Air conditioner energy consumption self-adaptive management system and method based on dynamic feature selection

Abstract

The invention discloses an air conditioner energy consumption adaptive management system and method based on dynamic feature selection, and relates to the technical field of air conditioner energy consumption management, and the method comprises the steps: building a system energy consumption digital twin model as a theoretical optimal energy consumption baseline; operating parameters are collected in real time, and key parameter subsets are screened through working condition recognition and feature importance dynamic evaluation; inputting the key parameters into the model to obtain theoretical energy consumption, and comparing the theoretical energy consumption with a measured value to generate an energy consumption deviation rate; smooth processing is carried out on the deviation ratio sequence, recognition is carried out in combination with a dynamic threshold value and various anomaly detection algorithms, and grading early warning is triggered; the system comprises four core modules, namely a digital twinborn model construction module, a key parameter dynamic screening module, an energy consumption deviation calculation module and a grading early warning triggering module. The method can adapt to different working conditions, accurately capture energy consumption abnormities, and effectively improve the intelligent level and accuracy of energy efficiency management of the air conditioning system.

Description

Technical Field

[0001] This invention relates to the technical field, specifically to an adaptive air conditioning energy consumption management system and method based on dynamic feature selection. Background Technology

[0002] Building energy conservation and consumption reduction have become a focus of social attention. As a major energy-consuming unit in public buildings, the energy efficiency management of central air conditioning systems is of paramount importance. Currently, common air conditioning energy management methods mainly rely on setting fixed energy efficiency benchmarks or comparing and analyzing static models based on historical data. However, air conditioning systems are complex systems with multiple variables, nonlinearity, and strong coupling; their energy consumption characteristics are dynamically affected by various factors such as outdoor environmental parameters, building internal load, and equipment operating status.

[0003] Existing technologies have significant limitations: First, static energy consumption models or benchmarks cannot accurately reflect the optimal operating state of the system under different operating conditions, leading to distortion of energy efficiency assessment benchmarks; second, among numerous operating parameters, the key parameters affecting energy consumption vary under different operating conditions, while existing methods typically use fixed parameter sets for analysis, failing to dynamically capture the core influencing factors under the current operating conditions, resulting in insufficient accuracy in energy consumption deviation analysis; third, traditional early warning mechanisms often use fixed thresholds, making it difficult to adapt to dynamic changes in system load and environmental conditions, easily leading to false alarms or missed alarms, and failing to effectively identify gradual or periodic energy efficiency degradation trends.

[0004] Therefore, developing an air conditioning energy consumption management method that can adapt to operating conditions, dynamically focus on key influencing parameters, and achieve intelligent and accurate early warning is of great significance for improving energy efficiency management and reducing system operating costs. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive air conditioning energy consumption management system and method based on dynamic feature selection, so as to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive air conditioning energy consumption management system and method based on dynamic feature selection. The methods include: Step S1: Based on the initial calibration data of the air conditioning room, establish a digital twin model of system energy consumption; Step S2: Collect the set of operating parameters of the air conditioning room in real time, and dynamically filter the subset of key parameters from the set of operating parameters based on the current operating conditions and through feature importance assessment; Step S3: Input the subset of key parameters into the system energy consumption digital twin model to calculate the theoretical optimal energy consumption under the current operating conditions; compare the theoretical optimal energy consumption with the measured total energy consumption to generate the energy consumption deviation rate; Step S4: Perform time series analysis on the energy consumption deviation rate. When the energy consumption deviation rate exceeds the dynamic threshold or shows a preset abnormal trend, trigger a graded early warning.

[0007] Furthermore, step S1 includes: Step S1-1: Collect historical operating data of the air conditioning room as initial calibration data; the initial calibration data includes at least environmental parameters, equipment operating parameters, system setting parameters, system load rate, and corresponding total energy consumption data; the environmental parameters include at least outdoor dry-bulb temperature and outdoor wet-bulb temperature; clean the collected data, handle outliers, and standardize it to form a dataset; Step S1-2: Select initial static feature variables that are strongly correlated with system energy consumption from the dataset; and select the gradient boosting decision tree algorithm as the architecture of the system energy consumption digital twin model; Step S1-3: Use the dataset to train and evaluate the digital twin model of the system energy consumption; randomly divide the training data into k parts, use k-1 parts as the training set and the remaining part as the validation set, repeat k times, and use this k-fold cross-validation method to evaluate the prediction accuracy and generalization ability of the model, and perform hyperparameter tuning based on the evaluation results until the model reaches the preset accuracy threshold. Step S1-4: The validated and best-performing digital twin model is determined as the baseline model, and the energy consumption prediction value of the baseline model under typical operating conditions is the theoretical optimal energy consumption baseline.

[0008] Furthermore, step S2 includes: Step S2-1: At a preset data collection time interval The system collects a set of operational parameters in real time, including environmental parameters, equipment operating parameters, and system status parameters; and aligns the data aggregation timestamps through an edge computing gateway. Step S2-2: Preprocess the set of operating parameters, including using the Kalman filter algorithm to reduce noise in continuous parameters, using interpolation methods based on similar operating conditions to fill in missing values, marking and replacing outliers, and applying the formula... Normalization is performed, where x is the original parameter value. , These are the minimum and maximum values ​​of the parameter's historical values, respectively. These are the normalized parameter values; Step S2-3: Based on historical operating data, with system load rate, outdoor dry-bulb temperature, and outdoor wet-bulb temperature as the core dimensions, a clustering algorithm is used to group historical data with similar values ​​of the three into the same typical operating condition category, forming several typical operating condition categories; calculate the Euclidean distance between the currently collected operating parameter set and the core features of each typical operating condition category, and take the operating condition corresponding to the category with the smallest distance as the current operating condition. Step S2-4: Construct a random forest regression model with total system energy consumption as the target variable, and calculate the feature importance weights of each parameter using a recursive feature elimination algorithm. The feature importance weights The calculation formula is ,in The amount by which the impurity of the node in the j-th decision tree is reduced. The impurity of the node in the j-th decision tree is reduced by the amount of the i-th parameter removed, where n is the total number of decision trees. Step S2-5: Set the dynamic filtering threshold θ, calculated using the following formula: ,in As the baseline threshold, The current system load rate and , Assign system load weights; filter out feature importance weights. Parameters that are greater than the dynamic filtering threshold θ form a subset of key parameters.

[0009] Furthermore, step S3 includes: Step S3-1: Match the dynamically filtered subset of key parameters by equipment type and input it into the corresponding equipment-level model in the system energy consumption digital twin model, with a preset time step. Trigger dynamic model solving; Step S3-2: Calculate the theoretical power consumption of each device using a variable step-size iterative algorithm, with the step size ranging from [value missing]. to The theoretical total energy consumption of the system is obtained by summing the theoretical power consumption of each device, and then the continuous energy consumption is calculated. The theoretical energy consumption at each time step is averaged to output the theoretical optimal energy consumption baseline value under the current operating condition. ; Step S3-3: Collect the actual power consumption of each device in real time and sum them up to obtain the total measured energy consumption of the system. Time windows are aligned based on the timestamps of key parameter acquisition, and time synchronization errors exceeding a preset threshold are eliminated. The data was analyzed, and its validity was verified by deviation detection compared to historical data from the same period. The historical data from the same period refers to the historical measured total energy consumption dataset under the same operating conditions. The deviation detection threshold was [missing information]. ; Step S3-4: According to the formula Calculate the real-time energy consumption deviation rate, if If it is 0, then set =0; store the energy consumption deviation rate. and corresponding timestamps to form a time series .

