An air conditioner energy consumption real-time optimization method based on environment perception

Abstract

The application relates to the technical field of intelligent control, and particularly discloses an air conditioner energy consumption real-time optimization method based on environment sensing, which aims to solve the problems of insufficient real-time control precision and instantaneous disturbance response lag of an existing air conditioner energy consumption model under a dynamic environment. The method is characterized in that a double-model predictive control framework is adopted, including a physical-experience hybrid baseline model for macro trend prediction and a real-time deviation correction model for micro instantaneous deviation correction, and the two are coupled through a model parameter self-correction unit to form a closed-loop control loop with adaptive learning capability. By adopting the above technical scheme, the present application can significantly reduce the energy consumption of an air conditioner system, improve the stability and comfort of an indoor environment, and ensure the high precision and high timeliness of a control strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a method for real-time optimization of air conditioning energy consumption based on environmental perception. Background Technology

[0002] In the energy consumption structure of modern buildings and industrial facilities, the heating, ventilation, and air conditioning (HVAC) system is the main energy-consuming unit, and its energy efficiency optimization is crucial to achieving energy conservation and emission reduction goals. Currently, energy consumption optimization methods based on environmental perception have become a research hotspot, the core of which lies in dynamically adjusting system operation strategies by monitoring environmental parameters in real time.

[0003] Existing technical solutions can be mainly divided into two categories: The first is a control method based on static energy efficiency parameters, which calculates cooling demand using a preset energy efficiency ratio (EER) and then adjusts equipment power accordingly. While simple to implement, this method ignores the dynamic changes in EER due to ambient temperature, humidity, and equipment status. It is prone to significant deviations under drastic fluctuations in operating conditions, leading to increased energy consumption or decreased comfort. The second method is a complex modeling approach based on historical data. This typically uses machine learning algorithms to train multidimensional historical data to predict system energy consumption trends. Although the model shows good accuracy in predicting macro trends, its reliance on long-term historical data and long model update cycles make it difficult to respond promptly to sudden, transient disturbances (such as dense crowds, open doors and windows, and sudden weather changes), resulting in significant control lag.

[0004] In general, simple models offer rapid responses but limited prediction accuracy, while complex models offer higher prediction accuracy but lack real-time performance. Therefore, achieving high-speed response to dynamic disturbances while ensuring model prediction accuracy has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time optimization method for air conditioning energy consumption based on environmental perception, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, a real-time optimization method for air conditioning energy consumption based on environmental perception is provided, including:

[0007] Step 1: Within a preset, high-frequency first time period (default value 1 second, configurable range 0.1-5 seconds), multi-dimensional status data is simultaneously collected using a multi-sensor array deployed inside and outside the target building space and on the air conditioning system itself. The multi-sensor array includes at least: an internal environment sensor group, set at multiple representative measuring points within the air conditioning service area, used to collect indoor temperature data, indoor humidity data, carbon dioxide concentration data, and illuminance data. Specifically, the temperature sensor uses a platinum resistance temperature sensor with a measurement accuracy of ±0.1 degrees Celsius; the humidity sensor uses a capacitive humidity sensor with a relative humidity measurement accuracy of ±2%; the carbon dioxide concentration sensor uses a non-dispersive infrared sensor with a measurement range of 0 to 5000 ppm and an accuracy of ±50 ppm; and the illuminance sensor measures the intensity of indoor artificial lighting and natural light.

[0008] The external environment data interface obtains localized meteorological data released by authoritative meteorological service agencies in real time through a web application interface, including outdoor dry-bulb temperature, outdoor wet-bulb temperature, atmospheric pressure, wind speed, wind direction, and direct and diffuse solar radiation intensity.

[0009] The sensor group integrated in the HVAC system includes: pressure and temperature sensors installed at the compressor suction and discharge ports to determine the real-time thermodynamic state of the refrigerant; and a high-precision three-phase power metering module connected in series with the main power supply circuits of the compressor, condenser fan, and evaporator fan, with a metering accuracy class of 0.2S, used to accurately measure the instantaneous active power and cumulative power consumption of each major energy-consuming component of the system.

[0010] After data acquisition, the data preprocessing unit is responsible for performing timestamp alignment, outlier removal based on median absolute deviation, and normalization of data of different dimensions to the [0, 1] interval on all acquired raw data, providing a high-quality, synchronous, and formatted input data stream for subsequent model calculations.

[0011] Step 2: The rationale model is based on the first law of building thermodynamics. It calculates the total cooling load required to maintain the indoor set temperature under given indoor and outdoor environmental conditions. The total cooling load consists of the following components: heat transfer load Q_trans through the building envelope (walls, roof, windows), load Q_solar generated by solar radiation, and internal heat source load Q_internal generated by indoor occupants, equipment, and lighting.

[0012] The formula for calculating the heat transfer load Q_trans is as follows:

[0013] Q_trans = Σ(U_i * A_i * ΔT_i);

[0014] U_i is calculated based on the heat transfer coefficient and thickness of the building envelope, specifically by dividing the building's heat transfer coefficient by its thickness. A_i is its area, and ΔT_i is the indoor-outdoor temperature difference.

