A Deep Learning-Based Decoupling Method for Multiple Building Energy Consumption Indicators

Abstract

This invention relates to the field of edge IoT technology, specifically a deep learning-based method for decoupling multiple indicators of building energy consumption. The method involves inputting periodic alignment features and a spatial adjacency matrix into an energy efficiency management analysis engine to generate a thermoelectric feature tensor mapping the spatial operating status, separating the baseline expectation matrix for routine operations from the net residual matrix of dynamic fluctuations. Compliance supervision is implemented by combining a thermal boundary mask matrix generated from the actual operating boundaries of the physical system, resulting in a boundary constraint tensor. Finally, a decoupling weight matrix is ​​constructed and subjected to a Hadamard product with the node load sequence to separate temperature control energy consumption values ​​and lighting energy consumption values, generating decoupled energy consumption assessment data for decision-making and overall planning. This invention internalizes the business flow logic and resource consumption patterns of building space through an energy efficiency management analysis engine, achieving a high degree of synergy between multi-dimensional IoT sensing data in the process of energy efficiency overall supervision and management.

Description

Technical Field

[0001] This invention relates to the field of edge IoT technology, specifically a deep learning-based method for decoupling multiple indicators of building energy consumption. Background Technology

[0002] In the energy consumption monitoring and resource management of large buildings, accurately breaking down the total global energy consumption monitoring into multiple sub-indicators such as temperature control energy consumption and lighting energy consumption (i.e., decoupling multi-dimensional energy consumption business indicators) is an important current technological application direction. Existing technologies generally adopt a purely data-driven energy consumption business prediction method, which directly monitors and collects environmental temperature time series and business node energy consumption statistical series, directly inputs these multi-dimensional data into a conventional neural network model for nonlinear feature fitting, and automatically identifies and outputs the decomposed business sub-energy consumption data. This is currently the mainstream technical means to achieve decoupling of energy consumption management indicators.

[0003] However, existing technical solutions have significant shortcomings in decoupling accuracy and physical interpretability. The underlying sequences used as input to the network suffer from temporal misalignment, failing to effectively handle the phase lag between temperature changes and power consumption response caused by building thermal inertia. Existing spatial relationships are mostly based on fixed physical distances or HVAC zoning, and their topology cannot be adjusted in real time with changes in personnel flow and business characteristics within the building. This makes it difficult to accurately match the linked energy consumption patterns in complex business flow scenarios (such as the transfer of people from the conference area to the catering area), leading to deviations in the extraction of node spatial features. The computational output of the decoupling process suffers from black-box nature and lack of constraints, and cannot be deeply adapted to the physical heterogeneous attributes of different energy consumption sub-items. This easily leads to the output of predictive values ​​that defy common sense under physical boundaries where the HVAC system is not actually operating, resulting in a lack of accuracy and physical consistency in the final generated energy consumption decoupling assessment data.

[0004] To address this, a deep learning-based method for decoupling multiple indicators of building energy consumption is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for decoupling multiple indicators of building energy consumption, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for decoupling multiple indicators of building energy consumption, comprising: Acquire environmental thermal field sequence, supply and return water temperature sequence, node load sequence and business flow map; perform sliding window translation calculation on environmental thermal field sequence and node load sequence, and output cycle alignment feature; construct spatial adjacency matrix based on business flow map; The spatial adjacency matrix and periodic alignment features are input into the energy efficiency management analysis engine for spatiotemporal aggregation, and the thermoelectric feature tensor is output. The thermoelectric feature tensor is input into a two-branch feature decoupling network. The baseline expectation matrix is ​​output through the baseline fitting branch, the bias matrix and mean matrix are output through the residual peaking branch, and the net residual matrix is ​​established through broadcast subtraction. The supply and return water temperature sequences are truncated with a preset water temperature threshold, and a thermal boundary mask matrix is ​​output. The thermoelectric feature tensor and the net residual matrix are linearly projected to output the key tensor, value tensor and query tensor. The thermal boundary mask matrix is ​​used for filtering, and dot product scaling and dot product calculation are performed to output the boundary constraint tensor. The baseline expectation matrix and boundary constraint tensor are spliced ​​together to construct the decoupling weight matrix; the decoupling weight matrix and the node load sequence are multiplied by Hadamard to output the temperature control energy consumption value and the lighting energy consumption value, thus generating energy consumption decoupling assessment data.

[0007] Preferably, the specific generation process of the periodic alignment feature includes: applying a fast Fourier transform to the environmental thermal field sequence and the node load sequence to extract the periodic thermal field evaluation quantity and the periodic electrical work evaluation quantity; calculating the cross-correlation evaluation product of the periodic thermal field evaluation quantity and the periodic electrical work evaluation quantity to obtain the cross-power spectral density tensor; processing the cross-power spectral density tensor using an inverse fast Fourier transform operation to generate a time-domain cross-correlation function sequence; extracting the peak index of the time-domain cross-correlation function sequence corresponding to each node to establish the response delay management vector; shifting the environmental thermal field sequence according to the response delay management vector to obtain the hysteresis-compensated thermal field sequence; performing constant-length sliding slice extraction calculations on the hysteresis-compensated thermal field sequence and the node load sequence respectively to obtain the windowed thermal field tensor and the windowed electrical work tensor respectively; and concatenating the windowed thermal field tensor and the windowed electrical work tensor in the channel dimension to construct the periodic alignment feature.

[0008] Preferably, the specific process of constructing the spatial adjacency matrix includes: based on the business flow graph, analyzing and extracting the spatial business flow probability and connection frequency between building space entity nodes; calculating the product of the spatial business flow probability and the connection frequency as the node collaborative management weight, and constructing an initial graph matrix accordingly; extracting the symmetric connection attribute of the initial graph matrix and establishing a symmetric connection tensor; filling the main diagonal elements of the symmetric connection tensor with preset constants to construct a self-collaborative evaluation tensor; performing numerical summation and aggregation operations on the self-collaborative evaluation tensor along the row direction to obtain a degree matrix tensor, and normalizing it to obtain a normalized degree matrix tensor; performing continuous matrix multiplication operations between the normalized degree matrix tensor and the self-collaborative evaluation tensor to generate a spatial adjacency matrix.

[0009] Preferably, the energy efficiency management analysis engine includes a spatial energy efficiency map supervision layer and a time-series energy consumption trend prediction layer. In the spatial energy efficiency map supervision layer, a preset spatial management evaluation weight tensor is used to perform linear mapping multiplication calculation on the periodic alignment features to obtain a dimensionality-reduced node feature matrix. The spatial adjacency matrix and the dimensionality-reduced node feature matrix are aggregated using standard matrix multiplication to extract a primary aggregated feature matrix. The primary aggregated feature matrix is ​​processed by nonlinear mapping calculation with leakage correction linear units to generate a primary spatial topology tensor. The primary spatial topology tensor is fed into a preset multi-head attention allocation mechanism to complete node distance allocation and obtain a weighted spatial feature tensor. Layer normalization numerical constraint calculation and residual addition operation are performed on the weighted spatial feature tensor in sequence to establish the spatial topology tensor. The spatial topology tensor is forward-passed to the time-series energy consumption trend prediction layer as a single-step topology input tensor.

[0010] Preferably, the specific process of generating the thermoelectric feature tensor includes: extracting the hidden state tensor and the core state evaluation tensor of the previous time step; concatenating the single-step topological input tensor and the hidden state tensor along the channel dimension to construct a joint spatiotemporal tensor; performing a nonlinear mapping on the joint spatiotemporal tensor using a preset set of gating weight matrices to extract the redundancy removal evaluation tensor, the trend introduction evaluation tensor, and the output gating tensor respectively; performing state update calculation by combining the trend introduction evaluation tensor, the redundancy removal evaluation tensor, and the core state evaluation tensor to obtain the current step core state evaluation tensor; calculating the activation mapping of the current step core state evaluation tensor and the Hadamard product of the output gating tensor to finally generate the thermoelectric feature tensor.

[0011] Preferably, the specific generation process of the baseline expectation matrix and the net residual matrix includes: processing the thermoelectric feature tensor using a multilayer perceptron for dimensionality reduction and fusion calculation, and extracting a shared hidden layer feature tensor; guiding the shared hidden layer feature tensor into the baseline fitting branch, and generating the baseline expectation matrix using matrix multiplication with a preset baseline fitting fully connected weight tensor; passing the shared hidden layer feature tensor through the residual peaking branch, and obtaining the bias matrix by matrix multiplication with a preset peaking fully connected weight tensor; performing mean summation calculation on the bias matrix along the feature channel dimension to obtain a residual mean scalar matrix; expanding the residual mean scalar matrix to the scale of the bias matrix along the channel dimension to construct the mean matrix; calculating the element-wise subtraction cancellation result between the bias matrix and the mean matrix to establish the net residual matrix.

