Building energy management system based on digital twinning

Abstract

The invention relates to the technical field of building energy management, and discloses a building energy management system based on digital twinning. The system obtains equipment operation parameters through a real-time energy consumption data acquisition module by means of a sensor network, synchronizes the equipment operation parameters to a digital twin modeling module, constructs a synchronous virtual energy model and outputs an overall energy efficiency characteristic spectrum. The feature sampling period setting module determines a time node and a period of feature extraction; the operation characteristic spectrum decomposition module performs multi-scale frequency domain decomposition on the spectrum to generate an operation spectrum characteristic vector set; the frequency domain entropy screening module screens out a feature vector subset through an optimization unit; the load characteristic interaction matching module calculates an energy load deviation coefficient through a cross-domain characteristic interaction network; the energy consumption change rate calculation module outputs a real-time energy consumption change rate in combination with the cycle; and the multi-equipment cooperative regulation and control module generates a cooperative regulation and control instruction for a heating ventilation air-conditioning system, an illumination system and an elevator system so as to realize high-efficiency energy management of the building.

Description

Technical Field

[0001] This invention relates to the field of building energy management technology, specifically a building energy management system based on digital twins. Background Technology

[0002] With the acceleration of urbanization, buildings, as the main carriers of energy consumption, have their energy management efficiency directly related to the rational utilization and sustainable development of energy resources. Currently, traditional building energy management methods mostly rely on the independent monitoring and control of single devices, lacking the ability to perceive and coordinate the overall energy consumption status of the building in real time. In actual operation, due to the large variety of equipment and complex operating modes within buildings, there are significant correlations and dynamic interactions in energy consumption between different devices. Local control of a single device often fails to optimize the overall energy efficiency of the building. In existing technologies, while some energy management systems have incorporated data acquisition and analysis functions, data acquisition is often limited to single-point sampling at fixed time intervals, failing to capture the dynamic changes in energy consumption and resulting in a lag between energy models and actual operating conditions. Furthermore, in the feature extraction and analysis stages, traditional methods often employ time-domain analysis or single-scale frequency-domain analysis, making it difficult to comprehensively uncover the multi-scale characteristics of building energy consumption, leading to significant discrepancies between energy load forecasts and actual demand. Existing systems lack effective consideration of the collaborative relationships between multiple devices during energy regulation. Regulation strategies are often based on experience and cannot be dynamically adjusted according to real-time energy consumption changes, leading to widespread energy waste. Especially in scenarios with large fluctuations in energy load, traditional systems struggle to respond quickly to load changes, easily resulting in a mismatch between energy supply and demand, further reducing the energy management efficiency of buildings. Summary of the Invention

[0003] The purpose of this invention is to provide a building energy management system based on digital twins to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a building energy management system based on digital twins, the system comprising: The real-time energy consumption data acquisition module is used to acquire the operating parameters of multiple energy-consuming devices within a selected time window through a sensor network deployed in the building, and synchronize the operating parameters to the digital twin modeling module; The digital twin modeling module is used to construct a virtual energy model synchronized with the physical building status based on the operating parameters transmitted by the real-time energy consumption data acquisition module, and output the overall energy efficiency characteristic map of the building. The feature sampling period setting module is used to determine the start time node and end time node of feature extraction according to the preset energy management strategy, and to calculate the feature sampling period between the end time node and the start time node. The feature spectrum decomposition module is used to receive the overall energy efficiency feature map of the building output by the digital twin modeling module, and perform multi-scale frequency domain feature decomposition on it to generate a set of building operation spectrum feature vectors; The frequency domain entropy filtering module is used to input the set of building operation spectrum feature vectors generated by the operation feature spectrum decomposition module into the feature optimization unit based on frequency domain information entropy, and filter to obtain a subset of optimized building operation spectrum feature vectors; The load feature interaction matching module is used to input a subset of the optimized building operation spectrum feature vector output by the frequency domain entropy value filtering module into a cross-domain feature interaction network based on tensor convolution to calculate and generate the building energy load deviation coefficient. The energy consumption change rate calculation module is used to divide the building energy load deviation coefficient generated by the load feature interaction matching module by the feature sampling period calculated by the feature sampling period setting module, and output the real-time energy consumption change rate of the building. The multi-device collaborative control module is used to compare the real-time energy consumption change rate of the building output by the energy consumption change rate calculation module with the preset energy consumption change rate threshold, and generate collaborative control instructions for the HVAC system, lighting system and elevator system based on the comparison results.

[0005] Preferably, the frequency domain entropy value filtering module includes: The spectrum feature entropy calculation unit is used to calculate the frequency domain information entropy value of each spectrum feature vector in the set of spectrum feature vectors of the building operation, forming a set of spectrum feature entropy values ​​of the building; An entropy reference vector generation unit is used to calculate the center vector of the entropy distribution of the building spectrum feature entropy set as a reference entropy vector; Entropy deviation measurement unit is used to calculate the statistical deviation between the benchmark entropy vector and each frequency domain information entropy value in the building spectrum feature entropy value set, and generate a building spectrum entropy deviation measurement value set. The feature filtering execution unit is used to determine a subset of the optimized building operation spectrum feature vector based on the comparison results of each deviation metric in the set of building spectrum entropy deviation metric values ​​with a preset entropy deviation threshold.

[0006] Preferably, the spectral feature entropy calculation unit includes: The frequency domain energy distribution analysis subunit is used to analyze the frequency band energy distribution of each spectral feature vector in the set of spectral feature vectors of the building operation, and generate the corresponding frequency domain energy distribution vector. The cross-entropy calculation subunit is used to calculate the relative entropy value between each frequency domain energy distribution vector and a preset reference energy distribution vector, thereby obtaining the set of spectral characteristic entropy values ​​of the building.

[0007] Preferably, the frequency domain energy distribution analysis subunit is specifically used for: Extract a specified spectral feature vector from the set of spectral feature vectors of the building's operation; The energy intensity of the specified spectral feature vector in different frequency bands is normalized to generate the corresponding frequency domain energy distribution vector.

[0008] Preferably, the cross-entropy calculation subunit is specifically used for: Calculate the energy percentage of the frequency domain energy distribution vector in each frequency band; Calculate the logarithmic product of the energy proportion of the frequency domain energy distribution vector in each frequency band and the reference proportion of the reference energy distribution vector in the corresponding frequency band; The logarithmic product is weighted and summed, and then negative to generate the frequency domain information entropy value in the set of building spectral feature entropy values.