[0015] The load Q_solar is obtained based on window orientation, the optical properties of the glass, and real-time solar radiation intensity data;

[0016] The load Q_internal is dynamically estimated based on the heat dissipation of personnel and carbon dioxide concentration data.

[0017] The empirical data sub-model uses the gradient boosting decision tree algorithm, which is trained by supervised learning based on historical operating data. Its input feature vector includes time series features such as indoor and outdoor temperature and humidity, carbon dioxide concentration, solar radiation intensity, actual power consumption of air conditioner at historical time, day of the week, and specific time. Its output is the predicted operating power of the air conditioning system under a specific load in the future control time domain.

[0018] The model fusion logic unit weights and fuses the theoretical cooling load calculated by the physical mechanism sub-model with the operating power predicted by the empirical data sub-model. Instead of a fixed value, it is dynamically adjusted based on the prediction errors of the two sub-models in the past period. The sub-model with the smaller error will be given a higher weight in the next period, thereby achieving complementary advantages.

[0019] The fusion weights are adjusted using the following formula:

[0020]

[0021] in, and These represent the mean absolute errors of the physical sub-model and the empirical sub-model within the period [t-1]. The smoothing coefficient (value 0.5) is used to normalize the weights to the [0,1] interval.

[0022] Step 3: Within the preset first time period T1, the real-time deviation correction model unit continues to run and calculates the prediction deviation Error(t). The calculation formula is: Error(t) = P_actual(t) - P_baseline(t), where P_actual(t) is the instantaneous actual power consumption of the system measured by the sensor group integrated in the HVAC system, and P_baseline(t) is the predicted power consumption at the current moment obtained from the baseline energy consumption prediction curve generated in Step 2.

[0023] The prediction bias Error(t) is used as the core input and fed into the Kalman filter. The Kalman filter is constructed as a single-variable linear dynamic system to filter, smooth, and predict the value of the noisy bias sequence at the next time step. The state equation and observation equation of the Kalman filter are as follows: State equation: x(k) = A * x(k-1) + w(k-1) Observation equation: z(k) = H * x(k) + v(k) Where, x(k) is the system state at time k, i.e., the true, noise-free prediction bias; z(k) is the observation value at time k, i.e., the calculated Error(t); A is the state transition matrix, H is the observation matrix, w(k-1) is the process noise, where the covariance of the process noise is Q, which is used to quantify the severity of the prediction bias's fluctuation over time; v(k) is the observation noise, where the covariance of the observation noise is R, where R is obtained by calculating the variance of the bias sequence, and Q is determined by estimating the process noise intensity through the autocorrelation function.

[0024] The Kalman filter operates iteratively through two standard steps: prediction and update. At the end of each first time period, it outputs the optimal estimate of the prediction deviation for the next time period, denoted as ΔP_correction(t+1). The optimal estimate accurately quantifies the energy consumption fluctuations that the baseline model fails to capture, caused by unmodeled, sudden, high-frequency disturbances (such as a window being suddenly opened or a group of people suddenly entering a meeting room).

[0025] Step 4: The model parameter self-calibration unit operates in the third time period between the first and second time periods. It analyzes the bias sequence smoothed by the Kalman filter. If the bias sequence consistently exhibits systematic positive or negative biases under specific conditions (such as periods of sustained high solar radiation), it indicates a mismatch in the corresponding parameters of the physics-empirical hybrid baseline model. For example, a persistent positive bias (actual power consumption greater than predicted power consumption) might mean that a parameter of the building's actual solar radiation heat gain coefficient Q_solar is underestimated, or that compressor efficiency has decreased due to aging. In this case, the model parameter self-calibration unit activates the recursive least squares parameter identification algorithm. The algorithm uses model parameters (such as U_i, or certain weighting coefficients in the empirical model) as the variables to be optimized, with the objective function of minimizing the sum of squares of long-term prediction biases, and iteratively updates these parameters online. This allows the physics-empirical hybrid model to learn from its short-term prediction errors and continuously improve itself.

[0026] Step 5: The optimal control strategy generation unit integrates the outputs from the physical-empirical hybrid baseline model and the bias correction model within the first time period to generate the final target power P_target(t+1)).

[0027] P_target(t+1) = P_baseline(t+1) + ΔP_correction(t+1);

[0028] Subsequently, the optimal control strategy generation unit converts the final target power into the target frequency that the compressor needs to operate at through a pre-calibrated "power-frequency" performance curve. At the same time, based on the deviation between the indoor temperature and humidity and the set value, the speed of the evaporator fan is adjusted through a proportional-integral-derivative controller to ensure that the air volume and cooling output match.

[0029] Step Six: The control commands issued by the optimal control strategy generation unit are received by the control execution module. The control execution module first encodes the control commands into a communication protocol that conforms to industrial standards (such as Modbus RTU or BACnet / IP). The control commands are then sent to the programmable logic controller (PLC or dedicated controller) of the HVAC system, which ultimately drives the frequency converter or motor to perform actions, thus completing the closed-loop control of the air conditioning system.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] This invention resolves the contradiction between macroscopic prediction accuracy and microscopic control response in traditional energy consumption models through its unique dual-model predictive control framework that combines physical mechanisms and empirical data, along with a model parameter self-calibration mechanism with closed-loop feedback capability. This method can not only accurately predict long-term macroscopic trends in energy consumption like a complex historical data model, but also respond quickly to microscopic environmental disturbances like a simple real-time controller. Furthermore, it combats model drift through continuous self-learning and calibration. Ultimately, it significantly reduces the energy consumption of the air conditioning system throughout its entire lifecycle while greatly improving the stability and comfort of the indoor environment, demonstrating high engineering practical value and broad market prospects. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a real-time optimization method for air conditioning energy consumption based on environmental perception, according to the present invention.