[0010] Furthermore, step S4 includes: Step S4-1: Process the energy consumption deviation rate time series Exponentially weighted moving average smoothing is performed using a window size of m, with time-reverse indices q = 0, 1, ..., m−1, where q = 0 corresponds to the latest time t. The smoothing energy consumption deviation rate at time t is calculated using the following formula. : ,in This represents the energy consumption deviation rate at time tq; thus, the smoothed energy consumption deviation rate time series is obtained. ; Step S4-2: Based on historical duration All energy consumption deviation rate data within ,use Outlier removal is performed using principles or box plot methods to obtain a dataset of baseline deviation rates. The kernel density estimation method is used based on Calculate threshold benchmark ; Step S4-3: Analyze the system load rate characterizing the current operating condition. and outdoor combined temperature Normalization was performed separately to obtain the normalized load factor. and normalized outdoor composite temperature The operating condition adaptation factor was determined using a multiple linear regression model. The calculation formula is: ,in , , For regression coefficients; with a fixed period The interval is based on the time before the current moment. Historical data within a time window are used to determine the regression coefficients of the multiple linear regression model. , , ; and then according to the formula The operating condition adaptation factor for the current cycle is calculated. ; Step S4-4: Perform anomaly detection, including: Mutation anomaly detection: based on threshold benchmark Adaptability Factor to Operating Conditions The dynamic threshold T is calculated using the following formula: ; Calculate continuous Deviation slope at each time point ,when And |s| is greater than the preset slope threshold. When the mutation is abnormal, it is determined to be abnormal. Progressive anomaly detection: using long short-term memory neural networks to predict the future. Energy consumption deviation rate over time, when continuous The actual deviation rate showed a monotonically increasing trend over the duration and the predicted value exceeded [the expected value]. When it is determined to be a gradual anomaly, among which The preset prediction value multiple threshold; Periodic anomaly detection: Periodic features of the deviation rate time series are extracted using Fourier transform; when the mean deviation rate of a certain period is continuous... This time it exceeded the historical average of this period and When the product of is used, it is determined to be a periodic anomaly, where The threshold is set as a multiple of the preset periodic average. Step S4-5: Trigger tiered alerts based on anomaly detection results: Level 1 warning: Triggered when a sudden abnormality or a single-cycle abnormality is detected, triggering an audible and visual alarm through the local monitoring system; Level 2 warning: When a gradual or continuous anomaly is detected. A maintenance work order is generated when a periodic anomaly occurs. Warning cancellation conditions: continuous Duration, energy consumption deviation rate All values ​​are within the dynamic threshold T, and the absolute value of the slope |s| in the mutation anomaly detection is consistently lower than the preset slope threshold. The predicted value in progressive anomaly detection is consistently lower than The periodic mean deviation in periodic anomaly detection is consistently lower than the historical mean. The product of.

[0011] The system includes: a digital twin model construction module, a key parameter dynamic screening module, an energy consumption deviation calculation module, and a graded early warning triggering module; The digital twin model building module is used to establish a digital twin model of system energy consumption based on the initial calibration data of the air conditioning room. The key parameter dynamic filtering module is used to collect the set of operating parameters of the air conditioning room in real time, and dynamically filter a subset of key parameters from the set of operating parameters based on the current operating conditions and through feature importance evaluation. The energy consumption deviation calculation module is used to input the subset of key parameters into the system energy consumption digital twin model, calculate the theoretical optimal energy consumption under the current operating conditions, compare the theoretical optimal energy consumption with the measured total energy consumption, and generate the energy consumption deviation rate. The graded early warning triggering module is used to perform time series analysis on the energy consumption deviation rate. When the energy consumption deviation rate exceeds the dynamic threshold or shows a preset abnormal trend, a graded early warning is triggered.

[0012] The digital twin model construction module includes: An initial calibration data processing unit is used to collect historical operating data of the air conditioning room as initial calibration data. The initial calibration data includes at least environmental parameters, equipment operating parameters, system setting parameters, system load rate, and corresponding total energy consumption data. The environmental parameters include at least outdoor dry-bulb temperature and outdoor wet-bulb temperature. The collected data is cleaned, outlier processed, and standardized to form a dataset. The initial static feature selection and model architecture selection unit is used to select initial static feature variables that are strongly correlated with system energy consumption from the dataset; and to select the gradient boosting decision tree algorithm as the architecture of the system energy consumption digital twin model. The model training evaluation and tuning unit is used to train and evaluate the digital twin model of the system energy consumption using the dataset. The training data is randomly divided into k parts, and k-1 parts are used as the training set and the remaining part is used as the validation set. This process is repeated k times. The prediction accuracy and generalization ability of the model are evaluated using the k-fold cross-validation method. Based on the evaluation results, the hyperparameters are tuned until the model reaches the preset accuracy threshold. The baseline model determination unit is used to determine the validated and best-performing digital twin model as the baseline model, and the energy consumption prediction value of the baseline model under typical operating conditions is the theoretical optimal energy consumption baseline.

[0013] The key parameter dynamic filtering module includes: The real-time acquisition and timestamp alignment unit for operating parameters is used to collect parameters at preset time intervals. The system collects a set of operational parameters in real time, including environmental parameters, equipment operating parameters, and system status parameters; and aligns the data aggregation timestamps through an edge computing gateway. The operating parameter preprocessing unit is used to preprocess the operating parameter set, including using a Kalman filter algorithm to reduce noise in continuous parameters, using interpolation methods based on similar operating conditions to fill in missing values, marking and replacing outliers, and processing according to the formula. Normalization is performed, where x is the original parameter value. , These are the minimum and maximum values ​​of the parameter's historical values, respectively. These are the normalized parameter values; The operating condition category determination unit is used to group historical operating data with system load rate, outdoor dry bulb temperature and outdoor wet bulb temperature as the core dimensions, and to use a clustering algorithm to group historical data with similar values ​​of the three into the same typical operating condition category, forming several typical operating condition categories; calculate the Euclidean distance between the currently collected real-time operating parameter set and the core features of each typical operating condition category, and take the operating condition corresponding to the category with the smallest distance as the current operating condition. The feature importance calculation unit is used to construct a random forest regression model with total system energy consumption as the target variable, and calculates the feature importance weights of each parameter through a recursive feature elimination algorithm. The feature importance weights The calculation formula is ,in The amount by which the impurity of the node in the j-th decision tree is reduced. The impurity of the node in the j-th decision tree is reduced by the amount of the i-th parameter removed, where n is the total number of decision trees. The key parameter subset filtering unit is used to set the dynamic filtering threshold θ, and the calculation formula is as follows: ,in As the baseline threshold, The current system load rate and , Assign system load weights; filter out feature importance weights. Parameters that are greater than the dynamic filtering threshold θ form a subset of key parameters.

[0014] The energy consumption deviation calculation module includes: The key parameter input and model solution triggering unit is used to match the dynamically filtered subset of key parameters according to equipment type and input them into the corresponding equipment-level model in the system energy consumption digital twin model, at a preset time step. Trigger dynamic model solving; The theoretical optimal energy consumption calculation unit is used to calculate the theoretical power consumption of each device using a variable step-size iterative algorithm, with a step size range of [missing information]. to The theoretical total energy consumption of the system is obtained by summing the theoretical power consumption of each device, and then the continuous energy consumption is calculated. The theoretical energy consumption at each time step is averaged to output the theoretical optimal energy consumption baseline value under the current operating condition. ; The measured total energy consumption acquisition and verification unit is used to collect the actual power consumption of each device in real time and sum them up to obtain the measured total energy consumption of the system. Time windows are aligned based on the timestamps of key parameter acquisition, and time synchronization errors exceeding a preset threshold are eliminated. The data was analyzed, and its validity was verified by deviation detection compared to historical data from the same period. The historical data from the same period refers to the historical measured total energy consumption dataset under the same operating conditions. The deviation detection threshold was [missing information]. ; Energy consumption deviation rate calculation and storage unit, used to calculate and store energy consumption deviation rate according to formula Calculate the real-time energy consumption deviation rate, if If it is 0, then set =0; store the energy consumption deviation rate. and corresponding timestamps to form a time series .