[0012] Preferably, the specific generation process of the boundary constraint tensor includes: linearly mapping the thermoelectric feature tensor and the net residual matrix using a preset set of projection weight matrices to extract the key tensor, the value tensor, and the query tensor; calculating the inner product multiplication result of the query tensor and the key tensor to establish an unscaled attention score matrix; performing additive masking processing on the unscaled attention score matrix using the thermal boundary mask matrix to filter illegal boundary association positions and obtain a masking score matrix; performing dot product scaling and exponential normalization calculation on the masking score matrix to construct an attention weight normalization matrix; and generating the boundary constraint tensor by performing matrix multiplication superposition calculation between the attention weight normalization matrix and the value tensor.

[0013] Preferably, the specific process of generating energy consumption decoupling evaluation data includes: concatenating the baseline expectation matrix and the boundary constraint tensor along the feature depth channel to construct a prediction weight matrix; feeding the prediction weight matrix into a preset multilayer perceptron feature upscaling network to perform nonlinear mapping upscaling calculation to extract the decoupling weight matrix; applying channel extraction and segmentation operations to the decoupling weight matrix to separate the temperature control weight channel matrix and the lighting weight channel matrix; calculating the Hadamard product of the temperature control weight channel matrix and the node load sequence, and performing global average pooling calculation to obtain the temperature control energy consumption value; calculating the Hadamard product of the lighting weight channel matrix and the node load sequence, and performing global average pooling calculation to obtain the lighting energy consumption value; and integrating the data structures of the temperature control energy consumption value and the lighting energy consumption value through serialization combination operations to generate energy consumption decoupling evaluation data.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By acquiring the environmental thermal field sequence and nodal load sequence, and performing cross-correlation and sliding window translation calculations on these sequences, the period-aligned features are output. This step locks in the true phase lag step between environmental temperature and power consumption response in the input data due to building thermal inertia. Furthermore, sliding window translation compensation is applied to the environmental thermal field during data preprocessing, eliminating feature extraction bias caused by temporal misalignment. This calculation process achieves a high degree of period alignment in the spatiotemporal data input to the energy efficiency management analysis engine, effectively reducing the difficulty of fitting and optimizing non-stationary lag sequences in neural networks, and providing a high-quality data foundation for subsequently capturing various energy consumption characteristics.

[0015] 2. After constructing a spatial adjacency matrix based on the business flow graph and performing spatiotemporal aggregation, the thermoelectric feature tensor is input into a dual-branch feature decoupling network. The baseline expectation matrix is ​​output through the baseline fitting branch, and the bias matrix and mean matrix are output through the residual peaking branch. The net residual matrix is ​​established through broadcast subtraction. The business flow graph is introduced to capture the linkage energy consumption pattern. The dual-branch structure is used to independently extract the stable lighting baseline features and the drastically fluctuating temperature control net residual features. Deep adaptation is performed according to the physical heterogeneity of different energy consumption sub-items, realizing the orthogonal decomposition of multiple energy consumption indicators in the feature space, effectively preventing mutual interference of different dimensions of energy consumption patterns in the feature extraction process.

[0016] 3. The supply and return water temperature sequences are truncated from the preset water temperature threshold, and a thermal boundary mask matrix is ​​output. When generating the boundary constraint tensor, this thermal boundary mask matrix is ​​used for filtering, and then the baseline expectation matrix is ​​spliced ​​to construct the decoupling weight matrix, generating energy consumption decoupling assessment data. This calculation process directly embeds key thermodynamic boundaries as hard filtering conditions into the underlying attention scaling mechanism. When the HVAC system is in a non-operating state, the mask matrix forcibly masks and filters illegal boundary association positions, cutting off unreasonable weight allocations. This effectively avoids the model outputting predictive values ​​that violate common sense under the physical boundaries where the HVAC system is not actually operating, ensuring that the decoupling calculation strictly follows the laws of building physics. Ultimately, this contributes to generating energy consumption decoupling assessment data with extremely high accuracy and physical consistency.

[0017] 4. By organically integrating time-domain lag compensation, spatial-domain business association, feature-dimensional bi-branch decomposition, and underlying thermodynamic boundary constraints, a closed-loop deep learning architecture inspired by physical mechanisms is formed. Periodic alignment features and spatial adjacency matrices lay an accurate spatiotemporal foundation for the model. The bi-branch network then independently extracts energy consumption features, while the thermal boundary mask serves as a crucial defense to ensure physical correctness. These interconnected and mutually supportive steps enable the data-driven network model to internalize the real operating patterns of building energy consumption systems, achieving high synergy of multi-dimensional data during feature fusion and comprehensively enhancing the model's generalization ability in complex energy environments and extreme operating conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of a deep learning-based decoupling method for multiple building energy consumption indicators proposed in an embodiment of this invention application; Figure 2 This is a flowchart of the spatiotemporal aggregation process of the energy efficiency management analysis engine proposed in an embodiment of this invention. Figure 3 This is a flowchart of the dual-branch feature decoupling and boundary mask constraint proposed in an embodiment of this invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1-3 The present invention provides a deep learning-based method for decoupling multiple indicators of building energy consumption, the specific steps of which are as follows: Acquire environmental thermal field sequence, supply and return water temperature sequence, node load sequence and business flow map; perform sliding window translation calculation on environmental thermal field sequence and node load sequence, and output cycle alignment feature; construct spatial adjacency matrix based on business flow map; The spatial adjacency matrix and periodic alignment features are input into the energy efficiency management analysis engine for spatiotemporal aggregation, and the thermoelectric feature tensor is output. The thermoelectric feature tensor is input into a two-branch feature decoupling network. The baseline expectation matrix is ​​output through the baseline fitting branch, the bias matrix and mean matrix are output through the residual peaking branch, and the net residual matrix is ​​established through broadcast subtraction. The supply and return water temperature sequences are truncated with a preset water temperature threshold, and a thermal boundary mask matrix is ​​output. The thermoelectric feature tensor and the net residual matrix are linearly projected to output the key tensor, value tensor and query tensor. The thermal boundary mask matrix is ​​used for filtering, and dot product scaling and dot product calculation are performed to output the boundary constraint tensor. The baseline expectation matrix and boundary constraint tensor are spliced ​​together to construct the decoupling weight matrix; the decoupling weight matrix and the node load sequence are multiplied by Hadamard to output the temperature control energy consumption value and the lighting energy consumption value, thus generating energy consumption decoupling assessment data.

[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1 This application discloses a deep learning-based method for decoupling multiple building energy consumption indicators; see reference [link to relevant documentation]. Figure 1The specific steps proposed in this invention include: S1, acquiring the environmental thermal field sequence, supply and return water temperature sequence, node load sequence, and business flow map; performing sliding window translation calculation on the environmental thermal field sequence and node load sequence to output period alignment features; constructing a spatial adjacency matrix based on the business flow map; S2, inputting the spatial adjacency matrix and period alignment features into the energy efficiency management analysis engine for spatiotemporal aggregation to output a thermoelectric feature tensor; inputting the thermoelectric feature tensor into a dual-branch feature decoupling network, outputting the baseline expectation matrix through the baseline fitting branch, and outputting the bias matrix and mean moment through the residual peaking branch for dimensionality reduction. The system first performs a linear projection on the thermoelectric feature tensor and the net residual matrix, and then establishes the net residual matrix through broadcast subtraction. Next, it truncates the supply and return water temperature sequences with a preset water temperature threshold, outputting a thermal boundary mask matrix. It then performs linear projection on the thermoelectric feature tensor and the net residual matrix, outputting a key tensor, a value tensor, and a query tensor. These are filtered using the thermal boundary mask matrix, and dot product scaling and dot product calculations are performed to output a boundary constraint tensor. Finally, it splices the baseline expectation matrix and the boundary constraint tensor to construct a decoupling weight matrix. The system then performs a Hadamard product on the decoupling weight matrix and the node load sequence, outputting temperature control energy consumption values ​​and lighting energy consumption values, generating energy consumption decoupling assessment data.

[0023] Further, the environmental thermal field sequence, supply and return water temperature sequence, node load sequence, and service flow map are obtained; a sliding window translation calculation is performed on the environmental thermal field sequence and node load sequence to output the period alignment feature; a spatial adjacency matrix is ​​constructed based on the service flow map; corresponding to step S1 above; the specific implementation process includes: A Fast Fourier Transform (FFT) is applied to the environmental thermal field sequence and the node load sequence to extract the periodic thermal field evaluation quantity and the periodic electrical work evaluation quantity. The cross-correlation evaluation product of the periodic thermal field evaluation quantity and the periodic electrical work evaluation quantity is calculated to obtain the cross-power spectral density tensor. The cross-power spectral density tensor is processed using the inverse FFT operation to generate a time-domain cross-correlation function sequence. The peak index of the time-domain cross-correlation function sequence corresponding to each node is extracted to establish the response delay management vector. The environmental thermal field sequence is translated according to the response delay management vector to obtain the hysteresis-compensated thermal field sequence. Constant-length sliding slice extraction calculations are performed on the hysteresis-compensated thermal field sequence and the node load sequence respectively to obtain the windowed thermal field tensor and the windowed electrical work tensor. The windowed thermal field tensor and the windowed electrical work tensor are concatenated in the channel dimension to construct a periodic alignment feature.