[0033] Figure 2 This is a functional block diagram of the control system described in this invention;

[0034] Figure 3 This is a flowchart of the baseline energy consumption prediction steps of the physical-empirical hybrid baseline model in this invention;

[0035] Figure 4 This is a flowchart of the instantaneous deviation correction steps of the real-time deviation correction model in this invention;

[0036] Figure 5This is a flowchart illustrating the steps of online identification and updating of model parameter self-calibration unit in this invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” and “described” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0039] Please see Figures 1-5 As shown, a real-time optimization method for air conditioning energy consumption based on environmental perception is provided. This method aims to achieve precise real-time control of air conditioning system energy consumption through a dual-model predictive control framework.

[0040] In a specific implementation, this method is achieved through a central processing unit comprised of a main controller. The main controller can be an industrial-grade computer or an embedded server, which contains software programs running the algorithm logic described in this invention. The system communicates bidirectionally with the multi-sensor array and actuators of the HVAC system deployed in the field via a standardized industrial communication bus, such as RS-485 or Ethernet.

[0041] Step 1: Within the preset first time period, i.e. 1 second, the multi-sensor array and data preprocessing unit continuously collect raw data streams from multiple dimensions, reflecting the building environment status and the working conditions of the air conditioning system itself.

[0042] The multi-sensor array is designed as a multi-layered, wide-coverage sensing network, comprising at least three subsystems: an internal environment sensor group deployed in representative areas within the building, such as open-plan office spaces, window-side areas, and high-density areas, ensuring at least one measuring point in each independent temperature-controlled area. Each measuring point is equipped with a high-precision composite sensor to collect indoor temperature, humidity, carbon dioxide concentration, and illuminance data. As a preferred implementation, the temperature sensor uses a PT100 platinum resistance temperature sensor with a measurement accuracy of ±0.1 degrees Celsius; the humidity sensor uses a capacitive humidity sensor with a measurement accuracy of ±2% relative humidity; the carbon dioxide concentration sensor employs non-dispersive infrared technology, covering a range of 0 to 5000 ppm with a measurement accuracy of ±50 ppm. The carbon dioxide concentration data is also used to estimate the number of people and their activity levels in real time and dynamically, thereby accurately quantifying the heat load from personnel within the internal heat sources. The illuminance sensor is used to quantify the heat generated by artificial lighting systems and natural light through windows.

[0043] The external environment data interface is a software module that obtains localized meteorological data from authoritative meteorological service agencies via API, including outdoor dry and wet bulb temperatures, atmospheric pressure, wind speed, wind direction, and solar radiation intensity (divided into direct and diffuse radiation, unit W / m²), providing a basis for calculating solar heat gain.

[0044] The air conditioning system's status sensor array is integrated into key nodes of the HVAC system. Specific deployment includes: pairing temperature and humidity sensors at the air inlets and outlets of both the evaporator and condenser; measuring temperature and humidity changes before and after airflow through the heat exchanger, combined with airflow data, to accurately calculate the system's actual sensible and latent heat cooling capacity; installing pressure and temperature sensors on the compressor's suction and discharge lines respectively; and using the collected data to accurately calibrate the refrigerant's thermodynamic state point on the refrigerant pressure-enthalpy diagram, thereby evaluating the compressor's actual operating efficiency; crucially, a 0.2S-level high-precision three-phase power metering module is connected in series in the system's main circuit and the circuits of major energy-consuming components (such as the compressor and fan) to measure instantaneous active power, reactive power, power factor, and cumulative power consumption once per second. The system's total instantaneous active power P_actual(t) is the key input for real-time deviation correction.

[0045] All heterogeneous raw data streams acquired from the multi-sensor array are immediately sent to the data preprocessing unit for purification and formatting. The data preprocessing unit performs standardized operations to ensure data quality. The first step is timestamp alignment, assigning a unified, high-precision timestamp to all data points from different sensors with different sampling periods, ensuring that all input data are synchronized in time during subsequent model analysis. The second step is outlier removal, using a robust statistical method based on median absolute deviation to identify and remove abnormal spikes or noise points in the data stream. Compared to methods based on mean and standard deviation, the median absolute deviation method is less sensitive to extreme outliers and more effectively preserves the true data distribution characteristics. The third step is data normalization, scaling all input data with different physical dimensions, such as temperature, concentration, and power, to the dimensionless [0, 1] interval through linear transformation (e.g., min-max normalization). This eliminates large differences in numerical scales between different features, preventing features with larger numerical ranges from dominating model results during subsequent model training and computation, thus ensuring that information from all dimensions can be used fairly by the model. After preprocessing, the system generates a high-quality, synchronous, and formatted multidimensional data stream, providing a reliable data foundation for the calculation of all subsequent models.

