the present invention, an air conditioning control method and system based on physical modeling and big data prediction are provided, aiming to solve the above-mentioned problems.

[0006] According to an embodiment of the present invention, an air conditioning control method based on physical modeling combined with big data prediction is provided, comprising: S1. Trigger the prediction task at a preset interval using a timed task scheduler; S2. Obtain the ambient temperature and humidity parameters monitored in real time by the air conditioning unit, and integrate the factory production load data, outdoor weather forecast data and air conditioning equipment health status data as the supporting operating parameters of the air conditioning system. Extract the corresponding feature vector based on the ambient temperature and humidity parameters and the supporting operating parameters of the air conditioning system. S3. Input the feature vector into a pre-set air conditioner start-up prediction model trained on random forest, and obtain the prediction duration of the air conditioner equipment in each air conditioner unit based on the air conditioner start-up prediction model; S4. Calculate the target time based on the predicted duration, and automatically issue an air conditioner start command when the current time reaches the target time. During the operation of the air conditioner, collect dynamic temperature and humidity change data in real time. If the deviation from the target trend is detected, dynamically adjust the air conditioner operating parameters.

[0007] According to an embodiment of the present invention, an air conditioning control system based on physical modeling and big data prediction is provided, comprising: The sensor layer includes outdoor temperature and humidity sensors, fresh air temperature and humidity sensors, indoor temperature and humidity sensors, production load sensors, equipment status sensors, meteorological data acquisition modules, and production plan docking modules. It is used to monitor temperature and humidity data at different locations in real time, and also to collect production load data, air conditioning equipment health status data, outdoor weather forecast data, and factory production plan data. The data acquisition layer is used to acquire data and convert protocols based on the sensor layer. It supports high-speed industrial data transmission and has edge data preprocessing capabilities, which can complete preliminary data cleaning and feature extraction at the edge. The data preprocessing and storage layer is used to clean, interpolate, extract features and store raw data. It adopts an edge-cloud collaborative storage architecture, with real-time high-frequency data stored at the edge and historical massive data stored in the cloud. It also supports bidirectional data synchronization and breakpoint resume. The algorithm service layer includes a random forest-based air conditioner start-up prediction model, a REST API service, a model inference acceleration module, a multi-source data fusion module, and a dynamic adjustment module for operating parameters. It is used to predict air conditioner start-up time and provide standard calling interfaces. It is also used to realize the fusion processing of multi-source data, hardware acceleration of model inference, and dynamic adjustment of air conditioner operating parameters. The upper-level system includes a visualization platform, management system, production planning linkage module, anomaly alarm module, and energy consumption statistical analysis module. It is used to display prediction results, operating status, and historical data, provide system configuration and operation control functions, realize real-time linkage with production plan, intelligent alarm for air conditioning operation anomalies, and refined statistical analysis of air conditioning system energy consumption.

[0008] By employing the embodiments of this invention, feature vectors are constructed by integrating multi-dimensional data. An improved random forest model is used to accurately predict the start-up time of air conditioners, and an adaptive reserved time mechanism eliminates redundant operating time caused by human experience, minimizing ineffective energy consumption of air conditioning units and reducing energy operating costs in industrial production. By monitoring temperature and humidity changes in real time and dynamically adjusting air conditioning operating parameters, it ensures that the workshop temperature and humidity are accurately and stably within the process requirements during production startup, avoiding the impact of substandard environmental conditions on material properties and process precision, effectively preventing batch quality problems and production order delays, and ensuring product quality uniformity. Thermodynamic physical mechanisms are incorporated into feature extraction, introducing physical mechanism feature weight constraints into the prediction model to address the insufficient generalization ability of purely data-driven models. Simultaneously, hierarchical sampling and weighted integration designs make the model more adaptable to complex industrial scenarios such as seasonal changes and production load fluctuations, resulting in more reliable prediction results. This approach transforms the tacit experience of seasoned operators into quantifiable and iterative explicit models and data assets, replacing the traditional manual estimation-based start-up mode. This eliminates the risk of single points of failure and, through a fully automated process design encompassing prediction, command issuance, and model self-monitoring and retraining, enhances the automation and lean management of the factory's air conditioning system control. The adaptive reserve time dynamically adjusts based on prediction errors, and the model performance monitoring and retraining mechanism promptly identifies model drift and performs iterative optimization. This allows the system to adapt to changes in operating conditions such as equipment aging, process adjustments, and sudden weather changes, ensuring the long-term accuracy and effectiveness of the air conditioning control strategy. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of an air conditioning control method based on physical modeling and big data prediction according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an air conditioning control system based on physical modeling and big data prediction according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0012] Method Implementation Examples According to embodiments of the present invention, an air conditioning control method based on physical modeling combined with big data prediction is provided. Figure 1 This is a flowchart of an air conditioning control method based on physical modeling and big data prediction according to an embodiment of the present invention. Figure 1 As shown, the air conditioning control method based on physical modeling and big data prediction according to an embodiment of the present invention specifically includes: S1. A prediction task is triggered at a preset interval by a timed task scheduler. The prediction task is triggered at a preset interval by a timed task scheduler. The timed task scheduler is deployed in the industrial control backend. The preset interval can be flexibly configured according to the factory production rhythm and the frequency of weather changes. The trigger command is synchronously pushed to the data acquisition module and the algorithm service module to prepare for subsequent data acquisition and model prediction.