[0009] Preferably, the entropy deviation measurement unit includes: An entropy difference calculation subunit is used to calculate the dimension-wise difference between the reference entropy vector and each frequency domain information entropy value; A covariance matrix generation sub-unit is used to calculate the covariance matrix of the set of spectral characteristic entropy values ​​of the building and to obtain its inverse matrix. The Mahalanobis distance calculation subunit is used to calculate the statistical deviation value in the set of building spectrum entropy deviation measurement values ​​based on the difference value output by the entropy difference calculation subunit and the inverse covariance matrix output by the covariance matrix generation subunit.

[0010] Preferably, the load characteristic interactive matching module includes: The feature tensor construction unit is used to reconstruct a subset of the optimized building operation spectrum feature vector into a three-dimensional feature tensor. Cross-domain convolutional kernel generation unit is used to generate learnable feature-interactive convolutional kernels based on the historical energy consumption patterns of buildings; The feature interaction response unit is used to perform tensor convolution operation between the three-dimensional feature tensor and the feature interaction convolution kernel to output the building energy load interaction response feature map. The deviation coefficient generation unit is used to perform global pooling processing on the building energy load interaction response feature map to generate the building energy load deviation coefficient.

[0011] Preferably, the feature interaction response unit is specifically used for: The three-dimensional feature tensor is sliced ​​in both the spatial and spectral dimensions. The sliced ​​feature subset is convolved block by block with the feature interaction convolution kernel; The convolution calculation results are stitched together to generate the building energy load interaction response feature map.

[0012] Preferably, the multi-device collaborative control module includes: The HVAC control submodule is used to adjust the chiller supply water temperature setting value according to the excess range when the real-time energy consumption change rate of the building exceeds the preset energy consumption change rate threshold. The lighting system control submodule is used to dynamically reduce the lighting brightness benchmark value based on the area population density data when the real-time energy consumption change rate of the building exceeds the preset energy consumption change rate threshold. The elevator system control submodule is used to extend the elevator standby state switching time and optimize the group control scheduling algorithm when the real-time energy consumption change rate of the building exceeds the preset energy consumption change rate threshold.

[0013] Preferably, the multi-device collaborative control module further includes: The control effect feedback unit is used to collect the actual energy consumption data after the HVAC control submodule, lighting system control submodule and elevator system control submodule are executed; The digital twin model update unit is used to update the virtual energy model parameters in the digital twin modeling module based on the actual energy consumption data collected by the regulation effect feedback unit. A threshold adaptive adjustment unit is used to recalculate the energy consumption change rate threshold based on the updated virtual energy model.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By leveraging the synergy of the real-time energy consumption data acquisition module and the digital twin modeling module, a virtual energy model synchronized with the physical building's status is constructed. This model intuitively presents the overall energy efficiency characteristics of the building, providing comprehensive and accurate status data for energy management. The feature sampling period setting module flexibly determines the sampling period based on energy management strategies, ensuring that subsequent feature analysis better aligns with actual management needs and avoiding information redundancy or loss that may result from fixed-period sampling. The feature spectrum decomposition module performs multi-scale frequency domain feature decomposition on the overall energy efficiency feature map, which can deeply explore the energy operation characteristics at different time scales. The generated set of spectral feature vectors contains richer dynamic information on energy consumption. The frequency domain entropy value filtering module uses the feature optimization unit of frequency domain information entropy to filter out key feature subsets from a large number of feature vectors, reducing the interference of irrelevant information and improving the targeting and efficiency of subsequent analysis. The load feature interaction matching module employs a cross-domain feature interaction network based on tensor convolution, which can effectively capture the correlation between the energy loads of different devices. The calculated energy load deviation coefficient can accurately reflect the difference between the actual load and the expected load. The energy consumption change rate calculation module combines the load deviation coefficient with the feature sampling period, and the resulting real-time energy consumption change rate can dynamically reflect the changing trend of energy consumption, providing timely dynamic reference for control decisions. The multi-device collaborative control module generates collaborative control commands for HVAC, lighting, and elevator systems based on a comparison of real-time energy consumption change rates with preset thresholds. This enables coordinated adjustments between multiple devices, making energy allocation more aligned with actual needs. The entire system forms a closed loop from data acquisition and model building to feature analysis and control execution. The modules work closely together to adapt to energy management needs in different scenarios, making energy utilization in buildings more adaptable and coordinated. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the building energy management system based on digital twins as described in this invention. Figure 2 Flowchart for the frequency domain entropy value filtering module; Figure 3 A flowchart showing the deviation of entropy values ​​from the measurement unit; Figure 4 A flowchart for the load characteristic interactive matching module; Figure 5 This is a flowchart showing the feedback from the multi-device collaborative control module. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 This invention provides a building energy management system based on digital twins, the system comprising: The real-time energy consumption data acquisition module continuously acquires operating parameters of energy-consuming equipment through a sensor network deployed within the building. This sensor network includes temperature sensors, power meters, and equipment status monitors, covering HVAC systems, lighting systems, and elevator systems. Within a selected time window, the sensors collect data on the equipment's current values, power factor, flow parameters, and operating status codes at fixed intervals. After verification, the collected data is synchronously transmitted to the digital twin modeling module via an industrial bus protocol. The verification mechanism includes data integrity checks and timestamp alignment.

[0018] The digital twin modeling module receives real-time operating parameters and executes the modeling process: based on the building's geometric model and physical rules, it maps the real-time parameters to the virtual space; it constructs a virtual energy model that is dynamically synchronized with the physical building, which includes equipment energy consumption mapping relationships, thermodynamic transfer links, and electrical topology; and it outputs an overall building energy efficiency characteristic map, which integrates the spatial dimension of energy intensity distribution and the temporal dimension of load change trends.

[0019] The feature sampling period setting module analyzes the preset energy management strategy and determines the start and end time nodes for feature extraction. The energy management strategy includes rules such as peak-valley electricity price period division and seasonal operation modes. The module calculates the time difference between the start and end nodes to generate the feature sampling period value, which serves as the system's global time base reference.

[0020] Run the feature spectrum decomposition module, input the energy efficiency feature map output by the digital twin modeling module, and perform multi-scale frequency domain decomposition: use a combination of Fourier transform and wavelet analysis to decompose the time-domain energy efficiency features into low-frequency, mid-frequency and high-frequency components; generate a set of building operation spectrum feature vectors containing spectral energy distribution, frequency band center frequency and spectral kurtosis.