[0024] Specifically, the generation process of the periodic alignment feature is as follows: Obtain environmental thermal field sequence, supply and return water temperature sequence, node load sequence, and business flow map: The environmental thermal field sequence was collected by an array of edge IoT sensors deployed on various floors and in independent spaces throughout the building. The data includes not only dry-bulb temperature but also estimated values ​​of relative humidity and infrared thermal radiation. Sampling was performed at discrete time steps, the selection of which was based on a comprehensive assessment of the sensor hardware's data acquisition and transmission performance, as well as the database's write capacity; in this embodiment, it was five minutes. The data dimensions of the entire sequence include the total time step, the total number of spatial nodes, and the number of thermal field feature channels.

[0025] Supply and return water temperature sequence: Read in real time from immersion temperature sensors on the manifold and distributor pipes in the building's HVAC room. This sequence consists of two dimensions: supply water and return water.

[0026] Node load sequence: Total active power readings collected by smart meters in distribution boxes of each area. This sequence is an undecoupled raw input source, and its data dimensions include time series length, total number of nodes, and power readings of individual indicators.

[0027] Business flow map: Structured map data extracted from the static spatial topology of the building information model and the dynamic flow trajectory of personnel recorded by access control and wireless network probes, used to reflect the load migration trend between spatial nodes.

[0028] In this embodiment, a nonlinear reconstruction operation of perceived temperature is introduced into the original environmental thermal field sequence. Multidimensional thermal comfort indices are extracted by combining micro-wind speed and radiation asymmetry, and sensor environmental noise is eliminated through multi-step mean filtering. Specifically, the dry-bulb temperature, relative humidity, and black-bulb temperature obtained from the infrared probe are extracted from the original environmental thermal field sequence. The difference between the black-bulb temperature and the dry-bulb temperature is calculated, multiplied by 2.4, and added to the dry-bulb temperature to obtain an estimated average radiation temperature. Indoor micro-wind speed monitoring values ​​are introduced. When the wind speed is no higher than 0.15 m / s, a static thermal reconstruction algorithm is activated: the dry-bulb temperature value in degrees Celsius is extracted, the product of the decimal form of the relative humidity (e.g., 0.6 for 60%) and 0.55 is subtracted, and the product of the average radiation temperature in degrees Celsius and 0.2 is added to calculate the baseline perceived temperature, where the coefficient 0.55 represents the temperature equivalent per unit change in humidity. When the wind speed exceeds 0.15 m / s, a dynamic thermal reconstruction algorithm is activated: Based on the baseline perceived temperature, the product of the square root of the wind speed (in meters per second) and 2.3 is subtracted to obtain the corrected perceived temperature, where 2.3 represents the equivalent cooling temperature per unit change in the square root of the wind speed. The continuously collected corrected perceived temperatures are then subjected to mean filtering calculations with a window size set to 5 time steps to filter out instantaneous fluctuation noise from infrared sensors caused by personnel movement, reconstructing the original multi-channel environmental thermal field sequence into a single channel. The dimension conversion coefficients for 0.55 and 2.3 mentioned above are based on the equivalent empirical model for thermal comfort of people with light activity in mid-latitude summers in the ASHRAE 55 standard, ensuring that the reconstructed indicators have a rigorous HVAC theoretical foundation. This process eliminates the discrepancy between physical temperature and the actual thermal perception of the human body, improving the accuracy of the subsequent energy consumption decoupling model in capturing temperature control load fluctuations caused by human regulatory behavior.

[0029] Fast Fourier Transform (FFT) was applied to both the environmental thermal field sequence and the nodal load sequence. Since the initial lengths of the environmental thermal field sequence and the nodal load sequence may differ during actual acquisition, zero-padding was performed on the tails of both sequences before the transform to ensure consistent length and expand them to an integer power of 2 equal to or greater than the current maximum length. The time-domain value of the zero-padding environmental thermal field sequence was multiplied by a complex exponential kernel function, and the result was discretely summed over all time steps of the entire time period to extract the periodic thermal field evaluation quantity. The base of the complex exponential kernel function is the natural constant, and the exponent is composed of the negative imaginary unit, twice pi, the frequency index, and the current time step number, multiplied sequentially and then divided by the total length of the zero-padding time sequence. Using the same calculation process, the time-domain value of the nodal load sequence was multiplied by the complex exponential kernel function, and the result was discretely summed along the time axis to extract the periodic electrical work evaluation quantity.

[0030] The conjugate complex form of the periodic power evaluation quantity is obtained, and element-wise multiplication is performed between it and the periodic thermal field evaluation quantity at the corresponding frequency component positions. The cross-correlation evaluation product of the two is calculated to obtain the cross-power spectral density tensor. The element value at each position in this tensor not only integrates the amplitude product of the two sequences at that frequency, but also implicitly encodes the phase difference information between them.

[0031] The cross-power spectral density tensor is processed using inverse Fast Fourier Transform (IFFT). The cross-power spectral density tensor is multiplied by an inverse complex exponential kernel function, and the result is integrated and summed over all frequency components in the frequency domain. The summation is then multiplied by the reciprocal of the total length of the time series after zero-padding to generate a time-domain cross-correlation function sequence. The exponential term of the inverse complex exponential kernel function has its negative sign removed to achieve inverse mapping. A global numerical search is performed on the time-domain cross-correlation function sequences corresponding to each node. Considering the possibility of multiple maxima in the cross-correlation function (i.e., multi-peak phenomenon), a priori constraint time window (e.g., limited to a physically reasonable delay range of 0 to 120 minutes) is introduced when extracting peak indices. Only within this constraint time window are the time step indices corresponding to the maximum absolute value of the function sequence retrieved and extracted as disambiguated peak indices. All peak indices are established as response delay management vectors.

[0032] Extract the peak index to establish a response delay management vector containing the sign; if the time delay step corresponding to the vector is positive, it indicates that the thermal field lags behind the electrical work, so the timestamp of each data point in the environmental thermal field sequence is uniformly subtracted by this time span (i.e., shifted backward); if the time delay step is negative, it indicates that the electrical work lags behind the thermal field, so the thermal field sequence is shifted forward on the time axis; if it is zero, no shift is performed, and the boundary gaps generated by the shift are filled by copying the nearest neighbor valid values ​​to obtain the hysteresis-compensated thermal field sequence.

[0033] Constant-length sliding slice extraction calculations are performed on the hysteresis-compensated thermal field sequence and the nodal load sequence, respectively. The size of the time sliding window is determined statistically based on the historical relaxation time data required for the target building envelope to reach thermodynamic equilibrium, for example, configured as 48 time steps; the sliding step size is set by grid search based on the memory capacity limitations of the computing hardware and the required amount of data augmentation samples. The one-dimensional time series data slices extracted within the sliding window are upscaled into independent time matrices, and the windowed thermal field tensor and windowed electrical work tensor are obtained respectively.

[0034] To eliminate the dimensional difference between the environmental thermal field and the electrical work sequence, Z-Score standardization is performed on the windowed thermal field tensor and the windowed electrical work tensor before channel splicing: the mean and standard deviation of the data within the sliding window of each node are calculated, the original value in the tensor is subtracted from the corresponding mean, and then divided by the corresponding standard deviation plus a minimum bias constant (e.g., 10). -5 To prevent division by zero, we obtain a standardized windowed thermal field tensor and a standardized windowed electrical power tensor that follow a zero-mean and unit-variance distribution. The data matrices of these two tensors are then directly combined in the last feature depth dimension, mapped to a multi-channel tensor, constructing a periodically aligned feature that balances time synchronization and multimodal characteristics.

[0035] By applying Fast Fourier Transform to the environmental thermal field and electrical power sequence, the cross power spectral density is calculated in the frequency domain and the true phase lag step is extracted, quantifying the signal delay caused by building thermal inertia. By using lag compensation, strict periodic alignment of the multidimensional sequence is ensured at the bottom layer, effectively reducing the network optimization deviation caused by timing misalignment.