[0013] S2. Multi-source data acquisition and feature vector extraction: First, the ambient temperature and humidity parameters monitored in real time by the air conditioning unit are acquired. Simultaneously, factory production load data, outdoor weather forecast data, and air conditioning equipment health status data are integrated, and these data are combined as the operating parameters for the air conditioning system. Based on the ambient temperature and humidity parameters and the operating parameters of the air conditioning system, the corresponding feature vectors are extracted through a multi-stage feature processing flow, specifically: (1) Basic feature extraction: Based on the operating mechanism of the air conditioning system, basic features are extracted from the original monitoring data containing environmental temperature and humidity parameters. The environmental temperature and humidity parameters include core parameters such as indoor temperature and humidity, outdoor temperature and humidity, fresh air temperature and humidity, mixed air temperature and humidity, and supply air temperature and humidity. (2) Generation of derived features: Based on the thermodynamic principle, the environmental temperature and humidity parameters are calculated to obtain the moisture content, enthalpy value and wet bulb temperature as physical derived features. At the same time, the unit area cooling / heating load derived features are calculated in combination with the production load data, and the equipment operating efficiency derived features are calculated in combination with the equipment health status data. The equipment health status data includes the air conditioning fan operating efficiency, chilled water heat transfer coefficient, steam humidification module fault-related parameters, etc. (3) Feature selection and vector formation: A filtering feature selection method is adopted, based on correlation analysis, to select a subset of features that have a significant impact on the target variable of start-up time. At the same time, the nonlinear correlation between the features and the target variable is verified by mutual information test. After removing features with low correlation, feature vectors are formed. The final feature vectors include basic state features, target gap features, time cycle features, thermodynamic derived features, production load features, and equipment health features; The basic state characteristics include the initial outdoor temperature, initial indoor humidity, and initial chilled water supply temperature. The initial outdoor temperature and initial indoor humidity are directly extracted from the ambient temperature and humidity parameters. The target gap characteristics include the initial indoor temperature difference ΔT, the initial indoor humidity difference ΔH, and the cooling / heating load difference ΔQ. The initial indoor temperature difference ΔT and the initial indoor humidity difference ΔH are calculated from the indoor environmental temperature and humidity parameters and the target temperature and humidity values ​​required by the process. The cooling / heating load difference ΔQ is calculated from the actual production load and the load required by the process. The time cycle characteristics include start-up time, start-up time attribute, production shift attribute, holiday attribute, and weather and seasonal attribute; Thermodynamic derivative characteristics include initial moisture content, initial air enthalpy, and wet-bulb temperature, which are calculated from ambient temperature and humidity parameters combined with standard thermodynamic formulas; production load characteristics include real-time production capacity and the number of workshop equipment in operation.

[0014] S3. Model Prediction to Obtain the Target Compliance Time: Input the feature vectors constructed above into a pre-set air conditioner start-up prediction model trained based on an improved random forest, and obtain the start-up target compliance prediction time for each air conditioner unit based on the model. The air conditioner start-up prediction model is a regression model built on an improved random forest algorithm. This improved random forest consists of multiple decision trees, and each decision tree is given a feature weight constraint based on physical mechanisms. Thermodynamically derived features and production load features are given higher initial weights, making the model prediction more consistent with the physical laws of industrial air conditioner operation. Each decision tree is trained based on a subset of samples obtained by Bootstrap sampling, and a hierarchical sampling method is used to ensure a balanced sample ratio for each season and production load range, avoiding model prediction bias caused by sample bias.

[0015] When splitting at each node of each decision tree, the optimal splitting feature is selected from a randomly selected subset of features. The splitting criterion adopts a mean squared error loss function that incorporates physical errors to improve the rationality of the splitting results.

[0016] After the feature vector is input into the model, each decision tree independently outputs a predicted value. The weighted arithmetic mean of the predicted values ​​of all decision trees is taken as the final output of the model, which is the predicted time for the system to meet the target. The weight of the decision tree is dynamically allocated according to its prediction accuracy on the validation set. The decision tree with higher prediction accuracy has a larger weight.

[0017] This improved random forest algorithm reduces the risk of overfitting through a dual mechanism of sample randomness and feature randomness, eliminates the prediction bias of a single decision tree through multi-tree ensemble averaging, and improves the interpretability and robustness of the model in complex industrial scenarios through physical mechanism constraints.