[0021] The frequency domain entropy filtering module inputs the set of spectral feature vectors into the feature optimization unit: based on the frequency domain information entropy theory, it calculates the energy distribution dispersion of each feature vector; by comparing entropy thresholds, it filters feature vectors with entropy values ​​lower than a preset threshold, forming an optimized subset of feature vectors. This subset retains the key frequency domain components that reflect the core energy consumption pattern.

[0022] The load feature interaction matching module reconstructs the optimized feature subset into a multi-dimensional tensor structure, which is then input into the cross-domain feature interaction network. The network structure includes tensor convolutional layers and feature fusion layers, which capture the energy consumption correlation characteristics between devices through learnable convolutional kernels. The output is the building energy load deviation coefficient, which represents the degree of statistical deviation between the real-time energy consumption status and the baseline pattern.

[0023] The energy consumption change rate calculation module performs algebraic operations: it divides the energy load deviation coefficient generated by the load characteristic interactive matching module by the sampling period value generated by the characteristic sampling period setting module, and outputs the real-time energy consumption change rate of the building. The calculation result is measured in units of change per unit time.

[0024] The multi-device collaborative control module compares the real-time energy consumption change rate with a preset threshold. When the rate exceeds the threshold, it generates a multi-system collaborative command. This command sends a temperature setpoint adjustment command to the HVAC system; a brightness adjustment command to the lighting system; and a standby strategy and group control parameter update command to the elevator system. Command transmission uses a proprietary communication protocol to ensure that control signals take effect in real time.

[0025] Example 1: See Figure 2 The frequency domain entropy value filtering module performs a multi-stage data processing flow on the set of building operation spectrum feature vectors generated by the running feature spectrum decomposition module. This process includes four main units: the spectrum feature entropy value calculation unit performs information entropy analysis on the input set, the entropy value benchmark vector generation unit establishes the evaluation benchmark, the entropy value deviation measurement unit quantifies the reliability of the features, and the feature filtering execution unit completes the construction of the final optimized subset.

[0026] The spectral feature entropy calculation unit receives an input set consisting of 128-dimensional spectral feature vectors, each vector corresponding to the building energy frequency characteristics of a specific time segment. This unit iterates through all vector elements in the set. For a single vector being processed, it first performs frequency band energy distribution analysis. Specifically, the vector is divided into fixed intervals in the frequency domain, including a 0-10Hz low-frequency band, a 10-30Hz mid-frequency band, and a 30-50Hz high-frequency band. The signal energy integral value of each frequency band is calculated to generate the original energy vector. Then, normalization is performed: the proportion of energy values ​​in each frequency band in the total is calculated, ensuring that the sum of the energy proportions of the three frequency bands is 1. This operation converts the original energy vector into a standard frequency domain energy distribution vector, whose elements represent the weights of different frequency bands in the overall energy distribution.

[0027] The cross-entropy calculation unit loads a preset baseline energy distribution vector as a reference standard. This baseline value is derived from historical data statistical analysis under normal building operation conditions. For the aforementioned generated frequency domain energy distribution vector, cross-entropy calculation is performed. The calculation process includes numerical preprocessing, component calculation, and result synthesis: First, the values ​​of each element in the baseline vector are checked; if a zero value appears, it is replaced with a very small positive number to avoid mathematical errors. Next, the product of the target vector element and the corresponding element of the baseline is calculated for each frequency band, where the second factor is the logarithm of the baseline element. Subsequently, a preset weighting coefficient is applied to the product results of each frequency band. Finally, the weighted results are summed and negative to generate an entropy scalar representing the degree of anomaly in the energy distribution of the current spectral feature vector. The entropy scalars of all vectors are stored in a temporary storage area in the order of processing, forming a set of building spectral feature entropy values.

[0028] The entropy benchmark vector generation unit reads all elements of the entropy value set. By calculating the arithmetic mean dimension-wise, it generates a benchmark vector with the same dimensions as the entropy value elements. This vector represents the expected entropy level of the building's energy distribution across frequency bands under normal conditions, serving as a reference center for subsequent deviation measurements. The calculation process employs a batch processing mode to avoid computational delays caused by element-wise operations. For example, in a 128-dimensional entropy space, the mean of all entropy values ​​in the first dimension, the mean of all entropy values ​​in the second dimension, and so on, until the mean calculation for all dimensions is completed, finally synthesizing the benchmark vector.

[0029] The entropy deviation measurement unit performs a two-stage quantitative analysis. First, difference calculation is performed: the numerical values ​​of each dimension of each element in the entropy value set are subtracted from the corresponding dimension values ​​of the baseline vector, generating a set of difference result vectors. This process produces 128 difference vectors with the same dimensions as the original entropy value elements. Next, covariance matrix calculation is performed: a 128×128-dimensional covariance matrix is ​​constructed based on the entropy value set to characterize the interconnected changes in entropy values ​​across different dimensions. The inverse covariance matrix is ​​obtained through matrix factorization to correct for the correlation effects between different dimensions. Finally, Mahalanobis distance calculation is performed: for each difference vector, a linear transformation of the inverse covariance matrix is ​​performed, followed by a dot product with the difference vector itself. The square root of the result yields the standardized deviation value. This result is stored in the building spectrum entropy deviation measurement set, and its magnitude is positively correlated with the degree of anomaly in the spectrum feature vector.

[0030] The feature selection execution unit loads a preset set of entropy deviation thresholds and deviation metrics. It initiates a traversal process according to a preset time period: initializing the optimized subset as an empty set; sequentially reading elements from the deviation metric set; comparing the current element value with the threshold; and adding the corresponding original spectral feature vector to the optimized subset for elements less than or equal to the threshold. This process continues until all 128 deviation metric elements have been processed. The final output set contains only spectral feature vectors with deviation metrics that meet the requirements, and its element count is approximately 60%-70% of the original set. The optimized subset is transmitted to the load feature interaction matching module via the data bus, simultaneously triggering a data cache cleanup operation to release storage resources for the next processing cycle.

[0031] The entire processing flow is designed as a pipelined architecture: 10 milliseconds after the spectral feature entropy calculation unit starts, the entropy baseline vector generation unit begins loading intermediate data; when the entropy baseline vector is 85% complete, the entropy deviation metric unit starts the differential calculation thread; the feature filtering execution unit starts pre-scanning when the deviation metric set is 60% generated. Through multi-threaded parallel management, the total processing time for a single 128-dimensional feature vector set is controlled within 6 minutes, meeting the system's real-time requirement of a 15-minute sampling period. Data transfer between units employs a dual verification mechanism, performing data integrity and boundary value checks at the interface layer to prevent abnormal data from entering the core calculation process. Temporary data generated during processing is automatically cleared after the final result is confirmed, and the system resource monitoring module dynamically allocates the memory cache size of the computing nodes.