[0015] The graded early warning triggering module includes: The energy consumption deviation rate smoothing unit is used to smooth the energy consumption deviation rate time series. Exponentially weighted moving average smoothing is performed using a window size of m, with time-reverse indices q = 0, 1, ..., m−1, where q = 0 corresponds to the latest time t. The smoothing energy consumption deviation rate at time t is calculated using the following formula. : ,in This represents the energy consumption deviation rate at time tq; thus, the smoothed energy consumption deviation rate time series is obtained. ; Threshold benchmark calculation unit, used for historical duration All energy consumption deviation rate data within ,use Outlier removal is performed using principles or box plot methods to obtain a dataset of baseline deviation rates. The kernel density estimation method is used based on Calculate threshold benchmark ; The operating condition adaptation factor calculation unit is used to calculate the system load rate, which characterizes the current operating condition. and outdoor combined temperature Normalization was performed separately to obtain the normalized load factor. and normalized outdoor composite temperature The operating condition adaptation factor was determined using a multiple linear regression model. The calculation formula is: ,in , , For regression coefficients; with a fixed period The interval is based on the time before the current moment. Historical data within a time window are used to determine the regression coefficients of the multiple linear regression model. , , ; and then according to the formula The operating condition adaptation factor for the current cycle is calculated. ; An anomaly detection unit, used for anomaly detection, includes: The tiered early warning execution unit is used to implement early warning based on threshold benchmarks. Adaptability Factor to Operating Conditions The dynamic threshold T is calculated using the following formula: ; Calculate continuous Deviation slope at each time point ,when And |s| is greater than the preset slope threshold. When the mutation is abnormal, it is determined to be abnormal. Progressive anomaly detection: using long short-term memory neural networks to predict the future. Energy consumption deviation rate over time, when continuous The actual deviation rate showed a monotonically increasing trend over the duration and the predicted value exceeded [the expected value]. When it is determined to be a gradual anomaly, among which The preset prediction value multiple threshold; Periodic anomaly detection: Periodic features of the deviation rate time series are extracted using Fourier transform; when the mean deviation rate of a certain period is continuous... This time it exceeded the historical average of this period and When the product of is used, it is determined to be a periodic anomaly, where The threshold is set as a multiple of the preset periodic average. The tiered early warning execution unit is used to trigger tiered early warnings based on anomaly detection results. Level 1 warning: Triggered when a sudden abnormality or a single-cycle abnormality is detected, triggering an audible and visual alarm through the local monitoring system; Level 2 warning: When a gradual or continuous anomaly is detected. A maintenance work order is generated when a periodic anomaly occurs. Warning cancellation conditions: continuous Duration, energy consumption deviation rate All values ​​are within the dynamic threshold T, and the absolute value of the slope |s| in the mutation anomaly detection is consistently lower than the preset slope threshold. The predicted value in progressive anomaly detection is consistently lower than The periodic mean deviation in periodic anomaly detection is consistently lower than the historical mean. The product of.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By introducing a dynamic feature selection mechanism, the accuracy and reliability of energy consumption monitoring are significantly improved. Traditional methods use a fixed set of parameters for analysis, which cannot adapt to the dynamic operating characteristics of air conditioning systems under different loads and environments. This invention, however, assesses the importance of each operating parameter to the current total energy consumption in real time and dynamically selects a subset of key parameters, ensuring the system always focuses on the most critical influencing factors. Simultaneously, a digital twin model based on a gradient boosting decision tree algorithm, rigorously validated, provides a more accurate theoretical optimal energy consumption baseline. The combination of these two aspects makes the theoretical energy consumption calculation and actual energy consumption deviation analysis more closely reflect the actual operating state of the system, laying a solid and reliable foundation for energy efficiency assessment.

[0017] 2. It achieves self-adaptation and intelligence throughout the entire process of air conditioning energy consumption management. From dynamic screening of key parameters and real-time calculation of theoretical energy consumption to dynamic adjustment of early warning thresholds, the entire management process does not rely on human experience to set fixed rules. It can automatically adjust and optimize according to changes in the external environment and internal load. This highly adaptive capability gives the management system judgment similar to expert experience, greatly reducing reliance on human intervention and promoting the transformation of energy consumption management from a static, experience-based model to a dynamic, data-driven intelligent model, ensuring that the system can operate continuously and efficiently under different operating conditions.

[0018] 3. Through tiered early warning and precise location tracking, the efficiency and targeted nature of operations and maintenance management are effectively optimized. The system triggers different levels of early warning based on the type, severity, and persistence of anomalies, helping the operations and maintenance team quickly prioritize faults and allocate maintenance resources rationally. More importantly, since energy consumption deviations are calculated based on current key parameters, these parameters themselves provide a clear direction for quickly locating the root cause of problems, significantly shortening fault diagnosis time, avoiding blind troubleshooting, and thus improving the overall efficiency of operations and maintenance work. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the air conditioning energy consumption adaptive management system based on dynamic feature selection according to the present invention; Figure 2 This is a schematic diagram of the key parameter dynamic screening module of the air conditioning energy consumption adaptive management system based on dynamic feature selection of the present invention; Figure 3 This is a flowchart illustrating the adaptive energy consumption management method for air conditioning based on dynamic feature selection according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Figure 1 , Figure 2 and Figure 3 As shown, the present invention provides a technical solution: an adaptive air conditioning energy consumption management system and method based on dynamic feature selection. The central air conditioning room of a commercial complex is equipped with 3 centrifugal chillers, 4 chilled water pumps, 4 cooling water pumps and 2 cooling towers. The system needs to meet the year-round temperature and humidity control requirements of the shopping mall and office area. The goal is to achieve adaptive energy consumption management and anomaly early warning through this method, and reduce ineffective energy consumption.

[0022] The methods include: Step S1: Based on the initial calibration data of the air conditioning room, establish a digital twin model of system energy consumption; Step S2: Collect the set of operating parameters of the air conditioning room in real time, and dynamically filter the subset of key parameters from the set of operating parameters based on the current operating conditions and through feature importance assessment; Step S3: Input the subset of key parameters into the system energy consumption digital twin model to calculate the theoretical optimal energy consumption under the current operating conditions; compare the theoretical optimal energy consumption with the measured total energy consumption to generate the energy consumption deviation rate; Step S4: Perform time series analysis on the energy consumption deviation rate. When the energy consumption deviation rate exceeds the dynamic threshold or shows a preset abnormal trend, trigger a graded early warning.