[0018] The training process of the air conditioner start-up prediction model includes an offline training process and an online inference process, wherein the offline training process is as follows: Historical operation records of air conditioners are queried from storage media such as the factory's Historian time-series database. The historical operation records are cleaned and labeled, and each complete start-up-run-achieve-stop process is automatically identified. The features X at the start-up time and the actual time Y from start-up to the target point are extracted. A random forest model is trained using K-fold cross-validation. After the model training is completed, the model performance is evaluated using indicators such as mean absolute error and coefficient of determination. Once the performance meets the standards, the model is serialized and deployed to the production server.

[0019] The online inference process is as follows: after the prediction task is triggered by the timed task scheduler, the multi-source data collected by the sensor is acquired in real time and a feature vector that meets the model input requirements is constructed. The trained and serialized model is loaded to predict the feature vector and quickly output the start-up time prediction for each air conditioning device.

[0020] The specific process of data cleaning and labeling is as follows: Automatically identify each complete power-on-run-standard-shutdown process, and define the time point when the system first enters the target temperature and humidity range and maintains it stably for a preset time as the standard point. The preset time is dynamically configured according to the precision requirements of the production process. For scenarios with extremely high temperature and humidity requirements, such as precision manufacturing, a short-duration threshold is configured, while for ordinary industrial scenarios, a long-duration threshold is configured, effectively filtering out false standard points caused by system oscillations. After extracting the feature X at the boot time, the feature X is normalized and standardized to eliminate dimensional differences and ensure the accuracy of model training. The actual time taken from the boot time to the target point is calculated as the target Y. If there are outliers in the actual time, the isolated forest algorithm is used to identify and remove them. At the same time, interpolation is used to supplement effective samples to improve the quality of the training dataset.

[0021] S4. Calculation of Start-up Target Time and Automatic Issuance of Command: Based on the start-up target prediction time output by the model, the air conditioner start-up target time is calculated. When the current time reaches the target time, the upper control system automatically issues a start-up command to the air conditioning unit. The specific calculation process is as follows: (1) Obtain the time point T_target that the workshop temperature and humidity must meet as specified in the production plan, and dynamically update the time point T_target in conjunction with the factory production plan adjustment information to ensure synchronization with the production rhythm; (2) Obtain the start-up compliance prediction time T_pred output by the air conditioner start-up prediction model, and correct the prediction time T_pred according to the health status of the air conditioner equipment. If the equipment is in a state of aging or deterioration, multiply it by the corresponding correction coefficient according to the degree of deterioration so that the prediction time is more in line with the actual operating capacity of the equipment. (3) Obtain the adaptive reservation time T_adaptive based on the dynamic calculation of historical prediction error. The adaptive reservation time T_adaptive is dynamically determined according to the statistical distribution of historical prediction error within the sliding window. The length of the sliding window is adaptively adjusted according to the frequency of meteorological changes and the degree of fluctuation of production load. The greater the meteorological changes or the fluctuation of production load, the shorter the length of the sliding window, ensuring a rapid response to error changes. (4) Calculate the target start time T_start of the air conditioner according to the formula T_start = T_target - T_pred - T_adaptive.

[0022] The specific steps for determining the adaptive reserve time T_adaptive are as follows: Continuously record the prediction error between the actual air conditioning compliance time and the model prediction time for each instance, forming an error time series. Smooth the error time series to eliminate abnormal errors caused by unforeseen factors such as sudden equipment failures or extreme weather, ensuring the effectiveness of error statistics. Use a sliding window method to statistically analyze the mean μ and standard deviation σ of the prediction error within the window, and simultaneously calculate the skewness coefficient Sk. Calculate the adaptive reserve time according to the formula T_adaptive = μ + k × σ + α × Sk, where k is a preset safety factor, and α is a skewness correction coefficient configured based on the process compliance tolerance rate. When the volatility of the prediction error increases, the standard deviation σ increases, automatically extending the adaptive reserve time. When model performance improves, causing the prediction error to converge, the adaptive reserve time automatically shortens. When the error distribution is left-skewed, the reserve time is increased through the skewness correction coefficient. This minimizes redundant air conditioning operation time while ensuring the reliability of production environment compliance.

[0023] Furthermore, this embodiment of the invention also includes dynamic parameter adjustment during air conditioning operation. After the air conditioning unit starts operating according to the start-up command, it collects real-time data on the dynamic changes in temperature and humidity in the workshop, compares the temperature and humidity change trends with the preset target trends of the model, and if the temperature and humidity changes are detected to deviate from the target trends, the dynamic parameter adjustment module of the algorithm service layer sends parameter adjustment commands to the air conditioning unit to dynamically adjust the chilled water supply temperature, fan frequency, steam supply pressure and other operating parameters of the air conditioner to ensure that the temperature and humidity in the workshop can accurately and stably reach the target range required by the production process.

[0024] In this embodiment, the air conditioning start-up prediction model also has a full lifecycle operation and maintenance optimization mechanism. During the long-term operation of the model, the deviation between the model's predicted value and the actual time of achievement is continuously recorded, and performance indicators such as mean absolute error and coefficient of determination are calculated periodically. When the performance indicators exceed the preset threshold, the system automatically triggers a drift alarm. At the same time, it supports one-click retraining of the model, using the latest air conditioning operation data, production load data, and meteorological data to retrain and optimize the model. After generating a new version of the model, it is directly replaced and deployed, forming an MLOps closed loop. This allows the model to adapt to changes in operating conditions such as equipment aging, process adjustments, and seasonal changes, ensuring the long-term stability of prediction accuracy.