[0036] Based on the business flow graph, the spatial business flow probability and connection frequency between building space entity nodes are analyzed and extracted; the product of the spatial business flow probability and the connection frequency is calculated as the node collaborative management weight, and an initial graph matrix is ​​constructed accordingly; the symmetric connection attribute of the initial graph matrix is ​​extracted, and a symmetric connection tensor is established; a preset constant is filled into the main diagonal elements of the symmetric connection tensor to construct a self-collaborative evaluation tensor; numerical summation and aggregation operations are performed on the self-collaborative evaluation tensor along the row direction to obtain a degree matrix tensor, which is then normalized to obtain a normalized degree matrix tensor; continuous matrix multiplication operations are performed between the normalized degree matrix tensor and the self-collaborative evaluation tensor to generate a spatial adjacency matrix.

[0037] Specifically, the process of generating the spatial adjacency matrix is ​​as follows: Based on the aforementioned business flow map, the spatial business flow probability and connection frequency between building space entity nodes are analyzed and extracted. The spatial business flow probability is calculated based on the historical displacement data of access control check-in logs and wireless network terminal probes within a specified time period. By statistically analyzing the personnel movement frequency from a specific starting node to a target node and dividing it by the total outflow frequency of that starting node, a relative proportion reflecting the personnel load migration trend is obtained. The specific conversion calculation rule for connection frequency is as follows: extract the thermal conductivity value of the physical isolation material, calculate the reciprocal of the thermal conductivity value as the basic weight of heat conduction; extract the initial heat conduction weight of all node pairs, and use the maximum value normalization algorithm (i.e., divide by the global maximum initial heat conduction weight) to convert it into a dimensionless heat conduction basic weight of 0 to 1. Determine whether there is a physical connection between the two nodes in the HVAC return air duct network. If there is a physical connection, obtain the historical highest and lowest operating extreme values ​​of the valve opening coefficient and fan frequency signal of this type in the HVAC building management system. Subtract the lowest operating extreme value from the currently collected valve opening coefficient or fan frequency signal, and divide by the difference between the highest and lowest operating extreme values. Calculate and map it to the dimensionless interval of zero to one using this linear interpolation, and use it as the dynamic connectivity identifier value. Otherwise, it is zero. The initial physical connectivity evaluation value of the corresponding physical heat transfer channel is calculated by summing the basic heat conduction weight and the dynamic connectivity identifier value equally (i.e., multiplying each by a constant of 0.5). To prevent causal reversal in feature updates, a final weighted design that separates static physical and dynamic business operations is adopted. The initial physical connectivity evaluation value of the above physical heat transfer channel is used as the static base weight and uniformly assigned a fixed weight coefficient of 0.7; the spatial business flow probability is used as the superposition correction term for dynamic thermal disturbances caused by personnel flow and uniformly assigned a fixed weight coefficient of 0.3; a weighted summation calculation is performed on the two to obtain the node collaborative management weight that balances physical structure and business dynamics. Based on this, the edge weights of all associated node pairs are mapped to the corresponding row and column index positions of the matrix, and the remaining unassociated positions are set to zero to construct the initial graph matrix.

[0038] The symmetric connectivity property of the initial graph matrix is ​​extracted to establish the symmetric connectivity tensor. Since the heat transfer and personnel return in the physical space have undirected graph characteristics, the constructed initial graph matrix is ​​added to its transpose. Then, all element values ​​in the resulting matrix are multiplied by a constant 1 / 2 to eliminate the weight skew caused by directed computation, thus establishing the symmetric connectivity tensor.

[0039] Considering the heat storage capacity and information self-retention mechanism of the physical space nodes, a preset constant is filled into the main diagonal elements of the symmetric connection tensor to construct a self-cooperative evaluation tensor. That is, by performing matrix addition on the symmetric connection tensor and the identity matrix of the same dimension, a value of 1 is uniformly added to the main diagonal to represent the node's own characteristic retention attribute.

[0040] In this embodiment, instead of using a preset constant, the spatial volume ratio of the building space entity and the thermal insulation coefficient of the enclosure structure are extracted. A comprehensive weight with a dynamic feature retention ratio is constructed by normalizing and adding these values, and this non-uniform value is used to fill the main diagonal. Specifically, the product of the actual length, width, and height of the building space corresponding to each node in the business flow graph is extracted to obtain the three-dimensional volume value of the independent node. The volume value of a single node is divided by the sum of the volumes of all nodes involved in the calculation to obtain the first proportional coefficient. The thickness value of the outer wall around the corresponding node is extracted simultaneously and multiplied by the reciprocal of the standard thermal conductivity of the building material to obtain the thermal insulation coefficient of the node. The maximum value of the thermal insulation coefficient among all nodes is found, and the thermal insulation coefficient of the current node is divided by this maximum value to obtain the second proportional coefficient. The first proportional coefficient and the second proportional coefficient are added together and multiplied by a constant of 0.5 for averaging to obtain the comprehensive feature retention weight. After establishing the symmetric connection tensor, instead of using a fixed constant to fill the main diagonal, the calculated comprehensive feature retention weight of each node is used to fill the corresponding main diagonal element positions of the tensor. The above process enables the spatial adjacency matrix to integrate physical spatial scale and thermal isolation properties, preventing nodes with small volume or poor insulation from losing their local unique energy consumption patterns during feature aggregation.

[0041] The self-cooperative evaluation tensor is summed along the row direction to obtain a degree matrix tensor, which is then normalized to obtain a normalized degree matrix tensor. Specifically, the sum of the elements in each row of the self-cooperative evaluation tensor is filled into the corresponding diagonal positions, and the off-diagonal elements are uniformly set to zero, thus constructing a diagonalized degree matrix tensor. To smooth out gradient explosion or vanishing phenomena caused by differences in node connectivity degrees, the degree matrix tensor is normalized: each non-zero value on its main diagonal is raised to the power of -1 / 2; during this process, to avoid division by zero errors caused by isolated nodes, a very small bias value (such as 10) is applied to the main diagonal elements to prevent numerical overflow. -5 ), to obtain a smooth normalized degree matrix tensor.

[0042] Following the framework of graph convolution theory, the normalized degree matrix tensor, the self-co-evaluation tensor, and the normalized degree matrix tensor are multiplied sequentially to generate the final spatial adjacency matrix.

[0043] In the real-time inference deployment phase of the energy efficiency management analysis engine, a preset graph time update cycle is set. This cycle is selected based on a comprehensive calculation of the target building's personnel flow frequency characteristics; for example, for a typical office building with frequent personnel flow, this cycle can be set to 15 minutes. Every time this graph time update cycle occurs, the latest displacement data from access control and wireless probes is incrementally acquired, the flow ratio and connectivity frequency between nodes are recalculated, and the aforementioned calculation process is executed to generate the spatial adjacency matrix for the current cycle. Subsequently, the average absolute value deviation of the differences between corresponding elements in the old and new matrices is calculated. If this average deviation exceeds a preset topology mutation threshold, the current spatial adjacency matrix is ​​overwritten into the engine's memory, achieving dynamic adaptive evolution of the feature topology. The topology mutation threshold is selected based on the total number of spatial nodes in the building and the basic random noise distribution probability of the access control system; for example, it is set to 0.05 to provide a 5% natural feature fluctuation margin, thereby avoiding frequent invalid global matrix update calculations triggered by extremely small-scale random personnel movement.

[0044] Based on the analysis of spatial business flow probability and connection frequency in the business flow graph to construct a spatial adjacency matrix, this method changes the model that relies solely on physical distance for graph construction. It transforms personnel activity trajectories and business flow logic into network topology weights, enabling the model to accurately extract energy consumption spatial features with real business linkage characteristics.

[0045] Furthermore, the spatial adjacency matrix and periodic alignment features are input into the energy efficiency management analysis engine for spatiotemporal aggregation, outputting a thermoelectric feature tensor. This thermoelectric feature tensor is then input into a dual-branch feature decoupling network. The baseline expectation matrix is ​​output through the baseline fitting branch, and the bias matrix and mean matrix are output through residual peaking branch for dimensionality reduction. Finally, the net residual matrix is ​​established through broadcast subtraction; this corresponds to step S2 above. (See also...) Figure 2 and Figure 3 The specific implementation process includes: The energy efficiency management analysis engine includes a spatial energy efficiency map supervision layer and a time-series energy consumption trend prediction layer. In the spatial energy efficiency map supervision layer, a linear mapping multiplication calculation is performed on the periodic alignment features using a preset spatial management evaluation weight tensor to obtain a dimensionality-reduced node feature matrix. The spatial adjacency matrix and the dimensionality-reduced node feature matrix are aggregated using standard matrix multiplication to extract a primary aggregated feature matrix. The primary aggregated feature matrix is ​​processed by nonlinear mapping calculation with leakage correction linear units to generate a primary spatial topology tensor. The primary spatial topology tensor is fed into a preset multi-head attention allocation mechanism to complete node distance allocation and obtain a weighted spatial feature tensor. Layer normalization numerical constraint calculation and residual addition operation are performed on the weighted spatial feature tensor in sequence to establish the spatial topology tensor. The spatial topology tensor is forward-passed to the time-series energy consumption trend prediction layer as a single-step topology input tensor.

