A smart building environment adaptive regulation method based on digital twinning and AI

Abstract

The application provides a kind of intelligent building environment adaptive regulation and control method based on digital twinning and AI, it is related to artificial intelligence technical field, including the following steps: based on time alignment after multi-source system data, extract dynamic characteristics to construct dynamic feature set;Multi-source system data is input to digital twinning, with the current building equipment state as boundary condition, iteratively simulates the key environmental parameters of multiple time steps in the future, and outputs multi-step prediction sequence;Again input model predictive controller, with the minimum overall energy consumption in the prediction period as the optimization goal, with the key environmental parameters not exceeding the comfort interval as the constraint condition, the equipment control instruction sequence in the first preset time in the future is solved;The first equipment control instruction in the equipment control instruction sequence is issued to the on-site actuator to realize the adaptive regulation and control of building environment, solve the disadvantage that the environment parameter is over-standard and then responds passively, significantly improve the rapid response capability of the system to sudden environmental changes.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of artificial intelligence, in particular to a smart building environment adaptive regulation method based on digital twinning and AI. BACKGROUND

[0002] Although the existing smart building operation analysis technology such as the scheme disclosed in CN120123757B has made progress in multi-source data fusion and causal diagnosis, its core limitation lies in the lack of real-time prediction and forward-looking regulation capability for building environment changes. The specific performance is as follows:

[0003] Through hidden variable modeling and transfer entropy analysis, the causal chain of people flow density CO2 concentration HVAC energy consumption can be effectively identified, and the surge of people flow density is located as the root cause, but this analysis is essentially a post-diagnosis mechanism. When the system identifies the root cause, the environmental parameters (such as CO2 concentration) have often exceeded the standard, and the HVAC system can only respond passively, resulting in regulation lag and sharp increase in energy consumption.

[0004] More importantly, this method cannot cope with sudden environmental changes caused by building thermal inertia, people flow impact conduction and other dynamic characteristics. For example:

[0005] During the morning rush hour, the heat load and CO2 emissions caused by the sudden increase in people flow will not be immediately reflected, but will be delayed and conducted with the building thermal inertia;

[0006] The traditional control system will start regulation only when the temperature or CO2 concentration actually exceeds the standard, at which time the refrigeration load has accumulated to the peak value, causing a sharp increase in energy consumption and a decrease in comfort.

[0007] Therefore, it is difficult to meet the rapid response and energy efficiency optimization needs of smart buildings to sudden environmental changes. SUMMARY

[0008] In view of the above defects or deficiencies in the prior art, the present application aims to provide a smart building environment adaptive regulation method based on digital twinning and AI, comprising the following steps:

[0009] Based on the time-aligned multi-source system data, dynamic features are extracted to construct a dynamic feature set; the multi-source system data includes environmental data, device operation data and people flow data; the dynamic features include thermal inertia delay features and people flow impact conduction features;

[0010] The multi-source system data is input into the digital twin, and the current building equipment state is used as the boundary condition to iteratively simulate the key environmental parameters at multiple time steps in the future, and output a multi-step prediction sequence; the key environmental parameters include indoor temperature and CO2 concentration;

[0011] The multi-step prediction sequence is input into the model prediction controller. With the goal of minimizing the overall energy consumption during the prediction period and the constraint that the key environmental parameters do not exceed the comfort range, the sequence of equipment control instructions for the first preset time period in the future is obtained.

[0012] The first equipment control command in the equipment control command sequence is sent to the field actuator to achieve adaptive control of the building environment;

[0013] The digital twin is a physical information neural network constructed by fusing the dynamic feature set with the building information physical model; the building information physical model is obtained by discretizing the thermodynamic differential equation, and the loss function of the physical information neural network includes data fitting terms and physical equation constraint terms.

[0014] According to the technical solution provided in this application, thermal inertial delay features are extracted based on time-aligned multi-source system data, including the following steps:

[0015] Based on the ambient temperature sequence and HVAC power sequence within the target historical time period, the mutual information value at different time offsets is calculated, and the time offset that maximizes the mutual information value is taken as the thermal inertia delay time of the building.

[0016] For each acquisition moment, the thermal inertia delay characteristic is obtained by combining the real-time ambient temperature acquired at that acquisition moment with the HVAC energy consumption that is delayed by the thermal inertia delay time at that acquisition moment.

[0017] According to the technical solution provided in this application, the mutual information value is obtained by the following formula:

[0018] ;

[0019] in, Represents mutual information value; Indicates the time offset; T represents the ambient temperature sequence; t represents the real-time ambient temperature; p represents the HVAC power sequence; The temperature value t represents the temperature at the i-th data acquisition time. i Delay power value at the j-th acquisition time The joint probability of simultaneous occurrence; Indicates the edge probability of temperature; The value represents the edge probability of delayed power; N represents the total number of data points in the ambient temperature sequence T; and M represents the total number of data points in the HVAC power sequence p.

[0020] According to the technical solution provided in this application, before calculating the mutual information value at different time offsets based on the ambient temperature sequence and HVAC power sequence within the target historical time period, the following steps are included:

[0021] Obtain the current day's pattern feature vector to determine the current building operation mode category; the pattern feature vector includes the daily average outdoor temperature, daily HVAC energy consumption peak, and a histogram of pedestrian density distribution;

[0022] Based on the current building operation mode category, the corresponding target historical time period is obtained.

[0023] According to the technical solution provided in this application, obtaining the current day's pattern feature vector and determining the current building operation mode category includes the following steps:

[0024] Obtain the pattern feature vector for the current day, retrieve and traverse the historical pattern database to obtain the building operation mode category corresponding to the pattern feature vector for the current day, and take it as the current building operation mode category;

[0025] The historical pattern database includes multiple historical time periods and the building operation mode category corresponding to each historical time period.

[0026] According to the technical solution provided in this application, the method further includes the following steps:

[0027] While the model predictive controller solves the equipment control command sequence, based on the digital twin, the expected operating state of key equipment when executing the equipment control command sequence is simulated synchronously. The expected operating state includes the expected load rate of the compressor and the expected speed of the fan.

[0028] After obtaining the sequence of device control commands within the first preset time period in the future, the method further includes the following steps:

[0029] Determine whether any of the critical devices continuously exceeds its safe operating boundary during the first preset time period;

[0030] The step of issuing the first equipment control command in the equipment control command sequence to the field actuator includes the following steps:

[0031] If not, the first equipment control command in the equipment control command sequence will be sent to the field actuator.

[0032] According to the technical solution provided in this application, after determining whether any of the key devices continuously exceeds its safe operating boundary within the first preset time period, the method further includes the following steps:

[0033] If so, the affected environmental control loop is identified, and based on the digital twin, the control of the environmental control loop is dynamically reconstructed to the reconstructed system architecture in the virtual environment. The reconstructed system architecture includes one or more healthy backup devices or redundant control loops.