[0032] Example 2: The spectral feature entropy calculation unit within the frequency domain entropy filtering module completes its core operations through two execution subunits. The frequency domain energy distribution analysis subunit is responsible for the transformation and processing of the original spectral features, while the cross-entropy calculation subunit performs quantitative evaluation. The two form a serial data processing pipeline.

[0033] The frequency domain energy distribution analysis subunit receives a set of building operation spectrum feature vectors from the operation feature spectrum decomposition module. This set contains several spectrum feature vectors arranged in a time series, each vector consisting of energy intensity values ​​of different frequency components. The processing flow begins with data slice scheduling: setting the time window index pointer to point to the first vector to be processed; loading the target vector data from the memory buffer into the register array; and activating the processing core after verifying data integrity.

[0034] The frequency components of the target vector are divided into three preset frequency bands: a low-frequency band (0 Hz to 10 Hz), a mid-frequency band (10 Hz to 30 Hz), and a high-frequency band (30 Hz to 50 Hz). The frequency band division employs a fixed boundary strategy, with boundary values ​​derived from prior knowledge of the building's equipment operating characteristics. During frequency band energy integration: three accumulator registers are created corresponding to each frequency band; all component values ​​in the target vector with frequencies less than 10 Hz are input to the first accumulator; component values ​​with frequencies between 10 and 30 Hz are input to the second accumulator; and component values ​​with frequencies greater than 30 Hz are input to the third accumulator. The integration process includes a numerical filtering stage: an amplitude threshold is set to filter environmental noise interference, accumulating only valid signal components exceeding the threshold.

[0035] After obtaining the raw energy integral values ​​for the three frequency bands, a normalization transformation is performed: the sum of the values ​​in the three accumulator registers is calculated; each accumulator value is divided by the sum to obtain the proportion of energy in that frequency band to the total energy. The normalization process incorporates an exception handling mechanism: when the sum falls below a preset threshold, a data re-acquisition process is triggered; the proportion coefficient calculation result is retained to six significant decimal places. Finally, a new vector composed of the three proportion coefficients is generated, namely the frequency domain energy distribution vector, whose physical meaning represents the weight relationship of different frequency intervals in the overall energy distribution. This process is executed iteratively on all input spectral feature vectors, converting each raw vector into a corresponding three-dimensional distribution vector.

[0036] The cross-entropy calculation subunit captures the distribution vector sequence generated by the frequency domain energy distribution analysis subunit in real time. During the initialization phase, a preset reference energy distribution vector is loaded. This reference value is stored in non-volatile memory and includes the proportions of low-frequency, mid-frequency, and high-frequency reference values. The reference values ​​are derived from historical data statistical analysis of the building's normal operating conditions and are updated quarterly.

[0037] The calculation process is divided into preprocessing and core operation stages: First, data availability is verified by checking whether each component of the current distribution vector is within the (0,1) interval. Then, the baseline vector undergoes normalization preprocessing: each baseline component value is scanned, and when a zero-value component is detected, it is replaced with a preset minimal positive constant to avoid abnormalities in subsequent logarithmic operations. This minimal positive value is much smaller than the effective signal amplitude but greater than the floating-point underflow threshold.

[0038] The core computation phase is performed independently for each input distribution vector: low-frequency, mid-frequency, and high-frequency component values ​​are extracted sequentially and paired with the corresponding components of the reference vector. First-frequency band computation is performed: the low-frequency component of the distribution vector is multiplied by the logarithmic value of the low-frequency component of the reference vector; mid-frequency band computation is performed: the mid-frequency component of the distribution vector is multiplied by the logarithmic value of the mid-frequency component of the reference vector; high-frequency band computation is performed: the high-frequency component of the distribution vector is multiplied by the logarithmic value of the high-frequency component of the reference vector. All logarithmic calculations use a binary base. In special cases where the reference component is replaced by a minimum value, the logarithmic result is taken from a preset constant table.

[0039] The calculation results for the three frequency bands are fed into a weighted calculation stage: each band is multiplied by its corresponding weighting coefficient; the three weighted results are accumulated and summed in an adder array; the final result is negative and used as the frequency domain information entropy value of the vector. The weighting coefficients are set according to the importance of each frequency band in energy consumption diagnosis, with different values ​​for low-frequency, mid-frequency, and high-frequency coefficients. The output entropy value retains four decimal places and is added to the building's spectral characteristic entropy value set.

[0040] The entire calculation process employs a hardware-accelerated design: the logarithmic calculation unit uses an architecture combining high-precision lookup tables and polynomial approximation; the multiplier uses a pipelined structure to support three-channel parallel computation; and the result register uses triple redundancy storage to prevent data loss. The control unit dynamically manages computing resources: when processing low-frequency data, the mid-to-high-frequency calculation paths enter a sleep state; and the intermediate result register is automatically cleared after all frequency band calculations for the current vector are completed.

[0041] An error detection mechanism is implemented in the data transmission channel: a parity check bit is added when the frequency domain energy distribution vector enters the calculation sub-unit; a range check is performed before the result is output, and abnormal entropy values ​​that exceed the theoretical range are marked and removed. The processing flow forms a closed-loop control: the time delay of the original spectral feature vector from entering the frequency domain energy distribution analysis sub-unit to outputting the entropy value is fixed, ensuring that the processing of the entire set is completed within the system-defined 15-minute sampling period.

[0042] Example 3: See Figure 3 The core function of the entropy deviation measurement unit is to quantify the degree of deviation between the entropy values ​​of each frequency domain information in the building spectral characteristic entropy value set and the benchmark entropy value vector through statistical analysis methods. This unit consists of three key sub-units: the entropy difference calculation sub-unit is responsible for calculating local deviations, the covariance matrix generation sub-unit analyzes global statistical characteristics, and the Mahalanobis distance calculation sub-unit comprehensively evaluates the degree of deviation.