[0023] Furthermore, step S1 includes: Step S1-1: First, collect historical operating data for the air conditioning room during winter heating (January-February), transitional season ventilation (April-May), and summer cooling (July-August) of a given year as initial calibration data. Specifically, this includes: environmental parameters such as outdoor dry-bulb temperature -5℃~40℃, outdoor wet-bulb temperature 0℃~32℃, indoor average temperature 20℃~26℃, and indoor average humidity 40%~65%; and equipment operating parameters such as chiller COP 2.8~5.2 and chilled water pump outlet pressure 0.6~0. The system parameters include: 8MPa pressure, cooling water pump flow rate 450~600m³ / h, cooling tower fan speed 800~1500r / min; system settings include chilled water supply and return temperature difference setpoint of 5℃, cooling water supply and return temperature difference setpoint of 4℃, and unit load rate setpoint range of 30%~100%; total energy consumption data includes the main meter reading in the computer room (kWh) and the sub-meter readings of each equipment; the collected data is preprocessed by first deleting invalid samples with a missing rate greater than 10%, and then using the IQR method to remove samples smaller than 10%. -1.5 IQR or greater Outliers with a value of +1.5 IQR were finally standardized using the Z-score method, and the mean was calculated. Standard deviation Standardized data This resulted in a dataset containing 18,200 valid samples. Step S1-2: To screen out variables that significantly affect energy consumption, the correlation between each parameter and the total system energy consumption is calculated using Pearson correlation coefficient analysis. Parameters with an absolute correlation coefficient greater than 0.6 are selected as initial static feature variables, ultimately determining six variables: outdoor dry-bulb temperature, outdoor wet-bulb temperature, chiller unit load rate, chilled water supply and return temperature difference, cooling water pump flow rate, and cooling tower fan speed. Considering the nonlinear relationship between air conditioning energy consumption and parameters, the gradient boosting decision tree algorithm is selected as the architecture of the system energy consumption digital twin model. The base learner uses a CART tree to better fit the nonlinear relationship, and the loss function is the mean squared error to minimize the deviation between the predicted and actual energy consumption values. Steps S1-3: Divide the dataset into a training set of 12740 records and a test set of 5460 records in a 7:3 ratio. Use 5-fold cross-validation to evaluate the model: divide the training set into 5 parts, with 4 parts as sub-training sets and 1 part as a sub-validation set, repeating this process 5 times to calculate the average prediction accuracy. Optimize hyperparameters using a grid search method, with the search range being a learning rate of 0.01~0.1, tree depth of 3~8, and number of trees of 100~500. The optimal hyperparameters are determined to be: a learning rate of 0.05, a tree depth of 5, and 300 trees. At this point, the model's mean absolute error on the test set is less than 3%, meeting the preset accuracy threshold. Step S1-4: The GBDT model that has been optimized by hyperparameters and meets the accuracy requirements is determined as the digital twin benchmark model for system energy consumption. This model can output the predicted energy consumption value under the corresponding operating condition based on the input operating condition parameters. This predicted value is the "theoretical optimal energy consumption baseline" under the operating condition, which serves as the core reference standard for subsequent calculation of energy consumption deviation and judgment of whether energy consumption is abnormal.

[0024] Furthermore, step S2 includes: Step S2-1: To obtain the real-time operating status of the computer room, set the data collection time interval. =5 minutes, the system collects a set of operating parameters in real time through sensors deployed in the computer room, including temperature sensors, pressure sensors, flow sensors, and smart meters. These parameters include: environmental parameters such as outdoor dry-bulb temperature and outdoor wet-bulb temperature; equipment operating parameters such as chiller COP, chilled water pump outlet pressure, cooling water pump flow rate, cooling tower fan speed, and real-time power of each device; and system status parameters such as chilled water supply and return temperature difference, cooling water supply and return temperature difference, and system total load rate. The timestamps of all parameters are aligned through an edge computing gateway to ensure that the time deviation of all parameters within the same collection period is less than 1 second.

[0025] Step S2-2: Preprocess the real-time collected operating parameter set. For continuous parameters such as outdoor temperature and water pump flow, Kalman filtering algorithm is used for noise reduction, with the process noise covariance Q set to 0.01 and the observation noise covariance R set to 0.1. For occasional missing parameters with a missing rate of less than 5%, the "same-condition interpolation method" is used to fill in the missing data. When the current condition is high load in summer, linear interpolation is performed based on the data of the same condition in the past hour. For outliers such as chilled water supply and return water temperature difference greater than 8℃ that are outside the physical range, the average value of the past 3 collection periods is replaced. The data is then processed according to the formula. All parameters were normalized, and historical values ​​of "outdoor dry bulb temperature" were used. -5℃ If the temperature is 40℃ and the current temperature is 25℃, then after normalization... =0.67; Step S2-3: Based on historical operating data from January to August, the K-means clustering algorithm is used to classify operating conditions: the number of clusters k=3 is determined by the elbow rule, and typical operating conditions are divided into three categories: low load, medium load, and high load, as shown in Table 1 below: specific parameters are described in the spatial region. Table 1: Spatial Region of Specific Parameters

[0026] Calculate the Euclidean distances between the current set of operating parameters (system load rate 55%, outdoor dry-bulb temperature 26℃, outdoor wet-bulb temperature 22℃) and the cluster centers of the three operating conditions. The distance to the cluster center of operating condition B is 0.12, indicating that the current condition belongs to "medium load in operating condition B".

[0027] Step S2-4: Using "total system energy consumption" as the target variable, construct a random forest regression model decision tree with a number of nodes n set to 100. The Gini coefficient is used as the node impurity index. Calculate the feature importance weights of each parameter using a recursive feature elimination algorithm. The calculation formula is: ,in The amount by which the impurity of the node in the j-th decision tree is reduced. The reduction in node impurity in the j-th decision tree after removing the i-th parameter, where n is the total number of decision trees; "Chiller load rate" =42, then its feature importance weight = =0.42; Similarly, the weights of other parameters are calculated as follows: the importance weight of outdoor dry-bulb temperature characteristic is 0.38, the importance weight of chilled water supply and return water temperature difference characteristic is 0.35, the importance weight of cooling water pump flow rate characteristic is 0.22, the importance weight of cooling tower fan speed characteristic is 0.18, and the importance weight of outdoor wet-bulb temperature characteristic is 0.15; Step S2-5: Set the baseline threshold =0.3, system load weight l=0.6, current system load rate =0.55, according to the formula Calculate the dynamic screening threshold: θ = 0.3 × (1 - 0.6 × 0.55) = 0.3 × 0.67 = 0.201; then select the feature importance weights. For parameters >0.201, the final subset of key parameters was determined as follows: chiller load rate characteristic importance weight 0.42, outdoor dry bulb temperature characteristic importance weight 0.38, chilled water supply and return temperature difference characteristic importance weight 0.35, and cooling water pump flow rate characteristic importance weight 0.22.