[0046] Specifically, the processing procedure for the space energy efficiency map monitoring layer is as follows: In the spatial energy efficiency map supervision layer, along the feature channel dimension, a pre-defined spatial management evaluation weight tensor is used to perform linear mapping multiplication on the input periodically aligned features (specifically, feature slices at each time step). This compresses the high-dimensional multimodal channels into the hidden layer representation space, reducing the feature dimensionality and extracting key information to obtain the dimensionality-reduced node feature matrix. The dimension of this spatial management evaluation weight tensor is dynamically allocated based on the number of forward input feature channels and the set low-dimensional hidden layer feature dimension. Its initial value is obtained by random sampling within a specified interval (e.g., between -0.1 and 0.1) using a uniform distribution.

[0047] Subsequently, within a complete sliding window period, the spatial adjacency matrix of the latest overwrite within the window is always used as the benchmark. The benchmark spatial adjacency matrix and the reduced-dimensional node feature matrix are subjected to standard matrix multiplication aggregation calculation to ensure that the feature aggregation at each time step within the window has a unified spatial topological reference, realize the weighted information interaction of features between adjacent nodes, and extract the primary aggregated feature matrix.

[0048] The primary aggregate feature matrix is ​​processed by nonlinear mapping calculation using a linear unit with leakage correction: when the matrix element value is greater than 0, the original value is retained; when the element value is less than 0, it is multiplied by a preset leakage slope parameter (set to 0.01 in this embodiment). This can effectively alleviate the problem of neuron death during training, thereby generating a nonlinear primary spatial topology tensor.

[0049] To further optimize feature allocation for long-distance spatial nodes, the aforementioned primary spatial topology tensor is fed into a pre-defined multi-head attention allocation mechanism. Feature weights are dynamically adjusted by calculating attention scores between nodes to obtain a weighted spatial feature tensor. Layer-normalized numerical constraints are then applied sequentially to the weighted spatial feature tensor to stabilize the internal data distribution of each layer. Residual addition is then performed, directly summing the input and output values ​​of the attention layer bit-by-bit to prevent degradation in deep graph networks. After these processes, the final spatial topology tensor is established and forward-propagated to the time-series energy consumption trend prediction layer. Within the time dimension corresponding to its sliding window, each time-step slice is extracted sequentially as the single-step topology input tensor.

[0050] In graph convolutional networks, node features are aggregated and combined with a multi-head attention mechanism to dynamically allocate node distances, realizing the computational fusion of spatial topological features and attention weights. While preserving physical connectivity attributes, the feature response intensity of a specific space is adaptively adjusted, enhancing the generalization of spatial feature extraction.

[0051] Extract the hidden state tensor and core state evaluation tensor from the previous time step; concatenate the single-step topological input tensor and the hidden state tensor along the channel dimension to construct a joint spatiotemporal tensor; perform a nonlinear mapping on the joint spatiotemporal tensor using a preset set of gating weight matrices to extract the redundancy removal evaluation tensor, trend introduction evaluation tensor, and output gating tensor respectively; perform state update calculation by combining the trend introduction evaluation tensor, the redundancy removal evaluation tensor, and the core state evaluation tensor to obtain the current step core state evaluation tensor; calculate the activation mapping of the current step core state evaluation tensor and the Hadamard product of the output gating tensor to finally generate the thermoelectric feature tensor.

[0052] Specifically, the processing steps of the time-series energy consumption trend prediction layer are as follows: In the time-series energy consumption trend prediction layer, temporal features are extracted and state updates are performed on the single-step topological input tensor. First, the hidden state tensor and core state evaluation tensor of the previous time step are extracted from memory. Along the feature channel dimension, the current single-step topological input tensor and the hidden state tensor of the previous time step are directly concatenated to construct a joint spatiotemporal tensor that integrates the current topological space and historical temporal information.

[0053] A pre-defined set of gating weight matrices is used to perform nonlinear mapping control on the joint spatiotemporal tensor. Specifically, the joint spatiotemporal tensor is multiplied by the forgetting gating weight matrix, the input gating weight matrix, and the output gating weight matrix, and then their respective bias vectors are added. The results of these three linear operations are then uniformly input into a sigmoid activation function to smoothly map the values ​​to between 0 and 1, extracting the redundancy removal evaluation tensor, the trend introduction evaluation tensor, and the output gating tensor. The initial values ​​of each gating weight matrix and bias vector are set based on random probability sampling from a standard normal distribution.

[0054] The state update calculation is performed by combining the trend-introduced evaluation tensor, the redundancy removal evaluation tensor, and the core state evaluation tensor of the previous time step. This process includes: calculating the Hadamard product of the redundancy removal evaluation tensor and the core state evaluation tensor of the previous time step (element-by-element product acquisition) to determine the retention ratio of historical information; simultaneously, performing a nonlinear mapping on the joint spatiotemporal tensor using a preset candidate state weight matrix, and inputting the mapping result into a hyperbolic tangent activation function to obtain a candidate core state evaluation tensor; calculating the Hadamard product of the candidate core state evaluation tensor and the trend-introduced evaluation tensor to determine the writing ratio of new topology information; and performing element-by-element addition on the above two product results to obtain the core state evaluation tensor of the current step.

[0055] Finally, the updated current-step core state evaluation tensor is processed by hyperbolic tangent activation mapping, and the Hadamard product of the activation mapping result and the output gating tensor is calculated to generate and output a thermoelectric feature tensor containing comprehensive spatiotemporal hidden layer information. In this module, the hidden layer feature dimension of the Long Short-Term Memory network is comprehensively calculated and configured in combination with the upper limit of the memory resources of the edge computing host and the historical fluctuation complexity index of the target electricity load sequence.

[0056] By extracting the hidden and core state evaluation tensors of historical steps and constructing a joint spatiotemporal tensor and using a gating weight matrix to perform state updates, the retention and forgetting ratio of historical features during network transmission is controlled, enabling the thermoelectric feature tensor to have excellent long-term tracking capabilities in the time dimension and extraction of temporal dependencies.

[0057] The thermoelectric feature tensor is processed by multilayer perceptron dimensionality reduction fusion computation to extract the shared hidden layer feature tensor; the shared hidden layer feature tensor is guided into the baseline fitting branch, and the baseline expectation matrix is ​​generated by matrix multiplication operation between it and the preset baseline fitting fully connected weight tensor; the shared hidden layer feature tensor is passed through the residual peaking branch, and the bias matrix is ​​obtained by matrix multiplication operation with the preset peaking fully connected weight tensor; the bias matrix is ​​calculated by mean summation along the feature channel dimension to obtain the residual mean scalar matrix; the residual mean scalar matrix is ​​expanded to the scale of the bias matrix in the channel dimension to construct the mean matrix; the element-wise subtraction cancellation result of the bias matrix and the mean matrix is ​​calculated to establish the net residual matrix.

[0058] Specifically, the process of establishing the net residual matrix is ​​as follows: The thermoelectric feature tensor is processed by dimensionality reduction and fusion computation using a multilayer perceptron to extract a shared hidden layer feature tensor after filtering out redundant noise. The number of network layers and nodes in each layer of the multilayer perceptron are dynamically determined using a Bayesian hyperparameter optimization search algorithm, combined with the measurement error results of the objective function on the validation set, to ensure the expressive power of the feature nonlinear transformation.

[0059] The shared hidden layer feature tensor is guided into the baseline fitting branch, and standard matrix multiplication is performed using it and the preset baseline fitting fully connected weight tensor to generate a baseline expectation matrix that characterizes the building’s normal basic energy consumption (such as conventional lighting and basic socket load).

[0060] Simultaneously, the shared hidden layer feature tensor is synchronously passed through the residual peaking branch, and the bias matrix is ​​obtained by combining it with the preset peaking fully connected weight tensor matrix multiplication operation.

[0061] To obtain the dynamic sensitive component that eliminates steady-state offset, the bias matrix is ​​summed and calculated along the feature channel dimension. Specifically, the values ​​of each independent spatial node and time step position in the bias matrix are summed across all feature channels, and the sum is divided by the total number of feature channels, thereby extracting the residual mean scalar matrix reduced to a single channel dimension.

[0062] The residual mean scalar matrix is ​​expanded along the channel dimension to the scale of the bias matrix to construct the mean matrix. In specific data structure operations, the tensor broadcasting mechanism of the deep learning framework is used to copy the residual mean scalar matrix of a single channel exactly along the direction of the original feature depth, restoring its size and structure to be completely consistent with the original bias matrix, thus completing the construction of the mean matrix.

[0063] The element-wise subtraction cancellation result between the bias matrix and the mean matrix is ​​calculated to extract the dynamic fluctuation characteristics after removing the static mean, thus establishing the final net residual matrix. This net residual matrix highlights anomalous abrupt changes in the time series.