[0025] This embodiment presents an air conditioning control method based on physical modeling and big data prediction. By integrating multi-source industrial data and constructing feature vectors using thermodynamic physical mechanisms, it employs an improved random forest model to accurately predict the start-up time of air conditioning units. Combined with an adaptive reserved time mechanism and dynamic parameter adjustment during operation, this method replaces the traditional air conditioning start-up decision-making mode based on human experience. This effectively reduces the redundant operating time of air conditioning units, significantly lowers energy consumption in industrial production, and ensures that the workshop temperature and humidity are accurately and stably up to standard during production startup, avoiding product quality problems and production delays caused by inconsistent environmental parameters. At the same time, it transforms the implicit experience of industrial air conditioning control into a quantifiable and iterative explicit data model, reducing reliance on experienced operators and improving the automation, intelligence, and lean management level of factory air conditioning system control, adapting to the complex operating conditions of various industrial production processes.

[0026] System Implementation Examples This invention also discloses an air conditioning control system based on physical modeling combined with big data prediction, used to execute the aforementioned air conditioning control method based on physical modeling combined with big data prediction. The system adopts a five-layer architecture design, consisting of a sensor layer, a data acquisition layer, a data preprocessing and storage layer, an algorithm service layer, and a host system layer from top to bottom. Communication connections and bidirectional data flow are established between each layer. At the same time, a model performance monitoring and retraining mechanism is configured to form a complete MLOps closed loop, which is suitable for the intelligent air conditioning control needs of industrial scenarios such as precision manufacturing, biopharmaceuticals, food processing, and high-end electronic assembly. It can effectively solve the problems of traditional industrial air conditioning control systems, such as reliance on manual labor, high energy consumption, and low prediction accuracy. The structure, function, and core mechanism of each layer are described in detail below.

[0027] Figure 2 This is a schematic diagram of an air conditioning control system based on physical modeling and big data prediction according to an embodiment of the present invention. Figure 2 As shown, the air conditioning control system based on physical modeling and big data prediction according to an embodiment of the present invention includes: The sensor layer includes outdoor temperature and humidity sensors, fresh air temperature and humidity sensors, indoor temperature and humidity sensors, production load sensors, equipment status sensors, meteorological data acquisition modules, and production plan docking modules. It is used to monitor temperature and humidity data at different locations in real time, and also to collect production load data, air conditioning equipment health status data, outdoor weather forecast data, and factory production plan data. Specifically, the sensor layer is the foundational data source layer of the entire system, serving as a prerequisite for physical modeling and data prediction. It employs fully industrial-grade sensing equipment to ensure the accuracy, stability, and interference resistance of data acquisition. This layer not only includes conventional temperature and humidity sensors but also integrates production load, equipment status-related sensing modules, and external data interface modules, specifically including: Outdoor temperature and humidity sensor, fresh air temperature and humidity sensor, indoor temperature and humidity sensor, production load sensor, equipment status sensor, meteorological data acquisition module and production planning docking module.

[0028] Temperature and humidity sensors: Deployed at key locations such as the outdoor end of the air conditioning unit, the fresh air inlet, the indoor workshop, the mixing air section, and the supply air section, they collect environmental temperature and humidity parameters such as outdoor temperature and humidity, fresh air temperature and humidity, indoor temperature and humidity, mixing air temperature and humidity, and supply air temperature and humidity in real time, providing core physical data for feature extraction and model prediction. Production load sensors: Deployed in workshop production lines, power equipment, and other locations to collect real-time production load data such as production capacity and the number of workshop equipment in operation, reflecting the actual cold / heat load demand of workshop production; Equipment status sensors: Deployed in core components of air conditioning units such as fans, chilled water heat exchange modules, and steam humidification modules, these sensors collect real-time health status data of air conditioning equipment, such as fan operating efficiency, chilled water heat exchange coefficient, and steam humidification module fault-related parameters, to characterize the actual operating capacity of the air conditioning unit. Meteorological data acquisition module: This is an integrated hardware and software module that supports communication with external meteorological platforms, acquires outdoor weather forecast data in real time (including temperature, humidity, wind speed, seasonal meteorological characteristics, etc.), and captures dynamic changes in the external environment. Production Planning Integration Module: This is an industrial communication integration module that connects with the Factory Production Management System (MES) to acquire real-time production planning data and adjustment information, ensuring that air conditioning control is synchronized with the production rhythm. All data acquisition devices and integration modules in the sensor layer use industrial standard interfaces, allowing direct communication with the lower data acquisition layer for real-time data upload.