[0046] Step 2: Baseline energy consumption prediction based on a physics-empirical hybrid baseline model. The baseline model generates cell rows, which are executed every 30 minutes. It utilizes historically accumulated knowledge and physical laws to generate a macroscopic trend prediction of the power consumption of the air conditioning system to maintain the set comfort level within one or more future control time domains (e.g., the next 2 hours), i.e., the baseline energy consumption prediction curve P_baseline(t).

[0047] The physical-empirical hybrid baseline model is structurally composed of two parallel sub-models: a physical mechanism sub-model and an empirical data sub-model.

[0048] The physical mechanism sub-model is constructed based on the first law of thermodynamics, calculating the total cooling load Q_total required to maintain a set indoor temperature under given indoor and outdoor environmental conditions. Q_total is decomposed into the sum of several main components, including the heat transfer load Q_trans through the building envelope, the load Q_solar generated by solar radiation, and the internal heat source load Q_internal generated by indoor occupants, equipment, and lighting systems. The heat transfer load Q_trans through the building envelope is calculated based on the steady-state heat transfer formula Q_trans = Σ(U_i * A_i * ΔT_i), where U_i is calculated based on the heat transfer coefficient and thickness of the building envelope representing each part (specifically, dividing the building's heat transfer coefficient by its thickness), A_i is the area of ​​the corresponding part, and ΔT_i is the effective temperature difference between indoors and outdoors. Parameters U_i and A_i are preset during system initialization based on the building's design drawings and material specifications. The heat dissipation power of equipment and lighting can be estimated based on their rated power and usage plan, while the heat dissipation load of personnel is dynamically estimated using the aforementioned carbon dioxide concentration data.

[0049] The relationship between carbon dioxide concentration and the number of people is expressed as a linear piecewise function, where the number of people N = max(0, a·[C] + b), and [C] is the real-time concentration of carbon dioxide. When [C] ≤ 800ppm, a = 0.02 people / ppm and b = 0; when 800ppm < [C] ≤ 2000ppm, a = 0.015 people / ppm and b = 4; a and b are coefficients.

[0050] The parallel empirical data sub-model employs a purely data-driven approach. As a preferred implementation, this embodiment uses a gradient boosting decision tree algorithm (such as LightGBM or XGBoost), trained on current and historical indoor and outdoor environmental data (temperature, humidity, carbon dioxide concentration, solar radiation). Input features include indoor and outdoor environmental parameters, historical system operation data, and time-series features, such as day-of-the-week features encoded using one-hot encoding, and specific times (e.g., hours of the day, days of the year) that have undergone sine / cosine transformations to express their periodicity. The output is the operating power of the air conditioning system in the future control time domain. Utilizing the empirical data sub-model, it is possible to learn from the data factors such as efficiency degradation due to equipment aging, subtle seasonal variations in the airtightness of the building envelope, and user behavior patterns that are difficult to model, thereby effectively supplementing and correcting the prediction results of the physical model.

[0051] The model fusion logic unit is responsible for dynamically weighting and fusing the theoretical cooling load calculated by the physical mechanism sub-model with the operating power directly predicted by the empirical data sub-model. Its fusion weight w has adaptive adjustment capabilities. The system continuously tracks the error between the predicted and actual values ​​of the two sub-models in the past cycle (e.g., the previous 30 minutes). In the fusion calculation of the next cycle, the sub-model with the smaller error will be assigned a higher weight. The final result after fusion is a smoothed baseline energy consumption prediction curve P_baseline(t).

[0052] The fusion weights are adjusted using the following formula:

[0053]

[0054] in, and These represent the mean absolute errors of the physical sub-model and the empirical sub-model within the period [t-1]. The smoothing coefficient (value 0.5) is used to normalize the weights to the [0,1] interval.

[0055] Step 3: The real-time deviation correction model unit runs with a 1-second cycle to predict instantaneous deviations. At the end of each first time cycle, the deviation calculation module calculates the predicted deviation Error(t) in real time.

[0056] Error(t) = P_actual(t) - P_baseline(t), where P_actual(t) is the instantaneous total active power of the system measured by the high-precision power metering module in the air conditioning system's main body state sensor group, and P_baseline(t) is the predicted power consumption value at the current moment obtained by interpolation from the baseline energy consumption prediction curve generated in step two.

[0057] The bias sequence Error(t) is fed into the Kalman filter as the core observation for processing. In this embodiment, the mathematical model of the Kalman filter is defined by the state equation and the observation equation: the state equation is: x(k) = A * x(k-1) + w(k-1); the observation equation is: z(k) = H * x(k) + v(k).

[0058] In this model, state x(k) represents the true, noise-free prediction deviation value at time k; observation z(k) is Error(t) calculated by the deviation calculation module. The state transition matrix A is set to 1, the observation matrix H is set to 1, w(k-1) is the process noise, and its covariance Q is used to quantify the severity of the natural fluctuation of the prediction deviation over time. v(k) is the observation noise, and its covariance R comprehensively reflects the noise level introduced by the sensor noise of the actual power consumption measurement P_actual(t) and the prediction error of the baseline model P_baseline(t). In this embodiment, R=0.25 is obtained by calculating the variance of the deviation sequence, and Q=0.01 is set by estimating the process noise intensity using the autocorrelation function.