[0064] A dual-branch network is used to generate the baseline expectation matrix and bias matrix, respectively, and the net residual matrix is ​​established by mean broadcast subtraction. This calculation achieves independent calculation and separation of basic stable load and sudden fluctuating load in the hidden layer feature space, avoiding feature interleaving and providing a clean input for accurate mapping of subsequent energy consumption weights.

[0065] Further, the supply and return water temperature sequences are truncated with a preset water temperature threshold, and a thermal boundary mask matrix is ​​output. Linear projections are performed on the thermoelectric feature tensor and the net residual matrix, respectively, outputting a key tensor, a value tensor, and a query tensor. These are then filtered using the thermal boundary mask matrix, and dot product scaling and dot product calculations are performed to output a boundary constraint tensor; this corresponds to step S3 above. The specific implementation process includes: A preset set of projection weight matrices is used to linearly map the thermoelectric feature tensor and the net residual matrix, respectively, to extract the key tensor, the value tensor, and the query tensor. The inner product multiplication result of the query tensor and the key tensor is calculated to establish an unscaled attention score matrix. The unscaled attention score matrix is ​​then subjected to additive masking using the thermal boundary mask matrix to filter illegal boundary association positions, resulting in a masked score matrix. Dot product scaling and exponential normalization are performed on the masked score matrix to construct an attention weight normalization matrix. Finally, a boundary constraint tensor is generated by superimposing matrix multiplication between the attention weight normalization matrix and the value tensor.

[0066] Specifically, the generation process of the boundary constraint tensor is as follows: Based on the factory performance calibration parameters of the HVAC units actually installed in the target building, the preset water temperature threshold is selected and determined. For conventional central air conditioning water systems, the physical extreme value range of chilled water supply and return water is set to 7℃ to 12℃; when the large temperature difference ice storage mode is activated, this extreme value range is expanded to 5℃ to 15℃; for the cooling water circulation loop, the lower limit temperature threshold of the chiller unit cooling water inlet is forcibly set to 15.5℃, and the absolute upper limit temperature threshold is set to 33℃.

[0067] The real-time supply and return water temperature sequence is logically compared with the preset water temperature threshold: if the supply and return water temperatures exceed the allowable physical extreme range, it indicates that the HVAC system is not performing effective heat exchange, and a very large negative value (such as -1×10) is output to the mask position corresponding to that time step. 9 If the water temperature is within the valid range, a zero value is output. A thermal boundary mask matrix is ​​constructed based on the time step and spatial node index. Furthermore, to avoid the mask failure problem caused by the main pipe water temperature being within the valid range but the terminal equipment in a specific area being actively shut down, the terminal fan start / stop flag matrix of each spatial node is collected simultaneously from the edge IoT network (0 for on, and a very large negative number for off). The thermal boundary mask matrix constructed based on the main unit water temperature is added element-wise to the terminal fan start / stop flag matrix. Only under the dual conditions of the main pipe main unit running and the local spatial terminal fan being on, will the mask at the corresponding position be zero, thus accurately adapting to the complex working conditions of buildings with multiple independent zones of cold and heat sources.

[0068] A preset set of projection weight matrices is used to linearly map and extract the forward-transmitted thermoelectric feature tensor and the net residual matrix. Specifically, the extracted net residual matrix is ​​multiplied with the query projection weight matrix to extract the query tensor used to initiate addressing in the feature space; simultaneously, the thermoelectric feature tensor containing complete spatiotemporal correlation information is multiplied with the key projection weight matrix and the value projection weight matrix to extract the key tensor and value tensor representing the target feature library. The specific values ​​of each projection weight matrix are determined by the iterative parameter updates during the network's pre-training and fine-tuning processes. To achieve strict dimensional alignment with the 16-dimensional baseline expectation matrix output by the dual-branch network front end, the feature projection dimensions of the key tensor, value tensor, and query tensor are all uniformly and forcibly configured to 16 dimensions.

[0069] A standard matrix multiplication operation is performed between the query tensor and the transposed key tensor to quantify the spatiotemporal correlation between features at different time steps and spatial nodes. Specifically, the query tensor is configured with dimensions of [batch size, number of spatial nodes, projection dimension], and the key tensor is configured with dimensions of [batch size, number of time steps, projection dimension]. The inner product of the two tensors establishes an unscaled attention score matrix with dimensions of [batch size, number of spatial nodes, time steps]. In terms of physical semantics, this cross-dimensional attention mechanism enables the net residual features of each spatial node to autonomously address and match time step features highly correlated with their fluctuation patterns on the time axis. Simultaneously, since the supply and return water temperatures of the building's heating and cooling sources (such as chillers) determine the global cooling or heating capacity limit of the entire building system, the one-dimensional time step thermal boundary mask is directly broadcast and expanded along the spatial node dimensions, allowing all spatial nodes to share this global physical legitimacy state at the same time step.

[0070] Using an initial one-dimensional time mask sequence generated according to time step indices, the tensor broadcasting mechanism of deep learning is used to replicate and expand it along the query dimension. This ensures that each time step inherits the physical validity marker when used as a query, constructing a hot boundary mask matrix with the same scale as the unscaled attention score matrix. At the underlying tensor operation level, element-wise addition is performed between the hot boundary mask matrix and the unscaled attention score matrix, adding extremely negative numbers to the scores associated with illegal boundaries to forcibly filter those positions. To avoid gradient sparsity and training instability caused by large-scale absolute masking, a progressive additive masking strategy is adopted during the backpropagation training phase of the model parameters: in the first 10 training rounds, only a soft penalty is applied (i.e., adding a small negative number, such as -5, to the attention score at illegal positions so that the gradient can still be weakly backpropagated), and in subsequent rounds, the absolute value of the negative penalty is linearly increased according to the training progress until it reaches an extremely negative number (such as -10 to the power of 9); while in the real-time inference deployment phase, a hard mask with an extremely negative number is directly applied. After subsequent exponential normalization, the attention weight of illegal positions with superimposed extremely negative numbers approaches zero, effectively severing unreasonable boundary associations and yielding a masking score matrix.

[0071] The values ​​of each element in the masking score matrix are scaled by dividing by the square root of the key tensor feature dimension to prevent excessively large inner product values ​​from causing subsequent calculations to fall into the activation function saturation region and gradient vanishing. Then, the scaled numerical matrix is ​​input into the exponential softmax function for normalization. Under exponential operation, the probability distribution weights of illegal positions masked by extremely large negative numbers will rapidly decay and approach zero, thus constructing an attention weight normalization matrix focused on effective cooling / heating periods.

[0072] By using the matrix multiplication superposition between the attention weight normalization matrix and the value tensor, the weighted aggregation of features is completed, generating a boundary constraint tensor that can lock the true temperature control power consumption features.

[0073] The query, key, and value tensors are extracted for inner product calculation. The unscaled fractions are then additively masked using a thermal boundary mask matrix. This transforms the system's thermodynamic state into a hard constraint and severs illegal boundary associations, eliminating erroneous energy consumption characteristic allocation within the non-operational physical range of the equipment and strictly ensuring the physical rigor of the calculation output.

[0074] Further, the baseline expectation matrix and the boundary constraint tensor are concatenated to construct a decoupling weight matrix; the decoupling weight matrix and the node load sequence are multiplied by a Hadamard product to output the temperature control energy consumption value and the lighting energy consumption value, generating energy consumption decoupling evaluation data; corresponding to step S4 above; the specific implementation process includes: The baseline expectation matrix and the boundary constraint tensor are concatenated along the feature depth channels to construct a prediction weight matrix. The prediction weight matrix is ​​fed into a preset multilayer perceptron feature upscaling network to perform nonlinear mapping upscaling calculations and extract the decoupled weight matrix. Channel extraction and segmentation operations are applied to the decoupled weight matrix to separate the temperature control weight channel matrix and the lighting weight channel matrix. The Hadamard product of the temperature control weight channel matrix and the node load sequence is calculated, and global average pooling is performed to obtain the temperature control energy consumption value. The Hadamard product of the lighting weight channel matrix and the node load sequence is calculated, and global average pooling is performed to obtain the lighting energy consumption value. The data structures of the temperature control energy consumption value and the lighting energy consumption value are integrated through a serialization combination operation to generate energy consumption decoupled evaluation data.

[0075] Specifically, the process for generating energy consumption decoupling assessment data is as follows: The decoupled hidden layer features are restored to specific physical energy consumption values. Along the channel axis of the feature depth, the extracted baseline expectation matrix (representing steady-state base load) and the boundary constraint tensor (representing dynamic temperature control load) constrained by thermal boundary mask are directly concatenated end to end to construct the prediction weight matrix, thereby aggregating the feature representations of all decoupled branches.