[0034] Using the reconstructed system architecture as the updated boundary conditions, the model predictive controller is rerun to solve for the sequence of alternative device control instructions within the second preset time period in the future.

[0035] The first equipment control command in the sequence of alternative equipment control commands is sent to the field actuator, and an equipment early warning and system reconfiguration report is generated simultaneously.

[0036] According to the technical solution provided in this application, the step of dynamically reconstructing the control rights of the environmental control loop to the reconstructed system architecture in a virtual environment based on the digital twin includes the following steps:

[0037] Based on the system's device topology, at least two alternative reconfiguration strategies are generated to allocate the load of the environmental control loop to healthy standby equipment or redundant control loops.

[0038] For each of the alternative reconfiguration strategies, a forward simulation is performed in the digital twin to simulate the execution of the alternative device control command sequence and to calculate the simulation performance index of each of the alternative reconfiguration strategies. The simulation performance index includes the impact intensity of the switching process on comfort, the additional energy consumption cost, and the device lifespan reduction factor.

[0039] Based on the aforementioned performance metrics, each of the candidate reconstruction strategies is comprehensively scored.

[0040] The restructured system architecture is obtained based on the highest-rated alternative refactoring strategy.

[0041] According to the technical solution provided in this application, obtaining the reconstructed system architecture based on the highest-scoring alternative reconstructing strategy includes the following steps:

[0042] Based on the digital twin, stress tests are conducted on the candidate reconstruction strategy with the highest score under various preset disturbance scenarios to obtain the performance retention of the candidate reconstruction strategy under each disturbance scenario. The preset disturbance scenarios include sudden fluctuations in pedestrian load and / or abrupt changes in outdoor meteorological parameters. The performance retention is the percentage of time that the comfort constraint can still be met under the disturbance.

[0043] If the performance retention of the alternative refactoring strategies meets the standard in each of the disturbance scenarios, then the architecture corresponding to the alternative refactoring strategy with the highest score will be used as the refactored system architecture.

[0044] According to the technical solution provided in this application, after obtaining the performance retention of the alternative reconstruction strategy under each of the disturbance scenarios, the method further includes the following steps:

[0045] If the performance retention of the alternative reconstruction strategy fails to meet the standard under any of the preset disturbance scenarios, then based on the characteristics of the preset disturbance scenarios that fail to meet the standard, the key devices that cause performance degradation and their operating periods are deduced in reverse in the digital twin.

[0046] Based on the equipment control command sequence obtained from the previous control cycle, a dynamic compensation command sequence for the key equipment during the operating period is generated. The dynamic compensation command sequence includes a progressive adjustment curve for the equipment output power and switching timing optimization.

[0047] The dynamic compensation instruction sequence is fused with the alternative reconstruction strategies to form an enhanced reconstruction strategy;

[0048] The enhanced reconstruction strategy is verified in the digital twin, and the architecture corresponding to the verification is determined as the reconstructed system architecture.

[0049] Compared with the prior art, the beneficial effects of this application are as follows:

[0050] First, by constructing a digital twin that integrates physical information neural networks, the system can perform high-precision iterative simulations of key parameters such as temperature and CO2 concentration at multiple future time steps based on the current equipment status and multi-source dynamic characteristics. This enables the system to move beyond post-hoc causal analysis and acquire multi-step predictive capabilities for environmental changes, providing a basis for proactive regulation.

[0051] Second, by extracting thermal inertia delay features and pedestrian impact conduction features, the system can accurately characterize the dynamic response process of the building environment to sudden changes. Combined with the predictive output of the digital twin, the model predictive controller can adjust the equipment operating status in advance before the heat load / CO2 concentration accumulates to the peak, fundamentally solving the control delay problem caused by building thermal inertia and pedestrian impact conduction.

[0052] Third, the multi-step prediction sequence output by the digital twin is input into the model prediction controller. The optimization objective is to minimize the overall energy consumption during the prediction period, while ensuring that the environmental parameters do not exceed the comfort range. The optimal equipment control command sequence is obtained and executed in real time, forming a complete intelligent control closed loop, realizing adaptive and refined control of the building environment.

[0053] Fourth, by combining physical information neural networks with model predictive control, the system maintains both the reliability based on physical mechanisms and the adaptability driven by data. It can dynamically optimize equipment operation strategies while meeting the comfort requirements of the indoor environment, avoiding energy consumption peaks. Compared with pure analysis solutions, it has more direct energy-saving benefits and engineering application value.

[0054] Fifth, for scenarios with sudden changes in traffic flow, such as the morning rush hour, the system uses a digital twin to predict the trend of environmental parameter changes in advance, and generates optimal control instructions through model prediction controllers. This effectively avoids the drawbacks of traditional methods that only respond passively after environmental parameters exceed the limits, and significantly improves the system's ability to respond quickly to sudden environmental changes. Attached Figure Description

[0055] Figure 1 A flowchart illustrating the steps of the adaptive control method for smart building environments based on digital twins and AI provided in this application. Detailed Implementation

[0056] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0057] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0058] Example 1

[0059] As mentioned in the background section, to address the problems in existing technologies, this application proposes a smart building environment adaptive control method based on digital twins and AI, such as... Figure 1 As shown, it includes the following steps:

[0060] S1. Based on time-aligned multi-source system data, extract dynamic features to construct a dynamic feature set; the multi-source system data includes environmental data, equipment operation data, and pedestrian flow data; the dynamic features include thermal inertia delay features and pedestrian flow impact conduction features;

[0061] Specifically, environmental data (temperature, humidity, CO2 concentration, light intensity), equipment operation data (such as instantaneous power of HVAC systems, lighting power, equipment current, etc., used to calculate energy consumption and related characteristics), and pedestrian flow data (density, speed, dwell time) are first collected through a sensor network deployed within the building. Since different sensors may have different sampling frequencies and clocks, time alignment processing is required first. A weighted least squares-based time axis mapping method is used to calculate the average time difference for each type of sensor data. The optimal alignment time point is found through nonlinear optimization, forming an alignment matrix with a unified time reference. Thermal inertia delay characteristics: reflecting the temperature response lag caused by the heat storage capacity of the building envelope, furniture, etc. The building-specific thermal inertia delay time is determined by analyzing the mutual information peaks of historical temperature sequences and HVAC power sequences. Pedestrian flow impact conduction characteristics: quantifying the conduction effect of sudden changes in pedestrian flow on environmental parameters. The conduction intensity and delay time are extracted by calculating the cross-correlation between the rate of change of pedestrian density and the rate of increase of CO2 concentration. The built environment is a complex system coupled across multiple time scales. Thermal inertia causes temperature regulation delays ranging from minutes to hours, while the impact of pedestrian flow exhibits spatial conduction characteristics. By extracting these dynamic features, the transient response behavior of the system can be described more accurately.