[0043] In the entropy difference calculation subunit, the source of the baseline entropy vector needs to be clarified first. This vector is obtained by the entropy baseline vector generation unit through statistical analysis of historical energy consumption data, representing the typical energy distribution pattern of a building under normal operating conditions. For each frequency domain information entropy value in the spectral feature entropy value set, it needs to be compared and analyzed with this baseline vector. Specifically, each frequency domain information entropy value is an n-dimensional vector, where n represents the total number of frequency bands divided by the system. The division of these frequency bands is determined based on the periodic characteristics of building energy consumption, and typically includes multiple frequency bands such as high frequency, mid frequency, and low frequency.

[0044] The work of the covariance matrix generation sub-unit is more complex. This sub-unit needs to handle the statistical characteristics of the entire set of spectral feature entropy values. First, it is necessary to calculate the mean vector of the entropy matrix, which represents the average level of all frequency domain information entropy values ​​across each frequency band. Then, by calculating the deviation matrix, the difference between each frequency domain information entropy value and the mean vector is obtained. This process is essentially establishing a correlation model between the various frequency bands, providing the necessary statistical basis for subsequent Mahalanobis distance calculations.

[0045] The Mahalanobis distance calculation subunit is the core of the entire deviation measurement process. This subunit uses the following formula to calculate the degree of deviation for each frequency domain information entropy value:

[0046] in, Represents the Mahalanobis distance value. This represents the vector of frequency domain information entropy values ​​that need to be evaluated. Represents the baseline entropy value vector. It is the inverse of the covariance matrix. Each symbol in this formula has a definite physical meaning: This represents the difference vector between the current energy consumption mode and the baseline mode. This is used to eliminate the correlation between frequency bands and to weight the importance of different frequency bands.

[0047] In practical calculations, the numerical stability of Mahalanobis distance requires special attention. Since the covariance matrix may exhibit ill-conditioned conditions, appropriate numerical methods, such as singular value decomposition or regularization techniques, are needed when performing inversion operations. Furthermore, to ensure computational efficiency, block-based or parallel computing methods can be employed to handle large-scale entropy sets.

[0048] The feature filtering unit operates based on Mahalanobis distance calculations. This unit needs to set a reasonable threshold to determine which frequency domain entropy values ​​should be retained and which should be discarded. Determining this threshold requires considering multiple factors, including the specific usage of the building, seasonal variations, and equipment operating status. In practical applications, this threshold can be fixed or dynamically adjusted based on historical data.

[0049] The implementation of the entire entropy deviation measurement unit requires a rigorous data processing workflow. First, the raw entropy data undergoes preprocessing, including data cleaning and normalization. Next comes the core statistical calculation process, including difference calculation, covariance matrix generation, and Mahalanobis distance calculation. Finally, there is the feature selection decision process. Each step requires strict quality control to ensure the accuracy and reliability of the calculation results.

[0050] At the system implementation level, the entropy deviation measurement unit requires efficient computational capabilities. Since building energy consumption data typically has a high sampling frequency and large data volume, optimized algorithm implementation, employing appropriate data structures and computational methods, is necessary. Furthermore, to meet real-time requirements, techniques such as incremental computation or sliding windows may be required to process continuous data streams.

[0051] From a functional perspective, the entropy deviation metric unit effectively establishes a multi-dimensional anomaly detection mechanism. By calculating Mahalanobis distance, the system can comprehensively assess the degree of anomaly in energy consumption patterns across multiple frequency bands, rather than focusing solely on changes in a single frequency band. This approach provides a more comprehensive reflection of building energy usage, improving the accuracy and robustness of anomaly detection.

[0052] In actual operation, the entropy deviation measurement unit needs to work closely with other modules. For example, the virtual energy model data provided by the digital twin modeling module is the basis for entropy calculation, and the results of feature selection will directly affect subsequent load feature interaction matching. Therefore, the data interfaces and interaction protocols between modules need to be fully considered during system design and implementation.

[0053] From an algorithmic perspective, Mahalanobis distance calculation provides an anomaly detection method based on statistical distribution. Compared to simple threshold comparisons or rule-based judgments, this method is better suited to the complexity and diversity of building energy consumption. Furthermore, by introducing the covariance matrix, the system can automatically learn and adapt to the correlations between different frequency bands without requiring manually defined complex judgment rules.

[0054] Regarding parameter settings, the entropy deviation from the measurement unit needs to be adjusted according to the specific application scenario. For example, different time windows can be used to calculate the covariance matrix, and the threshold for Mahalanobis distance can also be dynamically adjusted according to the season or usage scenario. Optimization of these parameters requires combining historical data and practical operating experience.

[0055] System maintenance and updates are also important aspects to consider during implementation. As building usage patterns change or equipment ages, the baseline entropy vector and covariance matrix may need to be updated periodically. Therefore, the system needs to be designed with appropriate mechanisms to automatically or semi-automatically update these parameters.

[0056] From a computational complexity perspective, the main computational overhead of entropy deviation from the metric unit is concentrated in the generation and inversion of the covariance matrix. For a system with n frequency bands, the covariance matrix has an n×n dimension. When n is large, these matrix operations may become a performance bottleneck. Therefore, in practical implementations, dimensionality reduction techniques or sparse matrix processing methods may be needed to improve computational efficiency.

[0057] Data visualization is also an aspect worth paying attention to during implementation. Although it is not a core function, a good visualization interface can help managers better understand the results of entropy deviation measurements. For example, heatmaps can be used to show the degree of deviation in different frequency bands, or trend graphs can be used to show how Mahalanobis distance changes over time.

[0058] Regarding exception handling mechanisms, the system needs a well-designed process to handle special situations that may arise during computation. For example, when input data contains missing values, there needs to be an appropriate strategy for imputation or ignoring them; when the covariance matrix is ​​not invertible, there needs to be a backup calculation method. These exception handling mechanisms are crucial for ensuring the robustness of the system.

[0059] From a system integration perspective, the entropy deviation measurement unit needs to interact with other management systems in the building. For example, it may need to obtain real-time operational status data from equipment monitoring systems or provide optimization suggestions to energy management systems. These cross-system integrations require consideration of multiple aspects, including data formats, communication protocols, and security mechanisms.

[0060] Example 4: See Figure 4 In the energy management system of a commercial complex, the optimized spectral feature vector subset, processed by the frequency domain entropy filtering module, contains 72 hours of continuous monitoring data. This data is sampled at 15-minute intervals, resulting in feature vectors for 288 time points. Each feature vector contains energy distribution information for six main frequency bands, corresponding to the operating characteristics of different energy devices. This data is first fed into the feature tensor construction unit for processing.