[0029] The data acquisition layer is used for data acquisition and protocol conversion based on the sensor layer. It supports high-speed industrial data transmission and has edge data preprocessing capabilities, enabling preliminary data cleaning and feature extraction at the edge. The data acquisition layer is the core layer for data transmission and protocol conversion connecting the sensor layer and the data preprocessing and storage layer. It undertakes the unified acquisition, protocol adaptation, and standardized transmission of multi-source data, solving the problem of incompatible communication protocols between different devices and modules in industrial scenarios. It also possesses industrial-grade high-speed data transmission capabilities and edge preliminary preprocessing capabilities. Specific functions are as follows: Multi-protocol data acquisition: Supports dual-protocol data acquisition using OPC UA industrial protocol and REST API interface. Real-time data from industrial-grade sensing devices and factory PLC / DCS systems can be directly acquired through the OPC UA industrial protocol. Third-party data sources for meteorological data acquisition modules and production planning modules can be accessed through the REST API interface to achieve unified aggregation of multi-source data. Protocol conversion: Standardize and unify the collected data with different protocols and formats, convert various heterogeneous data into a unified data format that the system can recognize, and ensure the consistency of subsequent data processing; High-speed industrial data transmission: Supports high-speed industrial data transmission methods such as industrial Ethernet and 5G to meet the transmission requirements of high-frequency real-time data at the sensor layer, ensuring low latency and high reliability of data transmission; Edge-end data preprocessing: It possesses lightweight edge-end data preprocessing capabilities, enabling preliminary data cleaning (such as removing obviously invalid and duplicate values) and basic feature extraction at the edge, reducing the computational burden on subsequent data preprocessing and storage layers and improving overall system processing efficiency. Standardized data processed by the data acquisition layer will be transmitted in real-time to the data preprocessing and storage layer for in-depth processing and persistent storage.

[0030] The data preprocessing and storage layer is used to clean, interpolate, extract features and store raw data. It adopts an edge-cloud collaborative storage architecture, with real-time high-frequency data stored at the edge and historical massive data stored in the cloud. It also supports bidirectional data synchronization and breakpoint resume. The data preprocessing and storage layer is the core layer for data processing and storage in the system. It receives standardized data from the data acquisition layer, performs deep cleaning, interpolation, feature extraction, and storage management. It adopts an edge-cloud collaborative storage architecture to balance fast access to real-time data and massive storage of historical data, while ensuring data integrity and traceability. The specific functional modules and designs are as follows: The data cleaning module performs deep cleaning on the preliminary processed data transmitted from the data acquisition layer, including outlier filtering based on business rules, outlier removal based on statistical methods, unified unit conversion, and data type validation. It also deduplicates duplicate data with the same timestamp to ensure data quality. Data interpolation module: To address data loss issues caused by sensor disconnection or acquisition failure, it employs multiple interpolation methods such as linear interpolation, forward imputation, and spline interpolation to complete the data. The interpolation method is adaptively selected based on the data type and variation characteristics to ensure the continuity and integrity of time series data. Feature Engineering Module: According to the feature extraction rules in the method embodiment of the present invention, the feature engineering process of the whole process is completed, including basic feature extraction based on the operation mechanism of the air conditioning system, physical derivative feature generation based on thermodynamic principles, derivative feature calculation combined with production load and equipment status, and feature subset screening through filter feature selection and mutual information verification, and finally generating feature vectors that meet the model input requirements, providing high-quality feature data for the algorithm service layer; Edge cloud collaborative storage module: It adopts a time-series database as the core storage medium, specifically the Historian database. This database is designed for industrial time-series data and features efficient data compression, fast time range query, and multi-dimensional data aggregation calculation, perfectly adapting to the time-series characteristics of air conditioning operation data. At the same time, it adopts an edge cloud collaborative architecture to store real-time high-frequency data at the edge, ensuring fast data access for real-time inference of the algorithm service layer, and storing massive historical data in the cloud to achieve long-term persistent data storage. It also supports bidirectional data synchronization and breakpoint resume between the edge and the cloud to avoid data loss due to network interruption.