[0062] S2. Input the multi-source system data into the digital twin, and use the current building equipment status as boundary conditions to iteratively simulate key environmental parameters for multiple future time steps, and output a multi-step prediction sequence; the key environmental parameters include indoor temperature and CO2 concentration;

[0063] Specifically, the digital twin employs a physical information neural network architecture, the core of which is the deep integration of building information physical models (BIMs) with data-driven methods. The BIM model is derived from the discretization of thermodynamic differential equations, describing physical processes such as heat transfer, airflow, and pollutant diffusion within the building envelope. The loss function of the physical information neural network includes a data fitting term (minimizing the difference between predicted and measured values) and a physical equation constraint term (ensuring that the prediction results conform to physical laws). By introducing physical equations as regularization terms into the loss function, the physical information neural network ensures that physically reasonable predictions are produced even in regions of sparse data. This fusion approach maintains the flexibility of data-driven models while inheriting the generalization ability of physical models.

[0064] Specific operation: Using the current status of building equipment (such as air conditioning set temperature, fan speed, and fresh air valve opening) as boundary conditions, input multi-source system data and dynamic feature sets into the digital twin, and generate a prediction sequence of key environmental parameters for multiple time steps in the future (such as the next 2 hours, with a step size of 15 minutes) through iterative simulation, including indoor temperature and CO2 concentration.

[0065] S3. Input the multi-step prediction sequence into the model prediction controller, take the lowest overall energy consumption during the prediction period as the optimization objective, and take the key environmental parameters not exceeding the comfort range as the constraint condition, and solve to obtain the equipment control instruction sequence within the first preset time period in the future.

[0066] Specifically, the model predictive controller aims to minimize overall energy consumption during the prediction period, while constraining key environmental parameters to remain within the comfort range (e.g., temperature 22-26℃, CO2 concentration <1000ppm). The optimization problem can be formulated as a constrained quadratic programming problem, solved using the interior-point method or the effective set method. The process is as follows: At each decision point, based on the current state and future predictions, the controller solves for the optimal sequence of equipment control commands for the first preset time period (e.g., the next hour), including adjustments to air conditioning set temperature, fresh air volume control commands, and lighting dimming levels. Model predictive control employs a rolling optimization strategy, resolving the finite-time domain optimization problem in each control cycle and implementing only the first control command. This strategy can promptly compensate for model errors and external disturbances, achieving closed-loop optimal control.

[0067] S4. Send the first equipment control command in the equipment control command sequence to the field actuator to achieve adaptive control of the building environment;

[0068] The digital twin is a physical information neural network constructed by fusing the dynamic feature set with the building information physical model; the building information physical model is obtained by discretizing the thermodynamic differential equation, and the loss function of the physical information neural network includes data fitting terms and physical equation constraint terms.

[0069] Specifically, the obtained equipment control command sequence is sent to the field actuators via building automation system protocols (such as BACnet and Modbus). Only the first command in the sequence is executed, and then prediction and optimization are performed again in the next control cycle.

[0070] In a preferred embodiment, thermal inertial delay features are extracted based on time-aligned multi-source system data, including the following steps:

[0071] Based on the ambient temperature sequence and HVAC power sequence within the target historical time period, the mutual information value at different time offsets is calculated, and the time offset that maximizes the mutual information value is taken as the thermal inertia delay time of the building.

[0072] For each acquisition moment, the thermal inertia delay characteristic is obtained by combining the real-time ambient temperature acquired at that acquisition moment with the HVAC energy consumption that is delayed by the thermal inertia delay time at that acquisition moment.

[0073] Specifically, the data preparation phase involves collecting ambient temperature and HVAC power sequences for the target historical time period (e.g., the past 30 days). Data must undergo time alignment and outlier handling to ensure data quality. The mutual information calculation phase involves calculating the mutual information value between the ambient temperature sequence and the delayed HVAC power sequence at different time offsets (e.g., 0-120 minutes, 5-minute step). Mutual information measures the statistical dependency between two variables. The time offset τ1 that maximizes the mutual information value is identified; this value represents the building's thermal inertia delay time. This process is achieved by scanning the mutual information at different τ values ​​to identify peak locations. For each acquisition moment, the real-time ambient temperature is extracted, and the HVAC energy consumption value delayed by τ1 is found. These two values ​​are combined to form the thermal inertia delay feature.

[0074] Furthermore, the mutual information value is obtained by the following formula:

[0075] ;

[0076] in, Represents mutual information value; Indicates the time offset; T represents the ambient temperature sequence; t represents the real-time ambient temperature; p represents the HVAC power sequence; The temperature value t represents the temperature at the i-th data acquisition time. i Delay power value at the j-th acquisition time The joint probability of simultaneous occurrence; Indicates the edge probability of temperature; The value represents the edge probability of delayed power; N represents the total number of data points in the ambient temperature sequence T; and M represents the total number of data points in the HVAC power sequence p.

[0077] Specifically, data discretization: The continuous ambient temperature sequence T and HVAC power sequence P are discretized into a finite number of intervals. Temperature can be discretized at 0.5℃ intervals, and energy consumption can be discretized as a percentage of rated power. The degree of discretization needs to be balanced between computational accuracy and computational complexity. Probability distribution estimation: Marginal probability distribution p(t): the frequency of statistical temperature values ​​falling in each discrete interval; Marginal probability distribution p(pτ): the frequency of statistical energy consumption values ​​falling in each discrete interval after a delay τ; Joint probability distribution p(t,pτ): the frequency of statistical temperature t and energy consumption pτ after a delay τ simultaneously appearing in the corresponding interval; Mutual information calculation: For each time offset τ, the mutual information value is calculated according to the above probability distributions. Care must be taken to handle the zero-probability case during calculation; Laplace smoothing is usually used to avoid numerical problems. Determination of maximum mutual information: Among all candidate τ values, the τ value that maximizes I(τ) is selected as the thermal inertia delay time. It should be further explained that cumulative concepts such as energy consumption data and total energy consumption are obtained by integration or accumulation based on equipment operating data (such as power data); while when extracting dynamic features, instantaneous or sequential data (such as power sequence) are used.