[0061] The feature tensor construction unit reassembles a one-dimensional sequence of spectral feature vectors into a three-dimensional feature tensor. During the reassembly process, the time dimension of the original data is preserved, while the frequency band dimension and equipment type dimension are added. For example, for the data of the aforementioned commercial complex, the constructed three-dimensional feature tensor has a size of 24×6×3, where 24 represents the time dimension after aggregating 288 time points by hour, 6 represents the frequency band dimension, and 3 represents the equipment type dimension (HVAC, lighting, elevator).

[0062] The cross-domain convolutional kernel generation unit is responsible for creating learnable feature-interacting convolutional kernels. These kernels are automatically generated by analyzing typical patterns in historical building energy consumption data. In a real-world system, multiple convolutional kernels of different sizes may be included to capture energy consumption features at different time scales. For example, small kernels (e.g., 3×3) are used to identify short-term fluctuation patterns, while large kernels (e.g., 7×7) are used to capture long-term trend changes. Each kernel contains adjustable weight parameters that are dynamically optimized based on actual energy consumption during system operation.

[0063] The Feature Interaction Response Unit performs operations on the 3D feature tensor and the feature interaction convolution kernel. This unit first slices the input 3D feature tensor, decomposing the complete tensor into multiple overlapping or non-overlapping sub-tensors. Each sub-tensor is computed block-by-block with its corresponding convolution kernel, generating intermediate feature maps. These intermediate feature maps are then reassembled to form a complete building energy load interaction response feature map. During this process, the system retains regions in the feature map that exhibit significant responses; these regions typically correspond to anomalous energy consumption patterns or significant operational state transitions.

[0064] The deviation coefficient generation unit performs global pooling on the feature map of the interaction response. The pooling operation is performed along the time and frequency band dimensions, compressing the three-dimensional feature map into a scalar value reflecting the overall energy load status. In practice, various pooling strategies can be employed, such as max pooling, average pooling, or hybrid pooling. For example, in the case of a commercial complex, the system uses a weighted average pooling method, assigning different weight coefficients based on the importance of different frequency bands, ultimately generating a building energy load deviation coefficient between 0 and 1. The larger this coefficient, the greater the deviation of the current energy consumption status from the normal pattern.

[0065] During system operation, the load characteristic interaction matching module needs to handle various special cases. For example, when the input data contains missing values, the module will automatically interpolate and complete the data using data from adjacent time points. When extreme outliers are detected, the system will initiate a verification process, confirming the authenticity of the anomaly through cross-comparison with other monitoring data. These mechanisms ensure that the module can produce reliable output results under various operating conditions.

[0066] The module's performance optimization is achieved primarily through two aspects: improved computational efficiency and enhanced feature extraction. Regarding computational efficiency, the system employs a parallel computing architecture to accelerate tensor operations, especially for processing large-scale feature tensors. For feature extraction, the module periodically updates the convolutional kernel parameters to adapt to gradual changes in building energy consumption patterns. This adaptive mechanism ensures that the system maintains high detection sensitivity during long-term operation.

[0067] Data visualization plays a crucial role in the load characteristic interactive matching module. The system provides various visualization tools to help managers understand the module's output. For example, the 3D heatmap of the feature tensor can intuitively display the energy consumption characteristic distribution at different times and frequency bands; the interactive response feature map can highlight abnormal energy consumption areas; and the time series plot of the deviation coefficient can reflect the dynamic changing trend of the building's energy load status. These visualization tools greatly enhance the system's interpretability and usability.

[0068] Real-time processing capability is one of the module's key features. In actual deployment, the system adopts a streaming processing architecture, enabling real-time analysis of continuously input energy consumption data. Data from each time window enters the processing flow immediately after acquisition, undergoing feature tensor construction, convolution operations, and pooling, generating deviation coefficients within a second-level latency. This real-time capability allows the system to promptly detect energy consumption anomalies and trigger control mechanisms.

[0069] Collaboration with other modules is also a crucial consideration during implementation. The load characteristic interaction matching module needs to maintain data synchronization with the digital twin modeling module to ensure that feature analysis is based on the latest virtual model. Simultaneously, the deviation coefficients output by this module need to be accurately transmitted to the energy consumption change rate calculation module to complete the subsequent energy management decision chain. This cross-module collaboration is achieved through carefully designed data interfaces and communication protocols.

[0070] System maintenance and upgrades are essential to ensure the long-term stable operation of the module. Regularly check the learning status of the convolutional kernels, evaluate the effectiveness of feature extraction, and adjust processing parameters according to changes in building usage patterns. These maintenance tasks can be completed through a combination of automated scripts and manual intervention. Regarding upgrades, the module supports flexible replacement of the convolutional kernel architecture, facilitating the subsequent introduction of more advanced deep learning models.

[0071] Security and reliability are fundamental requirements during implementation. All data interactions are encrypted, and critical computation processes incorporate redundant verification mechanisms. Modules possess self-detection capabilities, automatically switching to secure mode and notifying maintenance personnel in the event of anomalies. These designs ensure the system maintains normal basic functionality under various unforeseen circumstances.

[0072] In practical applications, the performance of the load characteristic interaction matching module is affected by various factors. The building scale and usage intensity determine the data scale of the characteristic tensor; the equipment type and operation mode affect the complexity of the characteristic interaction; and the environmental conditions and seasonal changes lead to the dynamic evolution of the energy consumption pattern. All these factors need to be fully considered during system design and parameter tuning to ensure that the module can achieve the expected effect in different application scenarios.

[0073] Example 5: Refer to Figure 5 , the core function of the multi-device collaborative control module is to coordinately control the HVAC system, lighting system, and elevator system based on the monitoring results of the real-time energy consumption change rate of the building, so as to optimize the energy use efficiency. Through the coordinated operation of multiple subsystems, this module establishes a complete control loop and can dynamically respond to the changes in the energy consumption state of the building.

[0074] The operation mechanism of the HVAC control sub-module is based on the comparison result between the real-time energy consumption change rate of the building and the preset threshold. When it is detected that the energy consumption change rate exceeds the threshold, the system will adopt a hierarchical control strategy according to the specific exceeding amplitude. The control measures mainly focus on adjusting the operation parameters of the chiller, including modifying the supply water temperature setting value, adjusting the pump operation frequency, and optimizing the fresh air ratio. The adjustment amplitude of these parameters is positively correlated with the exceeding degree of the energy consumption change rate, ensuring that the control intensity matches the actual demand. During the parameter adjustment process, the system will comprehensively consider the requirements of indoor environmental comfort to avoid affecting the building's usage function due to excessive energy conservation.