[0028] Furthermore, step S3 includes: Step S3-1: Match the selected subset of key parameters to the "equipment-level sub-models" of the baseline digital twin model according to equipment type: that is, match the chiller unit load rate to the chiller unit energy consumption sub-model; match the chilled water supply and return water temperature difference to the chilled water pump energy consumption sub-model; match the cooling water pump flow rate to the cooling water pump energy consumption model; and match the outdoor dry-bulb temperature to the cooling tower fan energy consumption model. =The model is dynamically solved with a time step of 5 minutes, and the theoretical energy consumption prediction value of each device is output every 5 minutes; Step S3-2: Calculate the theoretical power consumption of each device using a variable step-size iterative algorithm. The step size range of the variable step-size iterative algorithm is set as follows: =1 minute to =3 minutes, with a 1-minute step size for high-load conditions and a 3-minute step size for low-load conditions. The current load in condition B uses a 2-minute step size. Specific equipment: The theoretical power consumption of 3 chiller units is 85kW, 82kW, and 80kW respectively, totaling 247kW; the theoretical power consumption of 4 chilled water pumps is 15kW, 15kW, 14kW, and 14kW respectively, totaling 58kW; the theoretical power consumption of 4 cooling water pumps is 18kW, 18kW, 17kW, and 17kW respectively, totaling 70kW; the theoretical power consumption of 2 cooling tower fans is 12kW and 12kW respectively, totaling 24kW. The theoretical total energy consumption of the system = 247 + 58 + 70 + 24 = 399kW. A moving average is applied to the theoretical energy consumption over 12 consecutive time steps (1 hour) to finally output the theoretical optimal energy consumption baseline value under the current condition. It is 395kW; Step S3-3: Collect and sum the actual power consumption of each device in real time using smart meters to obtain the total measured energy consumption of the system. =410kW; Time window alignment is performed based on the key parameter acquisition timestamps, and time synchronization errors greater than 100kW are eliminated. =30 seconds of data; retrieve historical data from the same date, time period, and operating condition B; the historical measured total energy consumption dataset ranges from 385kW to 405kW. Current... =410kW, the deviation from the historical average of 395kW is ≈3.8%, which is less than the preset deviation detection threshold ±5%, so the data is considered valid; Step S3-4: According to the formula Calculate the real-time energy consumption deviation rate: ;like In extreme downtime conditions, δ is set to 0; the storage energy consumption deviation rate δ and the corresponding energy consumption deviation rate of 3.8% for the corresponding timestamp are used to form a time series. The percentages are 3.2%, 2.9%, 3.8%, 4.1%, 3.5%, and so on.

[0029] Furthermore, step S4 includes: Step S4-1: For Exponentially weighted moving average smoothing is used, with a window size m of 8, and the time indices q = 0, 1...7, where q = 0 corresponds to the latest time t. The calculation formula is as follows: With the latest 8 That is, 3.5%, 4.1%, 3.8%, 2.9%, 3.2%, 3.0%, 2.8%, and 2.7%, respectively, to calculate the smoothed time series. The percentages are 5.2%, 5.5%, 5.74%, and so on.

[0030] Step S4-2: Select historical duration Set to all within 30 days The dataset contains 8640 data points. Outliers were removed using the 3σ principle, where σ represents the standard deviation of the smoothed series and μ represents the mean of the smoothed series. Data points smaller than μ-3σ or greater than μ+3σ were removed, resulting in a baseline deviation rate dataset of 8520 data points. The kernel density estimation method was used to... Fit the probability density curve and take the deviation rate value corresponding to the maximum probability density as the threshold benchmark. It is 6%; Step S4-3: First, normalize the key operating parameters: the system load rate α is 55%, after normalization... Outdoor composite temperature According to the formula Calculation, where 26℃ The final temperature was 22℃. =26℃, based on the annual operation data of the data center, and the average outdoor temperature. The historical value range is 10℃~40℃, obtained =10℃, =40℃, and finally obtained the normalized result. Set update cycle =1 day, based on the current time before Regression coefficients were obtained by fitting a multiple linear regression model with 7 days of historical data. =0.08、 =0.55、 =0.32; according to the formula Calculate the working condition adaptation factor, and substitute the data to get γ=0.08+0.55×0.55+0.32×0.53≈0.08+0.3025+0.1696≈0.5521; Step S4-4: Anomaly Detection Mutation Anomaly Detection: Dynamic Threshold =6%×(1+0.5521)≈9.31%; Set =3, calculate the slope of the deviation rate at three consecutive time points. Substituting the data yields ≈0.27%, preset slope threshold =1%, because =5.74% is less than T and |s|=0.27% is less than It is not considered a mutation anomaly; Progressive anomaly detection: using a long short-term memory neural network to predict future duration. The energy consumption deviation rate is set to 24 hours, and the model input is the energy consumption deviation rate of the past 7 days. Output predicted values; set =6 hours, if 6 consecutive hours actually It shows a monotonically increasing trend, i.e., 5.2% → 5.5% → 5.74% → 5.9% → 6.1% → 6.3%, and is predicted to continue for 24 hours. =7.5%, preset =1.2, calculate T× ≈9.31%×1.2≈11.17%, since the predicted value of 7.5% is less than 11.17%, it is not judged as an asymptotic anomaly; Periodic anomaly detection: Extraction via Fourier transform Based on the cyclical characteristics, it was found that the main cycle is 24 hours, consistent with the mall's operating hours; [the following was set...] =2, if the average deviation rate of two consecutive cycles, the average of the first cycle (5.8%) and the average of the second cycle (6.0%), exceeds the historical average of 5.5% for that cycle. =1.1 product, pre-set =1.1, the historical average for this period is 5.5%, calculate the historical average and The product is 5.5% × 1.1 = 6.05%. Since 6.0% is less than 6.05%, it is not considered a periodic anomaly. Step S4-5: Tiered Early Warning and Lifting Warning Trigger: No warning will be triggered if no mutation, gradual abnormality, or periodic abnormality is detected; however, if a "continuous" abnormality is subsequently detected... = "2 periodic anomalies", with two-day periodic averages of 6.1% and 6.2% respectively, both greater than 6.05%, triggering a level 2 warning: generate a maintenance work order; Warning Lift: The duration of continuous time in the warning lifting conditions Set to 4 hours, to satisfy the requirement of 4 consecutive hours If all values ​​are within T (i.e., less than 9.31%), the slope |s| is less than 1%, the predicted value is less than 11.17%, and the periodic average is less than 6.05%, the warning will be automatically lifted.

[0031] The system includes: a digital twin model construction module, a key parameter dynamic screening module, an energy consumption deviation calculation module, and a graded early warning triggering module; The digital twin model building module is used to establish a digital twin model of system energy consumption based on the initial calibration data of the air conditioning room. The key parameter dynamic filtering module is used to collect the set of operating parameters of the air conditioning room in real time, and dynamically filter a subset of key parameters from the set of operating parameters based on the current operating conditions and through feature importance evaluation. The energy consumption deviation calculation module is used to input the subset of key parameters into the system energy consumption digital twin model, calculate the theoretical optimal energy consumption under the current operating conditions, compare the theoretical optimal energy consumption with the measured total energy consumption, and generate the energy consumption deviation rate. The graded early warning triggering module is used to perform time series analysis on the energy consumption deviation rate. When the energy consumption deviation rate exceeds the dynamic threshold or shows a preset abnormal trend, a graded early warning is triggered.

[0032] The digital twin model construction module includes: An initial calibration data processing unit is used to collect historical operating data of the air conditioning room as initial calibration data. The initial calibration data includes at least environmental parameters, equipment operating parameters, system setting parameters, system load rate, and corresponding total energy consumption data. The environmental parameters include at least outdoor dry-bulb temperature and outdoor wet-bulb temperature. The collected data is cleaned, outlier processed, and standardized to form a dataset. The initial static feature selection and model architecture selection unit is used to select initial static feature variables that are strongly correlated with system energy consumption from the dataset; and to select the gradient boosting decision tree algorithm as the architecture of the system energy consumption digital twin model. The model training evaluation and tuning unit is used to train and evaluate the digital twin model of the system energy consumption using the dataset. The training data is randomly divided into k parts, and k-1 parts are used as the training set and the remaining part is used as the validation set. This process is repeated k times. The prediction accuracy and generalization ability of the model are evaluated using the k-fold cross-validation method. Based on the evaluation results, the hyperparameters are tuned until the model reaches the preset accuracy threshold. The baseline model determination unit is used to determine the validated and best-performing digital twin model as the baseline model, and the energy consumption prediction value of the baseline model under typical operating conditions is the theoretical optimal energy consumption baseline.