[0078] For example, this example demonstrates the complete operation of a smart office building environment adaptive control system. The system is deployed in an office building with a floor area of ​​50,000 square meters, focusing on solving environmental quality control and energy consumption optimization issues during weekday morning peak hours. System Architecture and Data Foundation: The system constructs a complete sensor network and computing platform. The sensor network includes 86 temperature and humidity sensors, 45 CO2 concentration sensors, 12 HVAC system energy consumption monitoring meters, and 28 computer vision-based intelligent people flow statistics cameras. The computing platform consists of edge computing nodes and a cloud-based digital twin platform. The edge nodes are responsible for real-time data preprocessing and control command execution, while the cloud platform runs physical information neural networks and model predictive control algorithms. Preparation Phase Before Morning Peak: During the preparation phase from 7:00 to 7:30 AM, the system first performs time alignment processing on multi-source data. Data from each sensor is unified to a standard time base through a time axis mapping function, and abnormal data points are automatically identified and corrected, such as zero-point drift issues from a temperature sensor. The system then performs dynamic feature extraction. The optimal delay time for the building was determined to be 45 minutes by calculating the thermal inertia delay characteristics through mutual information analysis. The system combines the current temperature of 23.5℃ with the HVAC power data from 45 minutes ago to form the thermal inertia delay characteristics. Simultaneously, based on historical data analysis of morning peak pedestrian flow patterns, the impact transmission characteristics of pedestrian flow were extracted, predicting that the peak pedestrian flow in the core area will occur between 8:15 and 8:45. The digital twin performs multi-step predictions based on the current building status and external conditions. Inputting the current indoor temperature of 23.5℃, CO2 concentration of 450ppm, HVAC energy-saving operation mode, and the weather forecast of today's high temperature of 32℃, the system outputs a sequence of predicted future environmental parameters. The prediction shows that from 9:00, the CO2 concentration will exceed the comfort limit of 1000ppm, reaching 1250ppm by 9:15. Model predictive control optimization: The model predictive controller aims to minimize overall energy consumption over the next two hours, while constraining the temperature to not exceed 26℃ and the CO2 concentration to not exceed 1000ppm. By solving a constrained optimization problem, an optimal sequence of equipment control commands was generated: at 8:00, the air conditioning temperature was set to 24.5℃, the fresh air valve opening to 65%, and the fan speed to 75%; at 8:30, this was adjusted to 24.0℃, 80% opening, and 85% speed; at 9:00, it was further adjusted to 23.5℃, 95% opening, and 95% speed. During peak hours, the system demonstrated excellent proactive control capabilities. At 7:45, the system proactively adjusted the air conditioning temperature in the core area to 24.5℃; at 8:15, when the pedestrian density reached 0.3 people / square meter, the fresh air volume was increased to 80%; at 8:40, when the CO2 concentration rose to 850ppm, due to the proactive control measures, the concentration never exceeded the 1000ppm threshold. The system also demonstrated flexible adaptability.When the actual flow of people reached its peak 15 minutes later than predicted, the digital twin updated the prediction model in real time. The model's predictive controller then re-optimized subsequent instructions, appropriately reducing the fan speed after 9:00 AM to optimize energy consumption while ensuring comfort. Throughout the operation, the system continuously monitored the equipment's operating status. When the digital twin detected that the compressor load rate consistently exceeded the safety threshold of 85%, it automatically generated an alternative strategy, transferring part of the load to a standby chiller unit to ensure all equipment operated within safe limits. The system's ability to handle special scenarios was also enhanced, performing exceptionally well in a midday meeting scenario. At 12:00 PM, the system detected a large meeting scheduled for the afternoon through the meeting booking system and predicted, based on historical data, that the flow density during the meeting would surge from 0.1 to 0.8 people / square meter. The system began gradual pre-cooling at 1:30 PM to avoid discomfort caused by a sudden temperature drop, and the CO2 concentration remained stably below 950 ppm during the meeting. When dealing with extreme weather conditions, when the outdoor temperature suddenly rises to 36°C, the digital twin combines weather forecast information to increase cooling capacity one hour in advance and optimize the start-up and shutdown sequence of equipment through thermal inertia characteristic calculation, effectively avoiding the occurrence of peak power demand.

[0079] In a preferred embodiment, before calculating the mutual information values ​​at different time offsets based on the ambient temperature sequence and HVAC power sequence within the target historical time period, the following steps are included:

[0080] Obtain the current day's pattern feature vector to determine the current building operation mode category; the pattern feature vector includes the daily average outdoor temperature, daily HVAC energy consumption peak, and a histogram of pedestrian density distribution;

[0081] Based on the current building operation mode category, the corresponding target historical time period is obtained.

[0082] Specifically, the model feature vector is a multi-dimensional feature set that quantifies the daily operating status of a building. Its dimensions include: Average daily outdoor temperature: the statistical average of the temperature throughout the day, reflecting the intensity of external thermal disturbance; Daily HVAC energy consumption peak: the maximum instantaneous power of the HVAC system on that day, indicating the system's load limit; Pedestrian density distribution histogram: the frequency distribution of the average number of pedestrians in each equal-length interval after dividing the 24 hours, describing the temporal patterns of internal heat sources and CO2 emissions; Building operation mode category: typical operation scenario labels generated based on historical data clustering, used to distinguish the building's behavior patterns under different external environments and usage intensities, such as "summer weekdays" and "winter weekends"; Target historical time period: a historical data range dynamically selected according to the current model category, aiming to ensure consistency between the data used for thermal inertia calculations and current conditions.

[0083] The specific implementation method is described below: Obtaining the pattern feature vector for the current day: The system automatically collects and calculates the pattern feature vector from multiple data sources daily. Specifically:

[0084] Average daily outdoor temperature: Temperature data is collected throughout the day using a network of temperature sensors outside the building (such as PT100 platinum resistance thermometers or digital temperature sensors). The sampling frequency can be once every 5 minutes, and the arithmetic mean of all sampling points for the day is calculated. This value reflects the overall background heat load.

[0085] Daily HVAC Energy Peak: This is achieved by reading power time-series data (in kW) from smart meters or energy management systems to identify the maximum daily energy consumption and the time of its occurrence. This peak value characterizes the building's energy demand under extreme load conditions.

[0086] Pedestrian density distribution histogram: Pedestrian flow is counted using infrared counters or video analytics systems installed at the entrance, generating a 24-dimensional histogram vector with a 1-hour time slot. Each element represents the average number of people during that time period. This histogram reveals the temporal patterns of building usage intensity.

[0087] Data Fusion: After time alignment (e.g., unified to UTC timestamps), the above data is encapsulated into a pattern feature vector V=[Tavg,Pmax,H1,H2,...,H24], where Tavg is the average daily outdoor temperature, Pmax is the peak HVAC energy consumption, and Hi is the pedestrian density in the i-th hour. Determining the current building operation mode category: The system pre-stores a historical pattern database, which is generated by labeling historical data through cluster analysis (such as the K-means algorithm) and contains three typical patterns:

[0088] High-efficiency operation mode: corresponding to weekdays in mild seasons, characterized by moderate outdoor temperatures (such as 15-25°C), balanced peak energy consumption, and bimodal pedestrian flow distribution (morning and evening rush hours).

[0089] Extreme load mode: corresponding to extremely cold / extremely hot days, characterized by low / high temperatures (such as <5°C or >30°C), high energy consumption peaks and concentrated crowds (such as midday peaks).

[0090] Low load mode: corresponding to holidays or nighttime, characterized by comfortable temperature, low energy consumption peak and flat pedestrian flow distribution.

[0091] Matching logic: Calculate the Euclidean distance between the current daily feature vector and the center vectors of each category in the database, and select the category with the smallest distance as the current mode. For example, if the current vector has the smallest distance to the center vector of the efficient running mode, then it is classified as that mode.