[0059] The Kalman filter iterates once every first time period (1 second). In the prediction step, the state at the current time is predicted based on the state estimate from the previous time step. In the update step, the predicted value is corrected using the actual observation value Error(t) at the current time step, thus obtaining a more accurate optimal estimate of the current state.

[0060] Step 4: The model parameter self-calibration unit runs every 10-15 minutes in the preset third time period. It uses the recently accumulated deviation information, which has been smoothed by the Kalman filter, to fine-tune the key parameters of the physical-empirical hybrid baseline model in Step 2 online.

[0061] The model parameter self-calibration unit continuously analyzes the smoothed deviation sequence output by the Kalman filter. If the deviation sequence consistently and systematically exhibits positive or negative deviations under specific and identifiable operating conditions, and meets the preset trigger condition: triggering occurs when the average absolute deviation exceeds 5% of the baseline predicted energy consumption for 15 consecutive minutes, it indicates that one or more parameters in the physical-empirical hybrid baseline model are mismatched with physical reality. The average absolute deviation refers to the average absolute deviation of the deviation sequence after smoothing by the Kalman filter.

[0062] Once a preset trigger condition is detected, the model parameter self-calibration unit activates the recursive least squares parameter identification algorithm. One or more parameters in the baseline model suspected of being mismatched (e.g., the building envelope heat transfer coefficient U_i, solar heat gain correction coefficient in the physical model, or feature weights related to specific operating conditions in the empirical model) are used as variables to be optimized. The objective function is to minimize the long-term cumulative sum of squared prediction biases, and these parameters are iteratively updated online. In the recursive least squares method, the "forgetting factor" λ is set to 0.995.

[0063] Step 5: The optimal control strategy generation unit calculates the final target power value P_target(t+1) according to the following formula:

[0064] P_target(t+1) = P_baseline(t+1) + ΔP_correction(t+1)

[0065] P_baseline(t+1) represents the power prediction value at the next moment output by the baseline model generation unit, which is used to reflect the baseline power demand of the system under macroscopic trends; ΔP_correction(t+1) is the power correction amount at the next moment output by the real-time deviation correction model unit, which is used to compensate for the power deviation caused by instantaneous disturbances or unmodeled dynamic characteristics.

[0066] Subsequently, the optimal control strategy generation unit parses the target power value P_target into specific equipment operation instructions and sends them to the HVAC system for execution, which includes the following steps:

[0067] Based on the pre-calibrated compressor "power-frequency" performance curve or lookup table, P_target is converted to the target operating frequency value of the variable frequency compressor to directly adjust its cooling capacity output;

[0068] By comparing the current indoor temperature and humidity values ​​with the user-set values, if there is a deviation, the proportional-integral-derivative (PID) control algorithm is activated to generate adjustment commands for the evaporator fan speed or the variable air volume (VAV) terminal damper opening, so that the air supply volume matches the cooling capacity output by the compressor, avoiding overcooling or undercooling.

[0069] The optimal control strategy generation unit sends frequency commands and air volume adjustment commands to the system controller together to achieve coordinated control of multiple components of the HVAC system.

[0070] Step Six: All specific control commands generated by the optimal control strategy generation unit, such as the target frequency of the compressor, the target opening degree of each zone's dampers, and the target speed of the fan, are first sent to the control execution interface module. The control execution interface module encodes the internal control commands into communication protocol messages conforming to industrial fieldbus standards (such as Modbus RTU or BACnet / IP). After encoding, these command messages are sent to the HVAC system's field controller via the corresponding physical communication bus (such as RS-485 twisted pair or Ethernet). After receiving and parsing the commands, the controller drives its connected final actuators (such as frequency converters, electric damper actuators, and motors) to precisely execute the corresponding actions, thereby completing the closed-loop control of the air conditioning system's operating status. Thus, the complete process from data acquisition to command execution constitutes a cyclical, continuously self-optimizing closed-loop control system.

[0071] To further illustrate the technical effects of the present invention, specific embodiments and two comparative examples will be used for comparison and explanation below.

[0072] In a specific embodiment, the method of the present invention is applied to a modern independent office building with a total building area of ​​2,000 square meters, and the air conditioning system of the building is an advanced variable air volume (VAV) central air conditioning system.

[0073] The system deployment details of this embodiment are as follows. Firstly, regarding sensor deployment, a total of 50 indoor environmental monitoring points are evenly or strategically distributed throughout the office buildings, meeting rooms, corridors, and other areas. Each monitoring point is equipped with a three-in-one sensor module integrating high-precision temperature, humidity, and carbon dioxide concentration sensors. The temperature measurement accuracy is ±0.1°C, the relative humidity measurement accuracy is ±2%RH, and the carbon dioxide concentration measurement accuracy is ±50ppm. A small automatic weather station is installed on the building's roof to monitor outdoor environmental parameters in real time, including outdoor temperature, humidity, and total solar radiation intensity. The total solar radiation intensity is measured using a sensitive photovoltaic cell-type total radiation meter with a measurement accuracy of 5%. On the central air conditioning unit side, high-precision multi-functional energy meters with an accuracy class of 0.2S are installed on the main power input line, compressor main circuit, chilled water pump circuit, and cooling fan circuit. Furthermore, temperature and humidity sensors are installed on the central air conditioning's main supply air duct, return air duct, and branch air ducts leading to main areas to monitor the air supply status.