[0033] The key parameter dynamic filtering module includes: The real-time acquisition and timestamp alignment unit for operating parameters is used to collect parameters at preset time intervals. The system collects a set of operational parameters in real time, including environmental parameters, equipment operating parameters, and system status parameters; and aligns the data aggregation timestamps through an edge computing gateway. The operating parameter preprocessing unit is used to preprocess the operating parameter set, including using a Kalman filter algorithm to reduce noise in continuous parameters, using interpolation methods based on similar operating conditions to fill in missing values, marking and replacing outliers, and processing according to the formula. Normalization is performed, where x is the original parameter value. , These are the minimum and maximum values ​​of the parameter's historical values, respectively. These are the normalized parameter values; The operating condition category determination unit is used to group historical operating data with system load rate, outdoor dry bulb temperature and outdoor wet bulb temperature as the core dimensions, and to use a clustering algorithm to group historical data with similar values ​​of the three into the same typical operating condition category, forming several typical operating condition categories; calculate the Euclidean distance between the currently collected real-time operating parameter set and the core features of each typical operating condition category, and take the operating condition corresponding to the category with the smallest distance as the current operating condition. The feature importance calculation unit is used to construct a random forest regression model with total system energy consumption as the target variable, and calculates the feature importance weights of each parameter through a recursive feature elimination algorithm. The feature importance weights The calculation formula is ,in The amount by which the impurity of the node in the j-th decision tree is reduced. The impurity of the node in the j-th decision tree is reduced by the amount of the i-th parameter removed, where n is the total number of decision trees. The key parameter subset filtering unit is used to set the dynamic filtering threshold θ, and the calculation formula is as follows: ,in As the baseline threshold, The current system load rate and , Assign system load weights; filter out feature importance weights. Parameters that are greater than the dynamic filtering threshold θ form a subset of key parameters.

[0034] The energy consumption deviation calculation module includes: The key parameter input and model solution triggering unit is used to match the dynamically filtered subset of key parameters according to equipment type and input them into the corresponding equipment-level model in the system energy consumption digital twin model, at a preset time step. Trigger dynamic model solving; The theoretical optimal energy consumption calculation unit is used to calculate the theoretical power consumption of each device using a variable step-size iterative algorithm, with a step size range of [missing information]. to The theoretical total energy consumption of the system is obtained by summing the theoretical power consumption of each device, and then the continuous energy consumption is calculated. The theoretical energy consumption at each time step is averaged to output the theoretical optimal energy consumption baseline value under the current operating condition. ; The measured total energy consumption acquisition and verification unit is used to collect the actual power consumption of each device in real time and sum them up to obtain the measured total energy consumption of the system. Time windows are aligned based on the timestamps of key parameter acquisition, and time synchronization errors exceeding a preset threshold are eliminated. The data was analyzed, and its validity was verified by deviation detection compared to historical data from the same period. The historical data from the same period refers to the historical measured total energy consumption dataset under the same operating conditions. The deviation detection threshold was [missing information]. ; Energy consumption deviation rate calculation and storage unit, used to calculate and store energy consumption deviation rate according to formula Calculate the real-time energy consumption deviation rate, if If it is 0, then set =0; store the energy consumption deviation rate. and corresponding timestamps to form a time series .

[0035] The graded early warning triggering module includes: The energy consumption deviation rate smoothing unit is used to smooth the energy consumption deviation rate time series. Exponentially weighted moving average smoothing is performed using a window size of m, with time-reverse indices q = 0, 1, ..., m−1, where q = 0 corresponds to the latest time t. The smoothing energy consumption deviation rate at time t is calculated using the following formula. : ,in This represents the energy consumption deviation rate at time tq; thus, the smoothed energy consumption deviation rate time series is obtained. ; Threshold benchmark calculation unit, used for historical duration All energy consumption deviation rate data within ,use Outlier removal is performed using principles or box plot methods to obtain a dataset of baseline deviation rates. The kernel density estimation method is used based on Calculate threshold benchmark ; The operating condition adaptation factor calculation unit is used to calculate the system load rate, which characterizes the current operating condition. and outdoor combined temperature Normalization was performed separately to obtain the normalized load factor. and normalized outdoor composite temperature The operating condition adaptation factor was determined using a multiple linear regression model. The calculation formula is: ,in , , For regression coefficients; with a fixed period The interval is based on the time before the current moment. Historical data within a time window are used to determine the regression coefficients of the multiple linear regression model. , , ; and then according to the formula The operating condition adaptation factor for the current cycle is calculated. ; An anomaly detection unit, used for anomaly detection, includes: The tiered early warning execution unit is used to implement early warning based on threshold benchmarks. Adaptability Factor to Operating Conditions The dynamic threshold T is calculated using the following formula: ; Calculate continuous Deviation slope at each time point ,when And |s| is greater than the preset slope threshold. When the mutation is abnormal, it is determined to be abnormal. Progressive anomaly detection: using long short-term memory neural networks to predict the future. Energy consumption deviation rate over time, when continuous The actual deviation rate showed a monotonically increasing trend over the duration and the predicted value exceeded [the expected value]. When it is determined to be a gradual anomaly, among which The preset prediction value multiple threshold; Periodic anomaly detection: Periodic features of the deviation rate time series are extracted using Fourier transform; when the mean deviation rate of a certain period is continuous... This time it exceeded the historical average of this period and When the product of is used, it is determined to be a periodic anomaly, where The threshold is set as a multiple of the preset periodic average. The tiered early warning execution unit is used to trigger tiered early warnings based on anomaly detection results. Level 1 warning: Triggered when a sudden abnormality or a single-cycle abnormality is detected, triggering an audible and visual alarm through the local monitoring system; Level 2 warning: When a gradual or continuous anomaly is detected. A maintenance work order is generated when a periodic anomaly occurs. Warning cancellation conditions: continuous Duration, energy consumption deviation rate All values ​​are within the dynamic threshold T, and the absolute value of the slope |s| in the mutation anomaly detection is consistently lower than the preset slope threshold. The predicted value in progressive anomaly detection is consistently lower than The periodic mean deviation in periodic anomaly detection is consistently lower than the historical mean. The product of.

[0036] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An adaptive energy consumption management method for air conditioning based on dynamic feature selection, characterized in that: Includes the following steps: Step S1: Based on the initial calibration data of the air conditioning room, establish a digital twin model of system energy consumption; Step S2: Collect the set of operating parameters of the air conditioning room in real time, and dynamically filter the subset of key parameters from the set of operating parameters based on the current operating conditions and through feature importance assessment; Step S3: Input the subset of key parameters into the system energy consumption digital twin model to calculate the theoretical optimal energy consumption under the current operating conditions; compare the theoretical optimal energy consumption with the measured total energy consumption to generate the energy consumption deviation rate; Step S4: Perform time series analysis on the energy consumption deviation rate. When the energy consumption deviation rate exceeds the dynamic threshold or shows a preset abnormal trend, trigger a graded early warning.

2. The air conditioning energy consumption adaptive management method based on dynamic feature selection according to claim 1, characterized in that: Step S1 includes: Step S1-1: Collect historical operating data of the air conditioning room as initial calibration data; the initial calibration data includes at least environmental parameters, equipment operating parameters, system setting parameters, system load rate, and corresponding total energy consumption data; the environmental parameters include at least outdoor dry-bulb temperature and outdoor wet-bulb temperature; clean the collected data, handle outliers, and standardize it to form a dataset; Step S1-2: Select initial static feature variables that are strongly correlated with system energy consumption from the dataset; and select the gradient boosting decision tree algorithm as the architecture of the system energy consumption digital twin model; Step S1-3: Use the dataset to train and evaluate the digital twin model of the system energy consumption; randomly divide the training data into k parts, use k-1 parts as the training set and the remaining part as the validation set, repeat k times, and use this k-fold cross-validation method to evaluate the prediction accuracy and generalization ability of the model, and perform hyperparameter tuning based on the evaluation results until the model reaches the preset accuracy threshold. Step S1-4: The validated and best-performing digital twin model is determined as the baseline model, and the energy consumption prediction value of the baseline model under typical operating conditions is the theoretical optimal energy consumption baseline.