[0092] Retrieve the target historical time period:

[0093] Implementation: Each building operation mode category is associated with a preset time window strategy:

[0094] High-efficiency operation mode: Select data from all working days with the same pattern within the past 30 days to ensure that the samples have similar thermal dynamic characteristics.

[0095] Extreme load pattern: Select continuous data under the same weather conditions within the past 7 days to avoid seasonal bias.

[0096] Low load mode: Select all weekend or holiday data within the past 14 days.

[0097] Dynamic query: Based on the determined mode category, the system extracts the ambient temperature sequence and HVAC power sequence for the corresponding time period from the time series database, and uses them as input for calculating the thermal inertia delay characteristics.

[0098] In a preferred embodiment, obtaining the current day's pattern feature vector and determining the current building operation mode category includes the following steps:

[0099] Obtain the pattern feature vector for the current day, retrieve and traverse the historical pattern database to obtain the building operation mode category corresponding to the pattern feature vector for the current day, and take it as the current building operation mode category;

[0100] The historical pattern database includes multiple historical time periods and the building operation mode category corresponding to each historical time period.

[0101] Specifically, the pattern feature vector for the current day is obtained: the real-time data acquisition system automatically starts feature calculation at 00:00 every day.

[0102] The average daily outdoor temperature is obtained from the meteorological station's API interface. The sampled values ​​are averaged after being denoised by Kalman filtering.

[0103] HVAC energy consumption peaks are extracted from meter readings using the Modbus / TCP protocol, with a scanning frequency of 1 scan per minute.

[0104] The crowd density histogram is generated from the access control system logs, aggregating the number of people entering and exiting by hour.

[0105] Vector standardization: To prevent differences in dimensions, each feature value is normalized using Min-Max, for example, temperature is mapped to the [0,1] interval.

[0106] Retrieve and iterate through the historical pattern database:

[0107] Database structure: The historical schema database is a relational table (such as in MySQL), containing the following fields:

[0108] date (date), mode_class (mode class), temp_avg (average daily temperature), power_max (peak energy consumption), histogram (JSON histogram).

[0109] The data source is historical data from the past 3 years, and the pattern labels are updated monthly using a clustering algorithm.

[0110] Traversal Matching: The system executes an SQL query to retrieve all historical records, calculating the similarity between the current feature vector and historical vectors row by row. The similarity metric uses weighted cosine similarity: S = w1·cos(Tavg,Tavg′) + w2·cos(Pmax,Pmax′) + w3·cos(H,H′); where weights w1 = 0.4, w2 = 0.3, and w3 = 0.3 are assigned based on feature importance, and cos is the cosine similarity function. The historical record with the highest similarity is selected, and its mode_class field value is used as the current category. The matched category is output to the thermal inertia calculation module, triggering a query for the corresponding target historical time period.

[0111] In a preferred embodiment, the method further includes the following steps:

[0112] While the model predictive controller solves the equipment control command sequence, based on the digital twin, the expected operating state of key equipment when executing the equipment control command sequence is simulated synchronously. The expected operating state includes the expected load rate of the compressor and the expected speed of the fan.

[0113] After obtaining the sequence of device control commands within the first preset time period in the future, the method further includes the following steps:

[0114] Determine whether any of the critical devices continuously exceeds its safe operating boundary during the first preset time period;

[0115] The step of issuing the first equipment control command in the equipment control command sequence to the field actuator includes the following steps:

[0116] If not, the first equipment control command in the equipment control command sequence will be sent to the field actuator.

[0117] Specifically, the expected operating state of key equipment is simulated synchronously: In the parallel thread of the Model Predictive Controller (MPC) solving for the control command sequence, the digital twin receives the sequence as input and performs equipment-level simulation: Compressor expected load rate: Based on the refrigerant flow model and the start / stop signals in the commands, the load rate Lc = Qactual / Qrated × 100% is calculated every 5-minute time step, where Qactual is the actual cooling capacity and Qrated is the rated cooling capacity. Fan expected speed: Based on the fan law model N = (P represents input power) Derive the speed curve and cross-validate it with the inverter setpoint in the command. Generate a device status sequence perfectly aligned with the time axis of the control command sequence, including the load rate time series {Lc(t)} and the speed time series {N(t)}. Determine if the device status continuously exceeds the safe operating boundary: Boundary definition: The safe operating boundary is preset in the device knowledge base: The safe range for compressor load rate is [10%, 90%], and three consecutive time steps (15 minutes) > 90% is considered "continuously exceeding". The safe range for fan speed is [0, 1500 RPM], and an alarm is triggered if the instantaneous value exceeds 1500 RPM.

[0118] Judgment Logic: The system scans simulation data for the entire first preset duration (e.g., the next 2 hours):

[0119] For compressors: Check if there is a period where the load rate is >90% for a continuous time window.

[0120] For the fan: check if the speed at any sampling point is >1500 RPM.

[0121] If any condition is met, it is determined that "there is a continuous exceedance".

[0122] Conditional issuance of control orders:

[0123] Safe Path: If all device statuses are within limits, the system will send the first instruction in the device control instruction sequence obtained from the MPC solution (i.e., the instruction that should be executed immediately at the current moment) to the field PLC actuator via the OPC UA protocol. Exception Handling: If limits are exceeded, the following refactoring process will be triggered.

[0124] In a preferred embodiment, after determining whether any of the critical devices continuously exceeds its safe operating boundary within the first preset time period, the method further includes the following steps:

[0125] If so, the affected environmental control loop is identified, and based on the digital twin, the control of the environmental control loop is dynamically reconstructed to the reconstructed system architecture in the virtual environment. The reconstructed system architecture includes one or more healthy backup devices or redundant control loops.

[0126] Using the reconstructed system architecture as the updated boundary conditions, the model predictive controller is rerun to solve for the sequence of alternative device control instructions within the second preset time period in the future.

[0127] The first equipment control command in the sequence of alternative equipment control commands is sent to the field actuator, and an equipment early warning and system reconfiguration report is generated simultaneously.

[0128] Specifically, an environmental control loop refers to a complete control chain that adjusts specific environmental parameters, including sensors, controllers, actuators, and the controlled object. For example, a temperature control loop includes a temperature sensor, a PID controller, a VAV terminal unit, and the corresponding spatial area. Dynamic reconfiguration is the process of adjusting the control architecture in real time based on equipment status during system operation, including the transfer of control authority and the reallocation of equipment roles, ensuring that the system can maintain basic functionality even when some components malfunction. The reconfigured system architecture is the optimized and reorganized organizational form of the control system, clearly defining the control responsibilities and collaborative relationships of each device, typically represented by a device connection matrix and a control logic tree. Alternative device control command sequences are time sequences of device control commands re-optimized based on the reconfigured architecture, aiming to approximate the original control objective as closely as possible while ensuring safety.