[0074] In terms of control system hardware, an industrial-grade computer equipped with an Intel Core i7 processor and 16GB of RAM was selected as the main controller. This computer runs customized software that implements the dual-model predictive control framework and all related algorithms described in this invention. The main controller exchanges data at high speed and reliably with all the sensor modules and the core PLC controller of the air conditioning system on site via an RS-485 communication bus.

[0075] In terms of software algorithm implementation and parameter settings, the first time period in data acquisition and preprocessing was set to 1 second to ensure maximum real-time performance. The physical mechanism sub-model in the baseline model generation unit was pre-built based on the detailed architectural design drawings of the office building, which included the accurate material properties and heat transfer coefficients of all walls, roofs, glass curtain walls, and other building envelopes. The empirical data sub-model was implemented using the LightGBM algorithm framework and was fully trained offline using a year's worth of historical operating data accumulated before the system was put into formal operation. The update cycle of the physical-empirical hybrid baseline model, i.e., the second time period, was set to 30 minutes. In the real-time deviation correction model, the key parameters of the Kalman filter, process noise covariance Q and observation noise covariance R, were set to 0.01 and 0.25, respectively. These values ​​were optimized based on statistical analysis of deviation data from one week of system trial operation. The operating cycle of the model parameter self-calibration unit, i.e., the third time cycle, is set to 15 minutes. Its trigger condition is: when the average smoothing deviation over 15 consecutive minutes exceeds 5% of the current predicted power consumption, the RLS algorithm is activated to fine-tune the solar heat gain correction coefficient in the physical model and the weights of several time-related key features in the empirical data model. Finally, the optimal control strategy generation unit converts the fused target power into compressor output frequency commands and opening commands for each VAV terminal damper in real time.

[0076] To compare the effects, two comparison ratios were set up.

[0077] Comparative Example 1 uses the exact same hardware environment as the above embodiment, but replaces the control method with a control scheme based on static energy efficiency ratio (EER), which is more common in the prior art. The control logic of this scheme is as follows: the system presets a set of corresponding EER values ​​for the air conditioning system according to different load rates. This set of EER values ​​is usually manually updated quarterly by maintenance personnel based on experience. In real-time control, the system only collects the average indoor temperature. When the temperature deviates from the setpoint, it calculates the required theoretical cooling capacity, then calculates the target power based on the preset EER values, and starts or stops the air conditioning system or adjusts its speed accordingly.

[0078] Comparative Example 2 uses the exact same hardware environment as Example 1, but replaces the control method with a more advanced scheme that relies entirely on historical data models. This system utilizes a deep learning model, specifically a Long Short-Term Memory (LSTM) network. This network predicts energy consumption and indoor temperature trends for the next hour based on all environmental and system data collected over the past 24 hours, and pre-determines a control plan for the next hour based on this prediction. The LSTM model is retrained once a day. During operation, the system's real-time control commands strictly adhere to this pre-defined hourly plan, lacking the ability to adjust for unplanned events in real time.

[0079] To objectively evaluate the performance differences among the three systems, a typical summer workday with variable weather conditions was selected for a 24-hour continuous test. The weather characteristics of that day were: sunny with strong solar radiation in the morning; around 2:00 PM, the weather suddenly turned into a short-lived thunderstorm, causing a sharp drop in outdoor temperature and a significant reduction in solar radiation; after 4:00 PM, the weather cleared up again. Simultaneously, to simulate sudden changes in indoor load, an impromptu, unannounced meeting was held in a large conference room that could accommodate 100 people at 10:00 AM, causing a rapid increase in the number of people in the area within a short period. During the test, detailed records and comparisons were made of the key performance indicators of the three systems under identical conditions.

[0080] Table 1: Performance Comparison Table of Example 1, Comparative Example 1, and Comparative Example 2

[0081] Performance indicators Example 1 (This Invention) Comparative Example 1 (Static EER) Comparative Example 2 (Historical Model) Total energy consumption (kWh) 285.4 352.1 320.8 Average indoor temperature deviation (°C) ±0.22 ±0.85 ±0.51 Peak temperature overshoot (°C) 0.4 1.9 1.2 Response time (min) to a sudden increase in personnel at 10:00 AM <1 12 8 Response time (min) to the sudden weather change at 14:00 <1 15 >30 (within the hourly plan)

[0082] Analysis of the data in Table 1 clearly demonstrates the significant advantages of the technical solution of this invention. Regarding total energy consumption, Embodiment 1 of this invention consumes 285.4 kWh over 24 hours, the lowest among the three. Compared to Comparative Example 1's 352.1 kWh, the energy saving reaches 19.0%; compared to Comparative Example 2's 320.8 kWh, the energy saving also reaches 11.0%. This is mainly due to the dual-model framework of this invention, which can accurately grasp the macro-trend of energy consumption throughout the day through the baseline model, and can also compensate for various disturbances in real time through the deviation correction model, avoiding ineffective overcooling or overheating operation caused by inaccurate prediction or delayed response.