[0031] The algorithm service layer includes a random forest-based air conditioner start-up prediction model, a REST API service, a model inference acceleration module, a multi-source data fusion module, and a dynamic adjustment module for operating parameters. It is used to predict air conditioner start-up time and provide standard calling interfaces. It is also used to realize the fusion processing of multi-source data, hardware acceleration of model inference, and dynamic adjustment of air conditioner operating parameters. The algorithm service layer is the core intelligent analysis layer of the entire system. It is crucial for accurately predicting air conditioner operating time and dynamically adjusting operating parameters. It integrates the core prediction model, data fusion module, inference acceleration module, communication interface module, and operating parameter adjustment module of this invention. All functional modules adopt a modular design, supporting flexible updates and replacements. Specifically, it includes an air conditioner operating prediction model based on an improved random forest, a REST API service, a model inference acceleration module, a multi-source data fusion module, a dynamic operating parameter adjustment module, and an OPC UA client. The functions of each module are as follows: An improved random forest-based air conditioner start-up prediction model is the core module of the algorithm service layer. It adopts the improved random forest algorithm constructed by the method embodiment of this invention to construct a regression model. It introduces feature weight constraints based on physical mechanisms, assigns higher initial weights to thermodynamic derived features and production load features, and achieves accurate prediction of the start-up time of air conditioners by designing Bootstrap hierarchical sampling, combining the mean square error splitting criterion of physical errors, and decision tree weighted integration. At the same time, it supports batch feature vector input and parallel prediction of multiple devices, and adapts to the simultaneous control requirements of multiple air conditioning units. Multi-source data fusion module: It fuses feature vectors transmitted from data preprocessing and storage layer, as well as real-time supplemented production, meteorological, and equipment status data, to ensure the comprehensiveness and real-time nature of model input data, making model predictions more consistent with actual working conditions; Model inference acceleration module: It adopts dedicated hardware acceleration and model quantization technology to optimize the model inference process, significantly reduce model inference latency, meet the real-time requirements of industrial scenarios, and ensure that prediction results can be output quickly and support the subsequent control command issuance. REST API Service: The air conditioner start-up prediction function is encapsulated into a standard REST API interface, which can be called by the upper system layer and other external industrial systems in the factory to realize the standardized service output of the prediction function; OPC UA Client: It communicates with the factory OPC UA server. On the one hand, it subscribes to real-time sensor data from the factory OPC UA server to supplement the real-time data of the model prediction. On the other hand, it writes the predicted time for air conditioning to meet the standard output by the model back to the preset data node of the factory OPC UA server for the factory SCADA or HMI system to read, so as to achieve seamless connection between the algorithm service layer and the factory OT network. Dynamic adjustment module for operating parameters: It communicates with the control terminal of the air conditioning unit and receives real-time dynamic change data of temperature and humidity in the workshop during the operation of the air conditioning. It compares the data with the target trend preset by the model. If the temperature and humidity changes are detected to deviate from the target trend, it immediately generates and issues an operating parameter adjustment command to dynamically adjust parameters such as chilled water supply temperature, fan frequency, and steam supply pressure of the air conditioning to ensure that the temperature and humidity in the workshop are accurately and stably up to standard.

[0032] The upper-level system includes a visualization platform, management system, production planning linkage module, anomaly alarm module, and energy consumption statistical analysis module. It is used to display prediction results, operating status, and historical data, provide system configuration and operation control functions, realize real-time linkage with production plan, intelligent alarm for air conditioning operation anomalies, and refined statistical analysis of air conditioning system energy consumption.

[0033] The upper-level system layer is the human-machine interaction and control command issuance layer of the entire system. It serves as a bridge connecting the system's intelligent analysis with the actual production control of the factory. It integrates functions such as visualization, task scheduling, target time calculation, command issuance, production linkage, anomaly alarms, energy consumption analysis, and system management. It also features a strict security protection mechanism. Specifically, it includes a visualization platform, management system, timed task scheduling module, target time calculation module, command issuance module, production planning linkage module, anomaly alarm module, and energy consumption statistical analysis module. The functions of each module are as follows: The scheduled task module works in conjunction with the algorithm service layer to trigger prediction tasks at preset intervals, providing timed trigger signals for the intelligent prediction of the entire system. The preset intervals can be flexibly configured according to the factory's production rhythm. Target time calculation module: According to the start-up target time calculation rules in the method embodiment of the present invention, combined with the production plan target time point, model prediction duration, equipment health status correction coefficient and adaptive reserved time, the air conditioner start-up target time point T_start is accurately calculated to provide a time basis for instruction issuance; Command issuance module: It communicates with the control terminal of the air conditioning unit and the factory OT network. When the current time reaches the calculated start-up target time, it automatically issues a start-up command to the air conditioning unit. At the same time, it receives parameter adjustment commands from the dynamic adjustment module of operating parameters and forwards them to the air conditioning unit to realize the automation of air conditioning control. Visualization platform: Provides factory operators with a visual human-machine interface to display the system's operating status, the predicted start-up performance of each air conditioning unit, real-time temperature and humidity data in the workshop, model performance indicators (mean absolute error, coefficient of determination, etc.), historical data analysis results of air conditioning operation, and energy consumption statistics. It transforms implicit model predictions and data calculations into explicit visual information, making it convenient for operators to monitor in real time. Production planning linkage module: It links with the factory production management system (MES) in real time, automatically obtains the temperature and humidity target time points specified in the production plan, and dynamically updates the target time points according to the production plan adjustment information to ensure that the air conditioning control is highly matched with the production rhythm; Anomaly Alarm Module: Real-time monitoring of the operating status of each level of the system, the health status of the air conditioning unit, and the error status of the model prediction. If equipment failure, system anomaly, model drift, or temperature and humidity deviation from the target trend is detected, alarms will be issued immediately through sound and light, SMS, system pop-ups, etc., to remind operators to handle the situation in a timely manner. Energy consumption statistics and analysis module: performs detailed statistics and multi-dimensional analysis on the operating time and energy consumption data of air conditioning units, combines the optimization of redundant operating time predicted by the model, quantitatively displays the energy-saving effect of the system, and provides data support for the factory's energy management; Management System: Provides basic system configuration and access control functions, including preset interval time configuration, model parameter configuration, target threshold configuration, user hierarchical access control, operation log recording, etc. It also configures hierarchical access control mechanism and data-level security protection mechanism, assigns different operation permissions to different operators, encrypts and protects core system data and control commands, prevents unauthorized access and data tampering, and ensures the security of system operation. REST API Interface Integration Module: Communicates with the REST API service of the algorithm service layer to enable the calling of model prediction functions, and provides data integration interfaces for other management systems in the factory, realizing the openness and scalability of the system.