[0129] Specifically, identify affected environmental control loops: when the digital twin detects that the expected operating state of critical equipment continues to exceed the safety boundary, the system immediately initiates impact analysis:

[0130] Establish a device-loop mapping table: for example, compressor C1 serves "East Zone Temperature Control Loop R1", and fan F2 serves "North Zone Fresh Air Control Loop R2".

[0131] Topology propagation analysis: Based on the pipe network connections in the building BIM model, identify affected upstream and downstream equipment. For example, compressor overload can affect related components such as condensers and expansion valves.

[0132] Circuit importance assessment: Impact is tiered based on the area function served by the circuit (prioritizing critical areas such as operating rooms and data centers).

[0133] Output: Generates a list of affected circuits, labeling the fault level (minor / severe) and affected area for each circuit.

[0134] Dynamic reconfiguration of control based on digital twins:

[0135] Performing system architecture refactoring in the virtual environment of a digital twin:

[0136] Backup equipment activation: If the main compressor is detected to be overloaded, the backup redundant compressor will be automatically activated and the refrigerant distribution will be recalculated;

[0137] Redundant control loop switching: When a temperature control loop fails, control is transferred to a redundant loop in an adjacent area, and coverage is achieved by adjusting the air valve;

[0138] Load redistribution: Based on real-time calculation of equipment capacity, the load of overloaded equipment is proportionally distributed to multiple healthy equipment;

[0139] Refactoring verification: Simulate the refactored architecture in a virtual environment for 1-2 control cycles to confirm system stability;

[0140] Rerun the model predictive controller: using the refactored system architecture as the new boundary conditions.

[0141] Update the device status constraints of MPC: such as the power range of standby equipment and the control range of redundant loops;

[0142] Adjust and optimize target weights: Increase the weight of balanced equipment usage while minimizing energy consumption;

[0143] Solving the sequence of control commands for alternative equipment: The time range is usually shortened to half of the original first preset duration (e.g., adjusted from 2 hours to 1 hour) to improve response speed;

[0144] Issuing instructions and generating reports: The first instruction in the candidate instruction sequence is issued to the field actuator via industrial Ethernet, and a structured report is generated simultaneously, including:

[0145] Equipment warning: Describe in detail the overloaded equipment, the degree of exceeding the limit, and the expected impact;

[0146] System Refactoring Report: Refactoring strategy, list of enabled devices, performance expectations;

[0147] Maintenance recommendations: suggested inspection schedule, spare parts preparation, etc.

[0148] In a preferred embodiment, the step of dynamically reconstructing the control of the environmental control loop to the reconstructed system architecture in a virtual environment based on the digital twin includes the following steps:

[0149] Based on the system's device topology, at least two alternative reconfiguration strategies are generated to allocate the load of the environmental control loop to healthy standby equipment or redundant control loops.

[0150] For each of the alternative reconfiguration strategies, a forward simulation is performed in the digital twin to simulate the execution of the alternative device control command sequence and to calculate the simulation performance index of each of the alternative reconfiguration strategies. The simulation performance index includes the impact intensity of the switching process on comfort, the additional energy consumption cost, and the device lifespan reduction factor.

[0151] Based on the aforementioned performance metrics, each of the candidate reconstruction strategies is comprehensively scored.

[0152] The restructured system architecture is obtained based on the highest-rated alternative refactoring strategy.

[0153] Specifically, equipment topology relationships describe the network structure of physical connections and functional associations between components in a building equipment system, including pipe connections, electrical connections, and control signal transmission. Forward simulation simulates the future evolution of the system from its current state in chronological order, predicting system behavior by solving differential equations or using surrogate models. Pre-simulation performance indicators are a quantitative standard system used to evaluate the effectiveness of reconfiguration strategies, covering multiple dimensions such as comfort, economy, and equipment reliability. Comfort impact intensity quantifies the degree of disturbance to indoor environmental comfort caused by system reconfiguration, typically expressed as the integral of environmental parameters deviating from set values. Equipment lifespan reduction factor characterizes the impact of specific operating conditions on the remaining service life of equipment, calculated based on stress-life curves and cumulative damage models.

[0154] The specific implementation methods are described below: Multiple reconfiguration schemes are automatically generated based on the system's device topology: Topology modeling: Using graph theory, a system topology graph is constructed with devices as nodes and connections as edges. Weights represent physical parameters such as pipe resistance and cable capacity; Strategy generation algorithm: Load transfer strategy: Identify adjacent healthy devices of overloaded devices and calculate the transferable load; Functional replacement strategy: Find different but functionally equivalent device combinations, such as replacing a single large-capacity device with multiple small-capacity devices; Area reorganization strategy: Re-divide control areas and adjust the service range of each device; Strategy diversity: Ensure at least two fundamentally different strategies are generated, such as one focusing on device redundancy and another focusing on control loop reorganization; Forward simulation in digital twin: Perform full-process simulation in a virtual environment for each alternative strategy: Simulation environment setting: Copy all state parameters of the current actual system as the initial conditions for simulation; Execution process rehearsal: Simulate the system's dynamic response at high time resolution (e.g., 10 seconds / step) according to the sequence of control commands for alternative devices; Multi-dimensional data acquisition: Record environmental parameters such as temperature, humidity, and pressure during the simulation process, as well as operating parameters such as power and speed of each device. Performance metrics calculation: Quantitatively evaluate each strategy based on simulation data: Comfort impact intensity: Calculate the root mean square deviation of key environmental parameters (temperature, CO2 concentration) before and after strategy implementation to quantify the comfort impact. Additional energy consumption cost: Compare the energy consumption difference between the reconfigured strategy and the original strategy, and calculate the economic cost in conjunction with real-time electricity prices. Equipment lifespan depreciation coefficient: Calculate the fatigue damage increment of key equipment based on Miner's linear cumulative damage theory. Comprehensive scoring and strategy selection: Weight allocation: Dynamically adjust the indicator weights according to the current building operation mode. For example, comfort has the highest weight in hospital mode, while energy consumption has a higher weight in commercial building mode; Score calculation: Use the TOPSIS multi-objective decision method to calculate the relative closeness of each strategy to the ideal solution; Select the strategy with the highest comprehensive score as the final reconfiguration scheme.

[0155] In a preferred embodiment, obtaining the restructured system architecture based on the candidate restructure strategy with the highest score includes the following steps:

[0156] Based on the digital twin, stress tests are conducted on the candidate reconstruction strategy with the highest score under various preset disturbance scenarios to obtain the performance retention of the candidate reconstruction strategy under each disturbance scenario. The preset disturbance scenarios include sudden fluctuations in pedestrian load and / or abrupt changes in outdoor meteorological parameters. The performance retention is the percentage of time that the comfort constraint can still be met under the disturbance.

[0157] If the performance retention of the alternative refactoring strategies meets the standard in each of the disturbance scenarios, then the architecture corresponding to the alternative refactoring strategy with the highest score will be used as the refactored system architecture.