[0083] Regarding indoor temperature control accuracy, Example 1 achieved an average indoor temperature deviation of ±0.22°C and a peak temperature overshoot of only 0.4°C, providing the most stable and comfortable environment for indoor occupants. In contrast, Comparative Example 1, due to its static EER value, experienced severe control logic malfunctions when the weather suddenly changed at 2 PM and the outdoor load decreased sharply, resulting in significant overcooling for a prolonged period, with the peak temperature even falling below the setpoint by 1.9°C. This not only caused serious energy waste but also reduced occupant comfort. Although Comparative Example 2 performed better overall than Comparative Example 1, its lagging model updates prevented it from effectively responding to emergencies. Significant temperature fluctuations occurred during the sudden increase in meeting room occupants at 10 AM and the sudden weather change at 2 PM, with a peak overshoot reaching 1.2°C.

[0084] In terms of system response speed, Example 1 has a particularly prominent advantage. Its Kalman filter-based deviation correction model operates at a frequency of seconds, enabling it to complete detection, calculation, and control adjustment within one minute of a load change. Whether it's a sudden increase in internal personnel load or a drastic change in external weather conditions, its response time is less than one minute. The slow response of Comparative Example 1 stems from its control logic relying on the accumulation of indoor temperature changes to a certain level before being passively triggered. The slow response of Comparative Example 2 is a fundamental flaw in its control framework. Its hourly forecasting and control planning makes it unable to handle any unplanned instantaneous disturbances. For a sudden weather change at 14:00, the adjustment of its control strategy is effectively delayed to the next hour's planning cycle, with a response time far exceeding 30 minutes.

[0085] Furthermore, in long-term continuous observation, the model parameter self-calibration unit of Embodiment 1 of this invention demonstrated its key value. The long-term moving average of the baseline model prediction error consistently and stably approached zero, indicating that the unit effectively tracked slow changes in the system and environment, such as changes in the angle of sunlight due to seasonal changes and performance degradation of equipment after long-term operation, thereby maintaining the effectiveness and accuracy of the model throughout its entire lifecycle. In contrast, the models of Comparative Examples 1 and 2, without periodic manual recalibration or large-scale model retraining, exhibited a significant and monotonically increasing trend in prediction error over time.

[0086] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time optimization method for air conditioning energy consumption based on environmental perception, applied to a control system consisting of a main controller, a multi-sensor array, a model logic unit, a data processing unit, a real-time deviation correction model, a model parameter self-calibration unit, and an HVAC system, characterized in that... Includes the following steps: Within the preset first time period T1, multi-dimensional state data is collected and preprocessed in real time using a multi-sensor array. Within the preset second time period T2, a physical-empirical hybrid baseline model, based on the physical mechanism sub-model and the empirical data sub-model, generates a baseline energy consumption prediction curve (P_baseline(t)) to characterize the macroeconomic energy consumption trend. Within the first time period, the prediction deviation (Error(t)) between the instantaneous actual power consumption (P_actual(t)) of the HVAC system and the predicted power consumption of the baseline energy consumption prediction curve at the current moment is calculated based on the real-time deviation correction model. This prediction deviation is then filtered and state predicted to obtain the instantaneous deviation correction value at the next control moment. Within the preset third time period, the cumulative prediction bias in recent times is analyzed based on the model parameter self-calibration unit, and the key parameters of the physical-empirical hybrid baseline model are identified and updated online when a systematic bias is detected. Within the first time period, the predicted value of the baseline energy consumption prediction curve at the next control moment (P_baseline(t+1)) and the instantaneous deviation correction value (ΔP_correction(t+1)) are merged to generate the final target power P_target(t+1)), which is then parsed into specific control commands and sent to the HVAC system for execution. The generation of the baseline energy consumption prediction curve includes: a physical mechanism sub-model, based on the first law of building thermodynamics, calculates the total cooling load required to maintain the indoor set temperature under given indoor and outdoor environmental conditions. This total cooling load includes heat transfer load through the building envelope, load generated by solar radiation, and internal heat source load generated by indoor occupants, equipment, and lighting. An empirical data sub-model employs a gradient boosting decision tree algorithm, trained using supervised learning based on historical operating data, and predicts the operating power of the air conditioning system in the future control time domain based on current and historical environmental state data, system operating data, and time series characteristics. A model fusion logic unit weights and fuses the theoretical power corresponding to the theoretical cooling load calculated by the physical mechanism sub-model with the operating power predicted by the empirical data sub-model to generate the baseline energy consumption prediction curve. The generation of the baseline energy consumption prediction curve includes a weighted fusion of the theoretical power calculated by the physical mechanism sub-model and the operating power predicted by the empirical data sub-model. The weighted fusion adopts dynamic adjustment of weights so that the sub-model with smaller prediction error in the past period is given a higher weight in the fusion calculation of the next period. The steps for obtaining the instantaneous deviation correction value at the next control moment include: The prediction deviation is calculated in real time within each first time period by the deviation calculation module. The prediction deviation Error(t) is input as the observation sequence into the Kalman filter to filter and smooth the noisy prediction deviation sequence and predict its value at the next moment, so as to output the instantaneous deviation correction value (ΔP_correction(t+1)) and obtain the smoothed deviation sequence. The calculation formula is as follows: Error(t) = P_actual(t) - P_baseline(t).