3. The air conditioning energy consumption adaptive management method based on dynamic feature selection according to claim 1, characterized in that: Step S2 includes: Step S2-1: At a preset data collection time interval The system collects a set of operational parameters in real time, including environmental parameters, equipment operating parameters, and system status parameters; and aligns the data aggregation timestamps through an edge computing gateway. Step S2-2: Preprocess the set of operating parameters, including using the Kalman filter algorithm to reduce noise in continuous parameters, using interpolation methods based on similar operating conditions to fill in missing values, marking and replacing outliers, and applying the formula... Normalization is performed, where x is the original parameter value. , These are the minimum and maximum values ​​of the parameter's historical values, respectively. These are the normalized parameter values; Step S2-3: Based on historical operating data, with system load rate, outdoor dry-bulb temperature, and outdoor wet-bulb temperature as the core dimensions, a clustering algorithm is used to group historical data with similar values ​​of the three into the same typical operating condition category, forming several typical operating condition categories; calculate the Euclidean distance between the currently collected operating parameter set and the core features of each typical operating condition category, and take the operating condition corresponding to the category with the smallest distance as the current operating condition. Step S2-4: Construct a random forest regression model with total system energy consumption as the target variable, and calculate the feature importance weights of each parameter using a recursive feature elimination algorithm. The feature importance weights The calculation formula is ,in The amount by which the impurity of the node in the j-th decision tree is reduced. The impurity of the node in the j-th decision tree is reduced by the amount of the i-th parameter removed, where n is the total number of decision trees. Step S2-5: Set the dynamic filtering threshold θ, calculated using the following formula: ,in As the baseline threshold, The current system load rate and , Assign system load weights; filter out feature importance weights. Parameters that are greater than the dynamic filtering threshold θ form a subset of key parameters.

4. The air conditioning energy consumption adaptive management method based on dynamic feature selection according to claim 1, characterized in that: Step S3 includes: Step S3-1: Match the dynamically filtered subset of key parameters by equipment type and input it into the corresponding equipment-level model in the system energy consumption digital twin model, with a preset time step. Trigger dynamic model solving; Step S3-2: Calculate the theoretical power consumption of each device using a variable step-size iterative algorithm, with the step size ranging from [value missing]. to The theoretical total energy consumption of the system is obtained by summing the theoretical power consumption of each device, and then the continuous energy consumption is calculated. The theoretical energy consumption at each time step is averaged to output the theoretical optimal energy consumption baseline value under the current operating condition. ; Step S3-3: Collect the actual power consumption of each device in real time and sum them up to obtain the total measured energy consumption of the system. Time windows are aligned based on the timestamps of key parameter acquisition, and time synchronization errors exceeding a preset threshold are eliminated. The data was analyzed, and its validity was verified by deviation detection compared to historical data from the same period. The historical data from the same period refers to the historical measured total energy consumption dataset under the same operating conditions. The deviation detection threshold was [missing information]. ; Step S3-4: According to the formula Calculate the real-time energy consumption deviation rate, if If it is 0, then set =0; store the energy consumption deviation rate. and corresponding timestamps to form a time series .

5. The air conditioning energy consumption adaptive management method based on dynamic feature selection according to claim 1, characterized in that: Step S4 includes: Step S4-1: Process the energy consumption deviation rate time series Exponentially weighted moving average smoothing is performed using a window size of m, with time-reverse indices q = 0, 1, ..., m−1, where q = 0 corresponds to the latest time t. The smoothing energy consumption deviation rate at time t is calculated using the following formula. : ,in This represents the energy consumption deviation rate at time tq; thus, the smoothed energy consumption deviation rate time series is obtained. ; Step S4-2: Based on historical duration All energy consumption deviation rate data within ,use Outlier removal is performed using principles or box plot methods to obtain a dataset of baseline deviation rates. The kernel density estimation method is used based on Calculate threshold benchmark ; Step S4-3: Analyze the system load rate characterizing the current operating condition. and outdoor combined temperature Normalization was performed separately to obtain the normalized load factor. and normalized outdoor composite temperature The operating condition adaptation factor was determined using a multiple linear regression model. The calculation formula is: ,in , , For regression coefficients; with a fixed period The interval is based on the time before the current moment. Historical data within a time window are used to determine the regression coefficients of the multiple linear regression model. , , ; and then according to the formula The operating condition adaptation factor for the current cycle is calculated. ; Step S4-4: Perform anomaly detection, including: Mutation anomaly detection: based on threshold benchmark Adaptability Factor to Operating Conditions The dynamic threshold T is calculated using the following formula: ; Calculate continuous Deviation slope at each time point ,when And |s| is greater than the preset slope threshold. When the mutation is abnormal, it is determined to be abnormal. Progressive anomaly detection: using long short-term memory neural networks to predict the future. Energy consumption deviation rate over time, when continuous The actual deviation rate showed a monotonically increasing trend over the duration and the predicted value exceeded [the expected value]. When it is determined to be a gradual anomaly, among which The preset prediction value multiple threshold; Periodic anomaly detection: Periodic features of the deviation rate time series are extracted using Fourier transform; when the mean deviation rate of a certain period is continuous... This time it exceeded the historical average of this period and When the product of is used, it is determined to be a periodic anomaly, where The threshold is set as a multiple of the preset periodic average. Step S4-5: Trigger tiered alerts based on anomaly detection results: Level 1 warning: Triggered when a sudden abnormality or a single-cycle abnormality is detected, triggering an audible and visual alarm through the local monitoring system; Level 2 warning: When a gradual or continuous anomaly is detected. A maintenance work order is generated when a periodic anomaly occurs. Warning cancellation conditions: continuous Duration, energy consumption deviation rate All values ​​are within the dynamic threshold T, and the absolute value of the slope |s| in the mutation anomaly detection is consistently lower than the preset slope threshold. The predicted value in progressive anomaly detection is consistently lower than The periodic mean deviation in periodic anomaly detection is consistently lower than the historical mean. The product of.

6. An air conditioning energy consumption adaptive management system based on dynamic feature selection, used to execute the air conditioning energy consumption adaptive management method based on dynamic feature selection as described in any one of claims 1-5, characterized in that: The system includes: The module includes a digital twin model construction module, a key parameter dynamic screening module, an energy consumption deviation calculation module, and a graded early warning triggering module. The digital twin model building module is used to establish a digital twin model of system energy consumption based on the initial calibration data of the air conditioning room. The key parameter dynamic filtering module is used to collect the set of operating parameters of the air conditioning room in real time, and dynamically filter a subset of key parameters from the set of operating parameters based on the current operating conditions and through feature importance evaluation. The energy consumption deviation calculation module is used to input the subset of key parameters into the system energy consumption digital twin model, calculate the theoretical optimal energy consumption under the current operating conditions, compare the theoretical optimal energy consumption with the measured total energy consumption, and generate the energy consumption deviation rate. The graded early warning triggering module is used to perform time series analysis on the energy consumption deviation rate. When the energy consumption deviation rate exceeds the dynamic threshold or shows a preset abnormal trend, a graded early warning is triggered.