[0075] The workflow of the lighting system control sub-module integrates energy consumption monitoring and occupancy sensing data. This sub-module obtains real-time occupancy density information through a sensor network distributed in various areas of the building and correlates this data with the energy consumption change rate. When the system determines that it is necessary to reduce lighting energy consumption, it will implement a differential brightness adjustment strategy according to the actual usage conditions in different areas. Areas with frequent occupancy maintain the basic lighting demand, while areas with lower usage rates appropriately reduce the brightness reference value. The lighting control adopts a progressive adjustment method to avoid sudden brightness changes affecting the visual comfort of occupants. The system also dynamically adjusts the control strategy according to the changes in natural light conditions, making full use of daylight resources to reduce the demand for artificial lighting.

[0076] The control strategy of the elevator system control submodule focuses on optimizing and adjusting the operating mode. When the rate of energy consumption change exceeds the limit, the system will extend the elevator's standby time to reduce unnecessary operational losses. Simultaneously, the elevator group control algorithm will recalculate the optimal scheduling plan based on real-time monitored passenger flow patterns and energy consumption data, rationally allocating the transportation tasks of each elevator and improving overall operating efficiency. During off-peak hours, the system can automatically switch some elevators to energy-saving sleep mode, maintaining only the necessary number of elevators to provide basic services. These control measures effectively reduce the energy consumption of the elevator system while ensuring the building's vertical transportation needs are met.

[0077] The control effect feedback unit is responsible for collecting actual operational data from each subsystem after executing control commands. This unit continuously monitors energy consumption changes in the HVAC, lighting, and elevator systems through a building-wide sensor network, recording the operating parameters and status of key equipment. This data includes not only quantitative indicators of energy consumption but also various auxiliary parameters reflecting system operating status, such as indoor temperature and humidity, illuminance levels, and elevator response time. The frequency of feedback data collection is adapted to the dynamic characteristics of the control measures to ensure accurate capture of the system response process.

[0078] The digital twin model update unit corrects the parameters of the virtual energy model based on feedback data. This unit compares and analyzes the actual monitored system response with the model's predictions, identifies parameters with significant differences, and adjusts the relevant parameters in the model through iterative optimization algorithms. The model update process employs an incremental learning mechanism, fine-tuning only the necessary parameters each time to avoid model instability caused by large-scale parameter changes. The updated model can more accurately reflect the actual energy consumption characteristics of buildings, providing a reliable basis for subsequent control decisions.

[0079] The threshold adaptive adjustment unit recalculates the energy consumption change rate threshold based on the updated digital twin model. This unit analyzes historical energy consumption data of the building under different operating conditions, and dynamically determines the most suitable threshold level by combining this data with the current equipment operating status and environmental conditions. The threshold calculation process considers various influencing factors, including seasonal changes, workday patterns, and special event arrangements, ensuring that the set threshold adapts to the actual usage of the building. The system also establishes a threshold smoothing adjustment mechanism to avoid drastic fluctuations in control commands due to sudden threshold changes.

[0080] The multi-device collaborative control module operates according to a hierarchical decision-making architecture. The top layer is responsible for global strategy formulation, the middle layer handles subsystem coordination, and the bottom layer executes specific control commands. This architecture design ensures that each subsystem can implement optimal control based on its own characteristics, while also ensuring collaborative cooperation between different systems. The module has a robust anomaly handling mechanism that automatically activates backup plans or prompts manual intervention when control measures fail to achieve the expected results or unexpected situations occur.

[0081] Real-time performance is one of the module's key characteristics. From monitoring the rate of energy consumption change to generating and executing control commands, the entire process is completed within minutes, ensuring the system can respond promptly to rapid changes in energy consumption. This real-time performance is achieved through efficient data processing and optimized control algorithms, enabling the system to maintain stable performance in complex and ever-changing operating environments.

[0082] Data interaction with other modules is fundamental to the normal operation of the multi-device collaborative control module. The module needs to acquire virtual model data of the building from the digital twin modeling module, receive real-time monitoring results from the energy consumption change rate calculation module, and send control commands to the control systems of each device. This data interaction is achieved through standardized interface protocols, ensuring the accuracy and timeliness of information transmission. The module also has a data caching mechanism, enabling it to maintain basic control functions based on historical data during temporary communication interruptions.

[0083] The system's maintainability design takes into account the stability requirements of long-term operation. The module has built-in self-diagnostic functions, which can periodically check the operating status of each subsystem, identify potential problems, and generate maintenance prompts. All control decisions and system responses are recorded in the operation log, providing data support for troubleshooting and performance optimization. The module supports remote configuration and updates, facilitating adjustments to control parameters and strategies based on actual operating results.

[0084] In practical applications, the multi-device collaborative control module needs to adapt to various special scenarios. For example, when hosting large-scale events, the system needs to temporarily adjust control strategies to accommodate the special needs of dense crowds; under extreme weather conditions, a balance between energy efficiency and environmental comfort needs to be comprehensively considered. The module automatically detects these special situations through scene recognition algorithms and calls preset response plans to handle them. This flexibility enables the system to maintain good control performance under various complex conditions.

[0085] The module's human-machine interface design prioritizes practicality and ease of use. Operators can view system operating status, adjust control parameters, and set special scenario modes through the graphical interface. The interface provides multi-level detailed information display, including both a macro view of overall energy consumption trends and micro data on the operating status of specific equipment. All control operations are equipped with confirmation steps and safety restrictions to prevent accidental operation from causing system malfunctions.

[0086] From a system architecture perspective, the multi-device collaborative control module adopts a distributed design. Core decision-making functions are concentrated in the main control unit, while the control execution of each subsystem is distributed across local controllers. This architecture ensures both the uniformity of decision-making and improves system reliability and response speed. The module's continuous learning capability is a crucial guarantee for its long-term effective operation. The system records the implementation effect of each control measure, establishing a case library for analyzing the actual effectiveness of different strategies. Based on this historical data, the module can gradually optimize control parameters and decision rules, enabling the control effect to continuously improve over time. The learning process employs a conservative strategy to ensure the predictability and stability of system behavior.