[0158] Specifically, extreme condition tests are conducted on the highest-scoring candidate refactoring strategies: A disturbance scenario library is constructed: multiple typical disturbance modes are predefined: sudden fluctuations in pedestrian load: simulating scenarios such as the sudden disbandment of a large meeting or peak restaurant hours, with pedestrian flow increasing by 50%-100% within 5 minutes; abrupt changes in outdoor meteorological parameters: simulating cold fronts, sudden changes in solar radiation, etc., with temperature changing by 3-5°C within 10 minutes; mixed disturbance scenarios: the worst-case scenario of multiple disturbances occurring simultaneously; test execution: each disturbance scenario is applied sequentially in the digital twin, with each scenario's simulation duration covering the entire control cycle; performance retention is calculated: quantifying the strategy's performance under disturbances; comfort constraint checks: monitoring whether indoor temperature, CO2 concentration, etc., remain within the set range throughout the simulation; time percentage calculation: performance retention = (number of time points satisfying comfort constraints / total number of time points) × 100%; a threshold is set (e.g., ≥95%), below which the strategy is considered insufficiently robust; if all targets are met, the system architecture corresponding to the strategy is formally determined as the refactored system architecture.

[0159] In a preferred embodiment, after obtaining the performance retention of the alternative reconstruction strategy under each of the disturbance scenarios, the method further includes the following steps:

[0160] If the performance retention of the alternative reconstruction strategy fails to meet the standard under any of the preset disturbance scenarios, then based on the characteristics of the preset disturbance scenarios that fail to meet the standard, the key devices that cause performance degradation and their operating periods are deduced in reverse in the digital twin.

[0161] Based on the equipment control command sequence obtained from the previous control cycle, a dynamic compensation command sequence for the key equipment during the operating period is generated. The dynamic compensation command sequence includes a progressive adjustment curve for the equipment output power and switching timing optimization.

[0162] The dynamic compensation instruction sequence is fused with the alternative reconstruction strategies to form an enhanced reconstruction strategy;

[0163] The enhanced reconstruction strategy is verified in the digital twin, and the architecture corresponding to the verification is determined as the reconstructed system architecture.

[0164] Specifically, the key equipment and the period of impact are deduced in reverse based on the disturbance characteristics:

[0165] In-depth analysis of disturbance characteristics: The system first performs multi-dimensional feature analysis on disturbance scenarios that do not meet the standards. This includes analyzing the magnitude of the disturbance, the rate of change, the duration, and the propagation path of the disturbance within the building space. Special attention is paid to the directionality of the disturbance, such as whether the temperature rises or falls suddenly, or whether people rush in or disperse.

[0166] Sensitivity Mapping: A sensitivity analysis algorithm based on mathematical adjoint equations is run within the digital twin. This method constructs an adjoint model corresponding to the original system model, calculating the sensitivity coefficients of all device parameters on system performance in a single operation. The system identifies critical devices whose sensitivity coefficients exceed preset thresholds; the state changes of these devices play a decisive role in system performance degradation.

[0167] Precisely pinpointing performance degradation periods: Employing sliding time window analysis technology with a five-minute time resolution, this method precisely scans the specific time intervals during the entire simulation process where system performance begins to deteriorate, reaches its worst point, and gradually recovers. By comparing system response characteristics within multiple time windows, it accurately identifies the periods during which critical equipment experiences the greatest negative impact.

[0168] Specifically, a dynamic compensation instruction sequence is generated:

[0169] Progressive power adjustment curve design: Based on the equipment control command sequence obtained from the previous control cycle, a smooth power compensation curve is designed for the identified key equipment during its operating period. An S-shaped transient function is used to generate the adjustment trajectory of the equipment output power, ensuring that the power change is continuous and smooth rather than abrupt. This design allows the equipment output power to gradually increase to the compensation value at a preset rate and smoothly return to the normal level after the operating period ends. The system sets personalized maximum rate of change limits for different equipment based on their dynamic characteristics to prevent excessively rapid power changes from causing mechanical or electrical shocks to the equipment.

[0170] Multi-device collaborative timing optimization: Based on critical path analysis, the timing of actions of multiple devices requiring compensation is optimized and orchestrated. The system establishes a dependency graph of device actions to avoid system oscillations or resource conflicts that may be caused by multiple devices acting simultaneously. By introducing appropriate startup delays and action intervals, the execution of compensation instructions is ensured to be reasonably distributed in the time dimension, thereby guaranteeing the stability of the system throughout the compensation process.

[0171] Specifically, this integration forms an enhanced restructuring strategy:

[0172] Deep integration of control logic: The designed dynamic compensation instruction sequence is organically integrated with the original alternative reconfiguration strategies to form a hierarchical control architecture. In this architecture, the upper layer consists of global optimization instructions based on model predictive control, responsible for the overall system performance; the lower layer consists of local compensation instructions based on real-time state feedback, specifically addressing performance degradation under specific disturbances. The instructions at both levels are synthesized through a unified coordination mechanism to ensure consistency in control objectives.

[0173] Parameter Coordination and Conflict Resolution: The system establishes a dedicated parameter coordination module to detect and resolve potential conflicts between the compensation command and the original command. By setting priority rules and constraints, it ensures that no contradictory control requirements arise during command fusion. For detected potential conflicts, the system automatically adjusts based on multi-objective optimization principles to find the optimal compromise solution that satisfies all constraints.

[0174] Specifically, verification and confirmation in digital twins:

[0175] Full Regression Testing: The enhanced reconstruction strategy is fully validated by re-implementing it within the digital twin. First, the previously failed perturbation scenarios are rerun, with a focus on observing whether the system's performance under the same perturbation conditions meets the predetermined standards after the addition of the compensation mechanism.

[0176] Boundary performance testing: Test the stability of the enhancement strategy under more stringent operating conditions, including extreme tests such as applying multiple perturbations simultaneously, extending the duration of perturbations, and increasing the amplitude of perturbations, to verify the robustness of the strategy under extreme conditions.

[0177] Long-term stability testing: This involves simulating the enhancement strategy running continuously for multiple control cycles to check for accumulated errors or gradual performance degradation. Only enhancement strategies that pass all these verification stages and maintain a performance retention rate of over 97% will be ultimately approved and adopted.

[0178] This implementation method, through precise etiological diagnosis and targeted compensation design, can transform most substandard reconfiguration strategies into effective and usable control schemes, significantly reducing the manpower and material resources consumed by strategy abandonment and redesign. The system can maintain stable environmental parameters and reliable equipment operation when facing abnormal operating conditions such as sudden changes in population flow and drastic weather conditions.