[0034] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An air conditioning control method based on physical modeling combined with big data prediction, characterized in that... include: S1. Trigger the prediction task at a preset interval using a timed task scheduler; S2. Obtain the ambient temperature and humidity parameters monitored in real time by the air conditioning unit, and integrate the factory production load data, outdoor weather forecast data and air conditioning equipment health status data as the supporting operating parameters of the air conditioning system. Extract the corresponding feature vector based on the ambient temperature and humidity parameters and the supporting operating parameters of the air conditioning system. S3. Input the feature vector into a pre-set air conditioner start-up prediction model trained on random forest, and obtain the prediction duration of the air conditioner equipment in each air conditioner unit based on the air conditioner start-up prediction model; S4. Calculate the target time based on the predicted duration, and automatically issue an air conditioner start command when the current time reaches the target time. During the operation of the air conditioner, collect dynamic temperature and humidity change data in real time. If the deviation from the target trend is detected, dynamically adjust the air conditioner operating parameters.

2. The method according to claim 1, characterized in that, The extraction of the corresponding feature vector specifically includes: Based on the operating mechanism of the air conditioning system, basic features are extracted from the raw monitoring data containing environmental temperature and humidity parameters; Based on thermodynamic principles, the ambient temperature and humidity parameters are calculated to obtain moisture content, enthalpy, and wet-bulb temperature as physical derivative characteristics. At the same time, the unit area cold / heat load derivative characteristics are calculated by combining production load data, and the equipment operating efficiency derivative characteristics are calculated by combining equipment health status data. A filtering feature selection method is adopted to screen the feature subset that has a significant impact on the target variable based on correlation analysis, and to verify the nonlinear correlation between the features and the target variable through mutual information test, eliminating features with low correlation and forming feature vectors; The feature vector includes basic state features, target gap features, time period features, thermodynamic derived features, production load features, and equipment health features; The basic state characteristics include the initial outdoor temperature, the initial indoor humidity, and the initial chilled water supply temperature. The initial outdoor temperature and the initial indoor humidity are extracted from the environmental temperature and humidity parameters. The target difference features include the initial indoor temperature difference ΔT and the initial indoor humidity difference ΔH, which are calculated from the indoor ambient temperature and humidity parameters and the target temperature and humidity values. It also includes the cooling / heating load difference ΔQ, which is calculated from the actual load and the process requirement load. The time period characteristics include startup time, startup time attribute, production shift attribute, holiday attribute, and meteorological season attribute; The thermodynamic derived characteristics include initial moisture content, initial air enthalpy, and wet-bulb temperature, which are calculated from ambient temperature and humidity parameters combined with thermodynamic formulas. The production load characteristics include real-time production capacity and the number of workshop equipment in operation. The equipment health characteristics include air conditioning fan operating efficiency, chilled water heat exchange coefficient, and parameters related to steam humidification module malfunction.

3. The method according to claim 1, characterized in that, The air conditioner start-up prediction model specifically includes: An improved random forest algorithm is used as the core regression model. The improved random forest consists of multiple decision trees, and each decision tree is given a feature weight constraint based on physical mechanism. Thermodynamic derived features and production load features are given higher initial weights. Each decision tree is trained based on a subset of samples obtained by Bootstrap sampling, and a hierarchical sampling method is used to ensure a balanced proportion of samples in each season and production load range. When splitting at each node, the optimal splitting feature is selected from a randomly selected feature subset, and the splitting criterion adopts the mean square error loss function combined with physical error. For the input feature vector, each decision tree independently outputs a predicted value. The weighted arithmetic mean of the predicted values ​​of all decision trees is taken as the final prediction time. The weights of the decision trees are dynamically allocated according to their prediction accuracy on the validation set. The improved random forest algorithm reduces the risk of overfitting through a dual mechanism of sample randomness and feature randomness, eliminates the prediction bias of a single decision tree through multi-tree ensemble averaging, and improves the interpretability and robustness of the model in industrial scenarios through physical mechanism constraints.

4. The method according to claim 1, characterized in that, The training process of the air conditioner start-up prediction model includes an offline training process and an online inference process. The offline training process includes: querying historical operation records from the database, cleaning and labeling the data, identifying the complete startup, operation, and shutdown process, extracting features X and actual time consumption Y, training the random forest model using K-fold cross-validation, and evaluating the model performance. The online inference process includes: periodically triggering the prediction task, acquiring real-time sensor data, constructing feature vectors, and loading the trained model for prediction.