2. The method for real-time optimization of air conditioning energy consumption based on environmental perception according to claim 1, characterized in that, The multidimensional status data includes: indoor temperature data, indoor humidity data, carbon dioxide concentration data, and illuminance data collected by an internal environmental sensor array deployed within the air-conditioned service area; Local meteorological data is obtained in real time through a web application programming interface. The meteorological data includes outdoor dry-bulb temperature, outdoor wet-bulb temperature, atmospheric pressure, wind speed, wind direction, and the intensity of direct and diffuse solar radiation. The system's operating conditions are obtained through a sensor array integrated within the HVAC system, including the temperature and humidity at the evaporator and condenser inlets and outlets, the pressure and temperature at the compressor suction and exhaust ports, and the instantaneous active power of the system's major energy-consuming components as measured by a high-precision three-phase power metering module. The processing procedure is as follows: The data preprocessing unit performs timestamp alignment processing, outlier removal processing based on the median absolute deviation method, and normalization of data of different dimensions to a preset range on all the collected raw data to generate a high-quality, synchronous, and formatted input data stream.

3. The method for real-time optimization of air conditioning energy consumption based on environmental perception according to claim 2, characterized in that: The state equation and observation equation of the Kalman filter are as follows: State equation: x(k) = A * x(k-1) + w(k-1); Observation equation: z(k) = H * x(k) + v(k); Where x(k) is the true, noise-free prediction bias at time k, z(k) is the observed value at time k, i.e., the calculated prediction bias (Error(t)), A is the state transition matrix, H is the observation matrix, w(k-1) is the process noise, where the covariance of the process noise is Q, and the covariance Q is used to quantify the severity of the prediction bias's fluctuation over time, v(k) is the observation noise, and the covariance of the observation noise is R, where R is obtained by calculating the variance of the bias sequence, and Q is determined by estimating the process noise intensity through the autocorrelation function.

4. The method for real-time optimization of air conditioning energy consumption based on environmental perception according to claim 1, characterized in that: Online identification and updating of key parameters, including: During the third cycle, the smoothed deviation sequence is continuously analyzed to determine whether the smoothed deviation sequence continuously exhibits systematic positive or negative deviation under specific operating conditions and meets the preset trigger conditions: when the average absolute deviation for N consecutive minutes exceeds K% of the baseline predicted power consumption, it is triggered; when the trigger conditions are met, the parameter identification algorithm is activated, using one or more parameters in the physical-empirical hybrid baseline model as the variables to be optimized, and minimizing the sum of squares of long-term prediction deviations as the objective function, the variables to be optimized are updated online iteratively.

5. The real-time optimization method for air conditioning energy consumption based on environmental perception according to claim 4, characterized in that: The parameter identification algorithm is a recursive least squares parameter identification algorithm with a forgetting factor λ of 0.

995. The variables to be optimized include the heat transfer coefficient of the building envelope, the solar heat gain correction coefficient, or the feature weights related to specific operating conditions in the physical-empirical hybrid baseline model.

6. The real-time optimization method for air conditioning energy consumption based on environmental perception according to claim 4, characterized in that: The step of generating the final target power and parsing it into specific control commands specifically includes: The optimal control strategy generation unit calculates the final target power, and the calculation formula is as follows: P_target(t+1) = P_baseline(t+1) + ΔP_correction(t+1); The final target power (P_target(t+1)) is reverse-analyzed into the target frequency that the variable frequency compressor in the HVAC system needs to operate at by using a pre-calibrated "power-frequency" performance curve or lookup table. Meanwhile, based on the deviation between the current indoor temperature and humidity values ​​and the set values, the proportional-integral-derivative controller coordinates the adjustment of the evaporator fan speed or the opening of the variable air volume terminal device's damper to ensure that the total air volume matches the cooling capacity output by the compressor.

7. The real-time optimization method for air conditioning energy consumption based on environmental perception according to claim 1, characterized in that: The internal environment sensor group deployed in the air-conditioned service area uses sensors based on the non-dispersive infrared principle. The carbon dioxide concentration data collected by these sensors is used as a proxy variable in the physical mechanism sub-model. Through a preset correspondence between carbon dioxide concentration and the number of people, the real-time carbon dioxide concentration value is dynamically converted into an estimated number of people indoors, which is then multiplied by the standard per capita heat dissipation to calculate the load component generated by heat dissipation from people in the internal heat source load in real time and accurately. The correspondence between carbon dioxide concentration and the number of people adopts a linear piecewise function, where the number of people N = max(0, a·[C] + b), where [C] is the real-time concentration of carbon dioxide, and a and b are coefficients. When [C] ≤ 800ppm, a = 0.02 people / ppm and b = 0; when 800ppm < [C] ≤ 2000ppm, a = 0.015 people / ppm and b = 4.

Images