[0179] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A smart building environment adaptive control method based on digital twins and AI, characterized in that, Includes the following steps: Based on time-aligned multi-source system data, dynamic features are extracted to construct a dynamic feature set; The multi-source system data includes environmental data, equipment operation data, and pedestrian flow data; the dynamic characteristics include thermal inertia delay characteristics and pedestrian flow impact conduction characteristics. The multi-source system data is input into the digital twin, and the key environmental parameters for multiple future time steps are iteratively simulated using the current building equipment status as boundary conditions, outputting a multi-step prediction sequence; the key environmental parameters include indoor temperature and CO2 concentration. The multi-step prediction sequence is input into the model prediction controller. With the goal of minimizing the overall energy consumption during the prediction period and the constraint that the key environmental parameters do not exceed the comfort range, the sequence of equipment control instructions for the first preset time period in the future is obtained. The first equipment control command in the equipment control command sequence is sent to the field actuator to achieve adaptive control of the building environment; The digital twin is a physical information neural network constructed by fusing the dynamic feature set with the building information physical model; the building information physical model is obtained by discretizing the thermodynamic differential equation, and the loss function of the physical information neural network includes data fitting terms and physical equation constraint terms. Based on time-aligned multi-source system data, thermal inertial delay features are extracted, including the following steps: Based on the ambient temperature sequence and HVAC power sequence within the target historical time period, the mutual information value at different time offsets is calculated, and the time offset that maximizes the mutual information value is taken as the thermal inertia delay time of the building. For each acquisition moment, the thermal inertia delay characteristic is obtained by combining the real-time ambient temperature acquired at that acquisition moment with the HVAC energy consumption that is delayed by the thermal inertia delay time at that acquisition moment. The mutual information value is obtained by the following formula: ； in, Represents mutual information value; Indicates the time offset; T represents the ambient temperature sequence; t represents the real-time ambient temperature; p represents the HVAC power sequence; The temperature value t represents the temperature at the i-th data acquisition time. i Delay power value at the j-th acquisition time The joint probability of simultaneous occurrence; Indicates the edge probability of temperature; The value represents the edge probability of delayed power; N represents the total number of data points in the ambient temperature sequence T; and M represents the total number of data points in the HVAC power sequence p.

2. The adaptive control method for smart building environment based on digital twins and AI according to claim 1, characterized in that: Before calculating the mutual information values ​​at different time offsets based on the ambient temperature sequence and HVAC power sequence within the target historical time period, the following steps are included: Obtain the current day's pattern feature vector to determine the current building operation mode category; the pattern feature vector includes the daily average outdoor temperature, daily HVAC energy consumption peak, and a histogram of pedestrian density distribution; Based on the current building operation mode category, the corresponding target historical time period is obtained.

3. The adaptive control method for smart building environment based on digital twins and AI according to claim 2, characterized in that: The process of obtaining the current day's pattern feature vector and determining the current building operation mode category includes the following steps: Obtain the pattern feature vector for the current day, retrieve and traverse the historical pattern database to obtain the building operation mode category corresponding to the pattern feature vector for the current day, and take it as the current building operation mode category; The historical pattern database includes multiple historical time periods and the building operation mode category corresponding to each historical time period.

4. The adaptive control method for smart building environment based on digital twins and AI according to claim 1, characterized in that: The method also includes the following steps: While the model predictive controller solves the equipment control command sequence, based on the digital twin, the expected operating state of key equipment when executing the equipment control command sequence is simulated synchronously. The expected operating state includes the expected load rate of the compressor and the expected speed of the fan. After obtaining the sequence of device control commands within the first preset time period in the future, the method further includes the following steps: Determine whether any of the critical devices continuously exceeds its safe operating boundary during the first preset time period; The step of issuing the first equipment control command in the equipment control command sequence to the field actuator includes the following steps: If not, the first equipment control command in the equipment control command sequence will be sent to the field actuator.

5. The adaptive control method for smart building environment based on digital twins and AI according to claim 4, characterized in that: After determining whether any of the critical devices continuously exceeds its safe operating boundary within the first preset time period, the method further includes the following steps: If so, the affected environmental control loop is identified, and based on the digital twin, the control of the environmental control loop is dynamically reconstructed to the reconstructed system architecture in the virtual environment. The reconstructed system architecture includes one or more healthy backup devices or redundant control loops. Using the reconstructed system architecture as the updated boundary conditions, the model predictive controller is rerun to solve for the sequence of alternative device control instructions within the second preset time period in the future. The first equipment control command in the sequence of alternative equipment control commands is sent to the field actuator, and an equipment early warning and system reconfiguration report is generated simultaneously.

6. The adaptive control method for smart building environment based on digital twins and AI according to claim 5, characterized in that: The process of dynamically reconstructing the control rights of the environmental regulation loop to the reconstructed system architecture in a virtual environment based on the digital twin includes the following steps: Based on the system's device topology, at least two alternative reconfiguration strategies are generated to allocate the load of the environmental control loop to healthy standby equipment or redundant control loops. For each of the alternative reconfiguration strategies, a forward simulation is performed in the digital twin to simulate the execution of the alternative device control command sequence and to calculate the simulation performance index of each of the alternative reconfiguration strategies. The simulation performance index includes the impact intensity of the switching process on comfort, the additional energy consumption cost, and the device lifespan reduction factor. Based on the aforementioned performance metrics, each of the candidate reconstruction strategies is comprehensively scored. The restructured system architecture is obtained based on the highest-rated alternative refactoring strategy.

7. The adaptive control method for smart building environment based on digital twins and AI according to claim 6, characterized in that: The process of obtaining the restructured system architecture based on the highest-scoring candidate restructure strategy includes the following steps: Based on the digital twin, stress tests are conducted on the candidate reconstruction strategy with the highest score under various preset disturbance scenarios to obtain the performance retention of the candidate reconstruction strategy under each disturbance scenario. The preset disturbance scenarios include sudden fluctuations in pedestrian load and / or abrupt changes in outdoor meteorological parameters. The performance retention is the percentage of time that the comfort constraint can still be met under the disturbance. If the performance retention of the alternative refactoring strategies meets the standard in each of the disturbance scenarios, then the architecture corresponding to the alternative refactoring strategy with the highest score will be used as the refactored system architecture.

8. The adaptive control method for smart building environment based on digital twins and AI according to claim 7, characterized in that: After obtaining the performance retention of the alternative reconstruction strategy under each of the disturbance scenarios, the method further includes the following steps: If the performance retention of the alternative reconstruction strategy fails to meet the standard under any of the preset disturbance scenarios, then based on the characteristics of the preset disturbance scenarios that fail to meet the standard, the key devices that cause performance degradation and their operating periods are deduced in reverse in the digital twin. Based on the equipment control command sequence obtained from the previous control cycle, a dynamic compensation command sequence for the key equipment during the operating period is generated. The dynamic compensation command sequence includes a progressive adjustment curve for the equipment output power and switching timing optimization. The dynamic compensation instruction sequence is fused with the alternative reconstruction strategies to form an enhanced reconstruction strategy; The enhanced reconstruction strategy is verified in the digital twin, and the architecture corresponding to the verification is determined as the reconstructed system architecture.

Images