[0076] The predicted weight matrix is ​​fed into a preset multilayer perceptron feature upscaling network to perform nonlinear mapping upscaling calculation. At the end of the network output, exponential soft maximization normalization calculation is applied along the feature channel dimension to extract the decoupled weight matrix, so that the sum of the output values ​​of each channel is strictly equal to 1. In this process, although the feature upscaling network performs mixed mapping on the spliced ​​features, by subsequently introducing the physical sub-item meter calibration set to calculate the end-to-end supervised loss, the network is forced to be given a clear semantic optimization direction, thereby ensuring that the extracted channel 0 is always stably aligned with the temperature control prediction in physical semantics, and channel 1 is always stably aligned with the basic predictions such as lighting, preventing the interleaving and interchange of channel semantics.

[0077] Subsequently, a channel extraction and segmentation operation is applied to the decoupling weight matrix to separate the temperature control weight channel matrix and the lighting weight channel matrix. At this point, the element values ​​in each weight channel matrix represent the legitimate dimensionless proportion distribution of the corresponding energy-consuming equipment in the total active power. For stable basic energy consumption, including the third type of non-temperature-controlled loads, it is broadly classified into the lighting weight channel matrix in physical semantics.

[0078] The Hadamard product (i.e., element-wise product of corresponding elements) is calculated between the temperature control weighted channel matrix and the original input node load sequence to map the dimensionless weight proportions back to the actual physical power dimensions. Subsequently, global average pooling is performed, which involves summing the product results along the spatial dimension and dividing by the total number of spatial nodes. This eliminates measurement noise from individual nodes, yielding a more robust regional average temperature control energy consumption value. It should be noted that global average pooling is primarily suitable for homogeneous building scenarios where the power levels of spatial nodes are similar. For heterogeneous building scenarios with a mix of high- and low-power nodes, a non-uniform weighted pooling scheme is preferred.

[0079] In this embodiment, average pooling is not used. Instead, the historical peak load of each node is extracted as the energy benchmark. A dynamic energy density weight is generated by combining this with a real-time load volatility trigger threshold. The Hadamard product result is then subjected to non-uniform weighted summation and normalization. The maximum active power reading of the node load sequence over the past 30 days is extracted and set as the absolute value of the energy benchmark. The absolute difference between the instantaneous load value of each node at the current time step and the load value at the previous time step is calculated. This absolute difference is divided by the absolute value of the energy benchmark to obtain the real-time load volatility. It is determined whether the real-time load volatility exceeds a preset change threshold of 0.15. If it does not exceed this threshold, the node is assigned a base weight of constant 1. If it exceeds this threshold, the real-time load volatility is multiplied by a constant 5 and then added to a constant 1 to obtain an enhanced weighted value as the energy density weight of the node. After calculating the Hadamard product of the temperature control weighted channel matrix and the node load sequence, instead of directly performing pure arithmetic summation along the spatial dimension, each node element in the Hadamard product result is multiplied by its corresponding calculated energy density weight, and a weighted summation operation is performed. Finally, the weighted summation result is divided by the sum of all node energy density weights to eliminate the influence of uneven spatial distribution and extract the temperature control energy consumption value. This process avoids high-energy-consuming abrupt nodes being masked by global mean smoothing, and enhances the sensitivity and accuracy of the decoupling results for high-intensity start-up and shutdown characteristics of local equipment.

[0080] Using symmetric data restoration logic, the Hadamard product of the lighting weight channel matrix and the node load sequence is calculated to restore the physical dimensions of power; similarly, the summation of the product results along the spatial dimension is divided by the corresponding number of nodes to complete the global average pooling calculation and obtain the regional average lighting energy consumption value.

[0081] Through serialization and combination operations, the physical identifiers of each node in the space, the collection timestamps, the extracted temperature control energy consumption values, and the lighting energy consumption values ​​are integrated and packaged into a standardized data structure (such as standardized JSON or data table format) to generate the final energy consumption decoupling evaluation data.

[0082] By extracting the decoupled weight matrix in higher dimensions and applying channel extraction and segmentation operations, and combining the node sequence to perform Hadamard product and global average pooling respectively, the numerical convergence of the weights of each independent channel and the calibration of the entity energy consumption value were completed, realizing the mapping transformation from high-dimensional abstract features to specific physical quantities.

[0083] In this embodiment, the spatial energy efficiency map supervision layer (i.e., graph convolutional network layer) of the energy efficiency management analysis engine is established as an asymmetric three-layer progressive structure. The number of these three aggregation layers is determined based on the comprehensive effective topological receptive field of the physical and thermal conduction span within the target building and the cross-regional flow of personnel. Exceeding three layers will lead to over-smoothing of spatial topological features, resulting in the loss of spatiotemporal energy heterogeneity between nodes. Specifically, the first layer is the initial mapping layer, which maps the spliced ​​and fused periodically aligned feature dimensions to a 128-dimensional hidden space; the second layer is the deep aggregation layer, which maintains the 128-dimensional node features for spatiotemporal information interaction calculation; and the third layer is the topology compression layer, which reduces the dimensionality of the aggregated features to 64 dimensions to output a spatial topological tensor. After each GCN calculation, a nonlinear mapping with leakage correction linear unit and layer normalization operation with a slope parameter set to 0.01 are performed sequentially.

[0084] The time-series energy consumption trend prediction layer (i.e., the long short-term memory network layer) is established as a two-layer stacked architecture: the bottom LSTM network receives the 64-dimensional single-step topology input tensor, and its hidden state feature channel number is set to 128-dimensional, which is used to expand the receptive field and capture shallow short-term fluctuation patterns. This dimension scale is calculated based on the number of independent thermodynamic zones of the target building and the feature space complexity of the multi-dimensional start-stop mode of electrical equipment, thus distinguishing it from the general IoT time-series model; the top LSTM network receives the 128-dimensional time-series features from the bottom layer, and its hidden state feature channel number is compressed back to 64-dimensional, extracting high-order comprehensive information and outputting the final 64-dimensional thermoelectric feature tensor.

[0085] The hyperparameters of the multilayer perceptron used to process thermoelectric feature tensors were optimized in advance using a Bayesian optimization search algorithm combined with validation set error. The specific search configuration was as follows: the number of network layers was set to two to four, and the number of nodes in each layer was discretely selected from 32, 64, and 128; the expected improvement was used as the acquisition function, the initial number of random sampling points was set to 10, and the maximum number of search iterations was limited to 50. After the above optimization search process, the optimal structure of the multilayer perceptron module was finally set as a three-layer shared fully connected layer with a dimensional variation of "64-128-64". The baseline fitting branch and the residual peaking branch were established as independent linear mapping networks with dimensionality reduction variations of "64-32-16". The multilayer perceptron feature upscaling network used for the final output decoupling weights comprises a three-layer fully connected structure. It sequentially maps the 32-dimensional prediction weight matrix (composed of a 16-dimensional baseline expectation and a 16-dimensional boundary constraint) to 64 dimensions, then to 32 dimensions, and finally compresses it to a strictly 2-dimensional channel at the output, accurately aligning the decoupling objectives of temperature control and lighting. After the nonlinear activation operations of all the aforementioned deep hidden layers, a random deactivation mechanism with a dropout rate of 0.2 is uniformly enforced to prevent severe overfitting when fitting complex electricity consumption time-series patterns.

[0086] Before formally beginning iterative training of model parameters, the acquired periodic alignment features and node load sequences, among other full datasets, are divided into time-series subsets. To avoid future information leakage from the time-series data, the first 80% of the full dataset is divided into a training set, the next 10% into a validation set, and the last 10% into a test set, in chronological order. The training set is used for backpropagation and updating of the network's underlying weight parameters; the validation set does not participate in direct gradient updates and is specifically used to evaluate the model's generalization performance during training to guide hyperparameter optimization and trigger early stopping mechanisms; the test set, after model training and hyperparameter optimization are fully completed, is used to independently evaluate the physical constraint generalization ability of the final decoupled network model. This test set is strictly invisible throughout the entire training and validation lifecycle and does not participate in any parameter tuning.

[0087] This embodiment employs a self-supervised joint loss function to guide the backpropagation and iterative update of model parameters. During the training phase, the specific process of constructing the joint loss function includes: extracting the predicted temperature control energy consumption value and the lighting energy consumption value at each time step, calculating their arithmetic sum to obtain the predicted total active power sequence; calculating the mean square error between this predicted total active power sequence and the original input node load sequence (actual total power), and using this as the reconstruction conservation loss evaluation metric. This evaluation metric is used to constrain, at the macroscopic level, that the sum of the predicted sub-items of energy consumption conforms to the physical law of conservation of energy.