7. The air conditioning energy consumption adaptive management system based on dynamic feature selection according to claim 6, characterized in that: The digital twin model construction module includes: An initial calibration data processing unit is used to collect historical operating data of the air conditioning room as initial calibration data. The initial calibration data includes at least environmental parameters, equipment operating parameters, system setting parameters, system load rate, and corresponding total energy consumption data. The environmental parameters include at least outdoor dry-bulb temperature and outdoor wet-bulb temperature. The collected data is cleaned, outlier processed, and standardized to form a dataset. The initial static feature selection and model architecture selection unit is used to select initial static feature variables that are strongly correlated with system energy consumption from the dataset; and to select the gradient boosting decision tree algorithm as the architecture of the system energy consumption digital twin model. The model training evaluation and tuning unit is used to train and evaluate the digital twin model of the system energy consumption using the dataset. The training data is randomly divided into k parts, and k-1 parts are used as the training set and the remaining part is used as the validation set. This process is repeated k times. The prediction accuracy and generalization ability of the model are evaluated using the k-fold cross-validation method. Based on the evaluation results, the hyperparameters are tuned until the model reaches the preset accuracy threshold. The baseline model determination unit is used to determine the validated and best-performing digital twin model as the baseline model, and the energy consumption prediction value of the baseline model under typical operating conditions is the theoretical optimal energy consumption baseline.

8. The air conditioning energy consumption adaptive management system based on dynamic feature selection according to claim 6, characterized in that: The key parameter dynamic filtering module includes: The real-time acquisition and timestamp alignment unit for operating parameters is used to collect parameters at preset time intervals. The system collects a set of operational parameters in real time, including environmental parameters, equipment operating parameters, and system status parameters; and aligns the data aggregation timestamps through an edge computing gateway. The operating parameter preprocessing unit is used to preprocess the operating parameter set, including using a Kalman filter algorithm to reduce noise in continuous parameters, using interpolation methods based on similar operating conditions to fill in missing values, marking and replacing outliers, and processing according to the formula. Normalization is performed, where x is the original parameter value. , These are the minimum and maximum values ​​of the parameter's historical values, respectively. These are the normalized parameter values; The operating condition category determination unit is used to group historical operating data with system load rate, outdoor dry bulb temperature and outdoor wet bulb temperature as the core dimensions, and to use a clustering algorithm to group historical data with similar values ​​of the three into the same typical operating condition category, forming several typical operating condition categories; calculate the Euclidean distance between the currently collected real-time operating parameter set and the core features of each typical operating condition category, and take the operating condition corresponding to the category with the smallest distance as the current operating condition. The feature importance calculation unit is used to construct a random forest regression model with total system energy consumption as the target variable, and calculates the feature importance weights of each parameter through a recursive feature elimination algorithm. The feature importance weights The calculation formula is ,in The amount by which the impurity of the node in the j-th decision tree is reduced. The impurity of the node in the j-th decision tree is reduced by the amount of the i-th parameter removed, where n is the total number of decision trees. The key parameter subset filtering unit is used to set the dynamic filtering threshold θ, and the calculation formula is as follows: ,in As the baseline threshold, The current system load rate and , Assign system load weights; filter out feature importance weights. Parameters that are greater than the dynamic filtering threshold θ form a subset of key parameters.

9. The air conditioning energy consumption adaptive management system based on dynamic feature selection according to claim 6, characterized in that: The energy consumption deviation calculation module includes: The key parameter input and model solution triggering unit is used to match the dynamically filtered subset of key parameters according to equipment type and input them into the corresponding equipment-level model in the system energy consumption digital twin model, at a preset time step. Trigger dynamic model solving; The theoretical optimal energy consumption calculation unit is used to calculate the theoretical power consumption of each device using a variable step-size iterative algorithm, with a step size range of [missing information]. to The theoretical total energy consumption of the system is obtained by summing the theoretical power consumption of each device, and then the continuous energy consumption is calculated. The theoretical energy consumption at each time step is averaged to output the theoretical optimal energy consumption baseline value under the current operating condition. ; The measured total energy consumption acquisition and verification unit is used to collect the actual power consumption of each device in real time and sum them up to obtain the measured total energy consumption of the system. Time windows are aligned based on the timestamps of key parameter acquisition, and time synchronization errors exceeding a preset threshold are eliminated. The data was analyzed, and its validity was verified by deviation detection compared to historical data from the same period. The historical data from the same period refers to the historical measured total energy consumption dataset under the same operating conditions. The deviation detection threshold was [missing information]. ; Energy consumption deviation rate calculation and storage unit, used to calculate and store energy consumption deviation rate according to formula Calculate the real-time energy consumption deviation rate, if If it is 0, then set =0; store the energy consumption deviation rate. and corresponding timestamps to form a time series .

10. The air conditioning energy consumption adaptive management system based on dynamic feature selection according to claim 6, characterized in that: The graded early warning triggering module includes: The energy consumption deviation rate smoothing unit is used to smooth the energy consumption deviation rate time series. Exponentially weighted moving average smoothing is performed using a window size of m, with time-reverse indices q = 0, 1, ..., m−1, where q = 0 corresponds to the latest time t. The smoothing energy consumption deviation rate at time t is calculated using the following formula. : ,in This represents the energy consumption deviation rate at time tq; thus, the smoothed energy consumption deviation rate time series is obtained. ; Threshold benchmark calculation unit, used for historical duration All energy consumption deviation rate data within ,use Outlier removal is performed using principles or box plot methods to obtain a dataset of baseline deviation rates. The kernel density estimation method is used based on Calculate threshold benchmark ; The operating condition adaptation factor calculation unit is used to calculate the system load rate, which characterizes the current operating condition. and outdoor combined temperature Normalization was performed separately to obtain the normalized load factor. and normalized outdoor composite temperature The operating condition adaptation factor was determined using a multiple linear regression model. The calculation formula is: ,in , , For regression coefficients; with a fixed period The interval is based on the time before the current moment. Historical data within a time window are used to determine the regression coefficients of the multiple linear regression model. , , ; and then according to the formula The operating condition adaptation factor for the current cycle is calculated. ; An anomaly detection unit, used for anomaly detection, includes: The tiered early warning execution unit is used to implement early warning based on threshold benchmarks. Adaptability Factor to Operating Conditions The dynamic threshold T is calculated using the following formula: ; Calculate continuous Deviation slope at each time point ,when And |s| is greater than the preset slope threshold. When the mutation is abnormal, it is determined to be abnormal. Progressive anomaly detection: using long short-term memory neural networks to predict the future. Energy consumption deviation rate over time, when continuous The actual deviation rate showed a monotonically increasing trend over the duration and the predicted value exceeded [the expected value]. When it is determined to be a gradual anomaly, among which The preset prediction value multiple threshold; Periodic anomaly detection: Periodic features of the deviation rate time series are extracted using Fourier transform; when the mean deviation rate of a certain period is continuous... This time it exceeded the historical average of this period and When the product of is used, it is determined to be a periodic anomaly, where The threshold is set as a multiple of the preset periodic average. Step S4-5: Trigger tiered alerts based on anomaly detection results: Level 1 warning: Triggered when a sudden abnormality or a single-cycle abnormality is detected, triggering an audible and visual alarm through the local monitoring system; Warning cancellation conditions: continuous Duration, energy consumption deviation rate All values ​​are within the dynamic threshold T, and the absolute value of the slope |s| in the mutation anomaly detection is consistently lower than the preset slope threshold. The predicted value in progressive anomaly detection is consistently lower than The periodic mean deviation in periodic anomaly detection is consistently lower than the historical mean. The product of.

Images