5. The method according to claim 4, characterized in that... The data cleaning and labeling includes: The system automatically identifies each complete startup, operation, and shutdown process, defining the initial entry into the target range and the sustained stable operation for a preset time as the target point. The preset time is dynamically configured according to the precision requirements of the production process; a short-duration threshold is configured for precision manufacturing scenarios, while a long-duration threshold is configured for general industrial scenarios. The system extracts the feature X at the startup moment, and performs normalization and standardization on the feature X to eliminate dimensional differences. The actual time consumption is then calculated as the target Y. If there are outliers in the actual time consumption, the isolated forest algorithm is used to identify and remove them, while interpolation is used to supplement the effective samples.

6. The method according to claim 1, characterized in that, The calculation of the target time based on the predicted duration specifically includes: Obtain the target time T_target specified in the production plan and update it dynamically based on the production plan adjustment information; Obtain the prediction duration T_pred output by the air conditioner start-up prediction model, and correct T_pred according to the health status of the air conditioner equipment. When the equipment deteriorates, multiply it by the corresponding correction coefficient according to the degree of deterioration. Obtain the adaptive reservation time T_adaptive, which is dynamically calculated based on historical prediction errors; Calculate the power-on time point T_start using the following formula: T_start = T_target - T_pred - T_adaptive; The adaptive reservation time T_adaptive is dynamically determined based on the statistical distribution of historical prediction errors within the sliding window, and the length of the sliding window is adaptively adjusted according to the frequency of meteorological changes and the degree of fluctuation in production load; the greater the fluctuation, the shorter the window length.

7. The method according to claim 6, characterized in that, The adaptive reservation time T_adaptive is determined in the following ways: The prediction error between the actual time to achieve the target and the predicted time of the model is continuously recorded to form an error time series. The error time series is then smoothed to remove abnormal errors caused by sudden equipment failures or extreme weather. The mean μ and standard deviation σ of the prediction errors of the most recent N predictions are calculated using a sliding window method. Calculate the adaptive reservation time using the following formula: T_adaptive = μ + k × σ + α × Sk; Where k is the preset safety factor, Sk is the error skewness coefficient, and α is the skewness correction coefficient, which is configured according to the process compliance tolerance rate. When the volatility of the prediction error increases, the standard deviation σ increases, causing the adaptive reserve time to be automatically extended. When model performance improves and prediction errors converge, the adaptive reserve time is automatically shortened. When the error distribution is left-skewed, the reserve time is increased by the skewness correction coefficient to further improve the reliability of achieving the target.

8. An air conditioning control system based on physical modeling combined with big data prediction, characterized in that, The system for performing the method according to any one of claims 1 to 7 comprises: The sensor layer includes outdoor temperature and humidity sensors, fresh air temperature and humidity sensors, indoor temperature and humidity sensors, production load sensors, equipment status sensors, meteorological data acquisition modules, and production plan docking modules. It is used to monitor temperature and humidity data at different locations in real time, and also to collect production load data, air conditioning equipment health status data, outdoor weather forecast data, and factory production plan data. The data acquisition layer is used to acquire data and convert protocols based on the sensor layer. It supports high-speed industrial data transmission and has edge data preprocessing capabilities, which can complete preliminary data cleaning and feature extraction at the edge. The data preprocessing and storage layer is used to clean, interpolate, extract features and store raw data. It adopts an edge-cloud collaborative storage architecture, with real-time high-frequency data stored at the edge and historical massive data stored in the cloud. It also supports bidirectional data synchronization and breakpoint resume. The algorithm service layer includes a random forest-based air conditioner start-up prediction model, a REST API service, a model inference acceleration module, a multi-source data fusion module, and a dynamic adjustment module for operating parameters. It is used to predict air conditioner start-up time and provide standard calling interfaces. It is also used to realize the fusion processing of multi-source data, hardware acceleration of model inference, and dynamic adjustment of air conditioner operating parameters. The upper-level system includes a visualization platform, management system, production planning linkage module, anomaly alarm module, and energy consumption statistical analysis module. It is used to display prediction results, operating status, and historical data, provide system configuration and operation control functions, realize real-time linkage with production plan, intelligent alarm for air conditioning operation anomalies, and refined statistical analysis of air conditioning system energy consumption.

9. The system according to claim 8, characterized in that, The algorithm service layer also integrates an OPC UA client, which is used to subscribe to real-time sensor data from the factory OPC UA server and write the prediction results back to the preset data node for SCADA or HMI systems to read.

10. The system according to claim 8, characterized in that, The system also includes a model performance monitoring and retraining mechanism, which continuously records the deviation between the predicted value and the actual achievement time, calculates the mean absolute error and the coefficient of determination, triggers an alarm when the performance index exceeds a preset threshold and supports one-click retraining, forming an MLOps closed loop.

Images