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A building energy management system based on digital twins, characterized in that, include: The real-time energy consumption data acquisition module is used to acquire the operating parameters of multiple energy-consuming devices within a selected time window through a sensor network deployed in the building, and synchronize the operating parameters to the digital twin modeling module; The digital twin modeling module is used to construct a virtual energy model synchronized with the physical building status based on the operating parameters transmitted by the real-time energy consumption data acquisition module, and output the overall energy efficiency characteristic map of the building. The feature sampling period setting module is used to determine the start time node and end time node of feature extraction according to the preset energy management strategy, and to calculate the feature sampling period between the end time node and the start time node. The feature spectrum decomposition module is used to receive the overall energy efficiency feature map of the building output by the digital twin modeling module, and perform multi-scale frequency domain feature decomposition on it to generate a set of building operation spectrum feature vectors; The frequency domain entropy filtering module is used to input the set of building operation spectrum feature vectors generated by the operation feature spectrum decomposition module into the feature optimization unit based on frequency domain information entropy, and filter to obtain a subset of optimized building operation spectrum feature vectors; The load feature interaction matching module is used to input a subset of the optimized building operation spectrum feature vector output by the frequency domain entropy value filtering module into a cross-domain feature interaction network based on tensor convolution to calculate and generate the building energy load deviation coefficient. The energy consumption change rate calculation module is used to divide the building energy load deviation coefficient generated by the load feature interaction matching module by the feature sampling period calculated by the feature sampling period setting module, and output the real-time energy consumption change rate of the building. The multi-device collaborative control module is used to compare the real-time energy consumption change rate of the building output by the energy consumption change rate calculation module with the preset energy consumption change rate threshold, and generate collaborative control instructions for the HVAC system, lighting system and elevator system based on the comparison results.

2. The building energy management system based on digital twins according to claim 1, characterized in that, The frequency domain entropy value filtering module includes: The spectrum feature entropy calculation unit is used to calculate the frequency domain information entropy value of each spectrum feature vector in the set of spectrum feature vectors of the building operation, forming a set of spectrum feature entropy values ​​of the building; An entropy reference vector generation unit is used to calculate the center vector of the entropy distribution of the building spectrum feature entropy set as a reference entropy vector; Entropy deviation measurement unit is used to calculate the statistical deviation between the benchmark entropy vector and each frequency domain information entropy value in the building spectrum feature entropy value set, and generate a building spectrum entropy deviation measurement value set. The feature filtering execution unit is used to determine a subset of the optimized building operation spectrum feature vector based on the comparison results of each deviation metric in the set of building spectrum entropy deviation metric values ​​with a preset entropy deviation threshold.

3. The building energy management system based on digital twins according to claim 2, characterized in that, The spectral feature entropy calculation unit includes: The frequency domain energy distribution analysis subunit is used to analyze the frequency band energy distribution of each spectral feature vector in the set of spectral feature vectors of the building operation, and generate the corresponding frequency domain energy distribution vector. The cross-entropy calculation subunit is used to calculate the relative entropy value between each frequency domain energy distribution vector and a preset reference energy distribution vector, thereby obtaining the set of spectral characteristic entropy values ​​of the building.

4. The building energy management system based on digital twins according to claim 3, characterized in that, The frequency domain energy distribution analysis subunit is specifically used for: Extract a specified spectral feature vector from the set of spectral feature vectors of the building's operation; The energy intensity of the specified spectral feature vector in different frequency bands is normalized to generate the corresponding frequency domain energy distribution vector.

5. The building energy management system based on digital twins according to claim 4, characterized in that, The cross-entropy calculation subunit is specifically used for: Calculate the energy percentage of the frequency domain energy distribution vector in each frequency band; Calculate the logarithmic product of the energy proportion of the frequency domain energy distribution vector in each frequency band and the reference proportion of the reference energy distribution vector in the corresponding frequency band; The logarithmic product is weighted and summed, and then negative to generate the frequency domain information entropy value in the set of building spectral feature entropy values.

6. The building energy management system based on digital twins according to claim 5, characterized in that, The entropy deviation measurement unit includes: An entropy difference calculation subunit is used to calculate the dimension-wise difference between the reference entropy vector and each frequency domain information entropy value; A covariance matrix generation sub-unit is used to calculate the covariance matrix of the set of spectral characteristic entropy values ​​of the building and to obtain its inverse matrix. The Mahalanobis distance calculation subunit is used to calculate the statistical deviation value in the set of building spectrum entropy deviation measurement values ​​based on the difference value output by the entropy difference calculation subunit and the inverse covariance matrix output by the covariance matrix generation subunit.

7. The building energy management system based on digital twins according to claim 1, characterized in that, The load feature interactive matching module includes: The feature tensor construction unit is used to reconstruct a subset of the optimized building operation spectrum feature vector into a three-dimensional feature tensor. Cross-domain convolutional kernel generation unit is used to generate learnable feature-interactive convolutional kernels based on the historical energy consumption patterns of buildings; The feature interaction response unit is used to perform tensor convolution operation between the three-dimensional feature tensor and the feature interaction convolution kernel to output the building energy load interaction response feature map. The deviation coefficient generation unit is used to perform global pooling processing on the building energy load interaction response feature map to generate the building energy load deviation coefficient.

8. The building energy management system based on digital twins according to claim 7, characterized in that, The feature interaction response unit is specifically used for: The three-dimensional feature tensor is sliced ​​in both the spatial and spectral dimensions. The sliced ​​feature subset is convolved block by block with the feature interaction convolution kernel; The convolution calculation results are stitched together to generate the building energy load interaction response feature map.

9. The building energy management system based on digital twins according to claim 1, characterized in that, The multi-device collaborative control module includes: The HVAC control submodule is used to adjust the chiller supply water temperature setting value according to the excess range when the real-time energy consumption change rate of the building exceeds the preset energy consumption change rate threshold. The lighting system control submodule is used to dynamically reduce the lighting brightness benchmark value based on the area population density data when the real-time energy consumption change rate of the building exceeds the preset energy consumption change rate threshold. The elevator system control submodule is used to extend the elevator standby state switching time and optimize the group control scheduling algorithm when the real-time energy consumption change rate of the building exceeds the preset energy consumption change rate threshold.

10. The building energy management system based on digital twins according to claim 9, characterized in that, The multi-device collaborative control module also includes: The control effect feedback unit is used to collect the actual energy consumption data after the HVAC control submodule, lighting system control submodule and elevator system control submodule are executed; The digital twin model update unit is used to update the virtual energy model parameters in the digital twin modeling module based on the actual energy consumption data collected by the regulation effect feedback unit. A threshold adaptive adjustment unit is used to recalculate the energy consumption change rate threshold based on the updated virtual energy model.

Images