[0088] The baseline expectation matrix and the boundary constraint tensor are extracted from the feature space. After L2 norm normalization, the average absolute value of their inner product along the feature dimension is calculated as the evaluation metric for the orthogonal independent constraint loss. By minimizing this evaluation metric, a soft constraint condition is established to prevent excessive interweaving and confusion of lighting and temperature control energy consumption features while allowing the two branches to retain a certain degree of thermodynamic positive coupling correlation. Simultaneously, to avoid pseudo-orthogonality traps where numerical satisfaction is achieved but semantic separation is not achieved in the early stages of model training, a progressive weight warm-up strategy is adopted, linearly and smoothly increasing the orthogonal constraint loss weight from zero to 0.1 in the first 10 training epochs. This orthogonal constraint loss weight can be adaptively adjusted according to seasonal operating conditions, with the weight appropriately reduced under highly correlated coupling conditions such as summer cooling.

[0089] The lighting energy consumption values ​​are extracted from multiple consecutive time steps, and the average of the absolute values ​​of the first-order differences between the lighting energy consumption values ​​of adjacent time steps is calculated as the baseline smoothness loss assessment metric. This constraint forces the network output of conventional lighting and infrastructure load to remain relatively stable, conforming to the normal steady-state operation physical laws of the infrastructure.

[0090] To construct a self-supervised, decoupled physical calibration anchor, physical sub-meters were temporarily installed in representative areas of the target building (e.g., selecting 10% of typical floors) for a one-month basic data collection period. The sample size of this calibration set was required to cover at least 10% of the physical sub-meter data of typical floors. These 10% of typical floors needed to be selected through stratified balanced sampling based on the building's functional zoning (e.g., including office areas, meeting areas, and computer rooms). At the same time, the one-month basic data collection time window had to be set in the month between the cooling season and the transitional season to ensure that the data covered both high-frequency start-up and complete shutdown of the HVAC system. The calibration set data was directly extracted proportionally from the aforementioned 80% training set to ensure a reasonable matching ratio between the model parameters and the supervision information, preventing overfitting.

[0091] Because the daily variation coefficient of conventional lighting load is usually much lower than that of temperature-controlled load according to the energy consumption statistics of typical commercial buildings, but in building scenarios with a large number of dimmable lighting systems, unsupervised constraints alone can easily lead to confusion. Therefore, the mean square error of the temperature-controlled energy consumption prediction value output by the model is compared with the actual temperature-controlled electricity meter reading. By introducing this part of the real calibration calculation, the model is forced to complete the correct semantic alignment between the stable baseline and dynamic fluctuations in the feature latent space, and the self-supervised constraints are guided to converge in the physically correct direction.

[0092] A grid search was performed on the validation set with uniformly distributed weights for the three factors, within the range of [0.05, 2.0] and a step size of 0.05. The minimum total loss on the validation set was used as the selection criterion, ultimately determining the reconstruction conservation loss weight to be 1.0, the orthogonal constraint loss weight to be 0.1, and the baseline smoothness loss weight to be 0.05. That is, the formula for calculating the total loss function on the validation set is explicitly set as: total loss equals reconstruction conservation loss multiplied by 1, plus orthogonal constraint loss multiplied by 0.1, plus baseline smoothness loss multiplied by 0.05. For building types with significantly different temperature control load fluctuation characteristics, it is recommended to appropriately lower the baseline smoothness loss weight to the range of 0.02 to 0.03 to accommodate the strong load fluctuations caused by frequent equipment start-ups and shutdowns.

[0093] During the parameter optimization phase, an adaptive moment estimation optimizer is used to perform backpropagation based on the total loss value, calculating the gradient distribution of all fully connected weight matrices, gated weight matrices, and projected weight matrices in the network. The initial learning rate is set to 0.001, and a cosine annealing learning rate decay strategy is introduced to ensure smooth convergence in the later stages of training. The batch size for training is set to 32, and the total number of training iterations is set to 100. An early stopping mechanism is also introduced: if the total loss value on the validation set does not decrease for 10 consecutive iterations, the training process is terminated early, and the network weight parameter file with the minimum validation set loss is saved. This completes the internalization and convergence of the data-driven network model to the real operating rules of the building energy consumption system.

[0094] This invention provides a deep learning-based method for decoupling multiple indicators of building energy consumption. By acquiring environmental thermal field sequences and performing cross-correlation and sliding window translation, it eliminates temporal phase deviations in the input sequences. A spatial adjacency matrix is ​​constructed based on a business flow graph, enabling the model to accurately capture the spatial energy consumption patterns of personnel and business interactions. Thermoelectric feature tensors are input into a dual-branch network, and baseline expectation and net residual matrices are extracted independently, achieving orthogonal decomposition of different energy consumption sub-items. Furthermore, a thermal boundary mask matrix is ​​generated by combining supply and return water temperature thresholds, forcibly filtering illegal boundary association positions during the attention scaling stage. This scheme organically integrates spatiotemporal alignment, feature decoupling, and thermodynamic constraints, effectively avoiding the intertwining interference of multiple energy consumption features during extraction and predictive calculations that violate physical principles. This fundamentally contributes to the accuracy and physical consistency of the final decoupled energy consumption data.

[0095] Example 2 This embodiment describes the application of a deep learning-based decoupling method for building energy consumption across multiple indicators to a large comprehensive hospital, which faces the challenges of high-intensity constant temperature and humidity requirements and the routine operation of high-energy-consuming medical equipment.

[0096] First, using a fixed 10-minute time step, dry-bulb temperature and relative humidity in each clinic and ward were collected via a wireless LAN probe and environmental sensor array to construct an environmental thermal field sequence. Simultaneously, supply and return water temperature sequences were read from the manifold in the underground cold and heat source room. The raw total active power was read from smart meters in the distribution boxes on each floor to form a node load sequence. Furthermore, by combining the hospital's static building information model with the outpatient registration and ward personnel flow system, a business flow map was extracted.

[0097] Given the significant thermal inertia of the hospital's thick brick walls, a Fast Fourier Transform (FFT) was applied to the environmental thermal field sequence and the nodal load sequence to extract the frequency domain tensor. The cross-correlation product was then calculated to obtain the cross-power spectral density. The peak value of the time-domain cross-correlation was established through an inverse transform, determining the hospital's response delay management vector to be 6 time steps (i.e., 60 minutes). Based on this, the thermal field sequence was shifted and compensated. Subsequently, a sliding window of length 24 was used to extract the windowed tensor, which was then concatenated along the channel dimension to generate periodically aligned features.

[0098] The process involves analyzing the workflow graph, such as calculating the probability of patient flow from the outpatient hall to the radiology department per unit time, and combining this with the frequency of shared fresh air duct connections between adjacent wards within the building space. Multiplying these two values ​​yields the node collaborative management weights, constructing an initial graph matrix. Symmetrical connectivity attributes are then extracted sequentially, and the main diagonal is filled to construct a self-collaborative evaluation tensor. This tensor is then combined with the degree matrix to perform continuous matrix multiplication, ultimately generating a high-fidelity spatial adjacency matrix.

[0099] In the spatial energy efficiency map supervision layer, the periodic alignment features are reduced in dimensionality using a pre-defined weight tensor and aggregated with the spatial adjacency matrix. The output features are then processed by linear unit activation with leakage correction and multi-head attention distance allocation to generate a spatial topology tensor. This tensor is forward propagated to the time-series energy consumption trend prediction layer, where the hidden state and cell state from the previous time step are combined. Through nonlinear mapping of forget gates, input gates, and output gates, the cell state is updated, and a thermoelectric feature tensor incorporating spatiotemporal characteristics is output.

[0100] Multilayer perceptron is used to process thermoelectric feature tensors to extract shared hidden layer features. On one hand, the features are guided into the baseline fitting branch, and a baseline expectation matrix representing the power consumption of basic lighting and normal medical instruments is generated through fully connected matrix multiplication. On the other hand, the bias matrix is ​​obtained through the residual peaking branch, and the mean is calculated along the channel. Through element-wise cancellation operation, the pure net residual matrix is ​​extracted.

[0101] To address the alternating heating and cooling characteristics of northern hospitals in winter and summer, a threshold of 45-60 degrees Celsius was set for hot water supply in winter and 5-15 degrees Celsius for cold water supply in summer. Real-time water temperature was compared to these thresholds; if it exceeded the allowable range, a maximum negative value was output to generate a thermal boundary mask matrix. Key-value and query tensors were generated using linear projection. After calculating the unscaled attention score, illegal physical-thermal relationships were filtered out using this mask matrix. Finally, a boundary constraint tensor was generated through exponential normalization.

[0102] The baseline expectation matrix and boundary constraint tensor are concatenated along the feature depth channels and fed into a feature upscaling network to extract the decoupling weight matrix. Through channel extraction, the temperature control weight channel and the conventional lighting and medical equipment weight channel are separated. Both are then subjected to Hadamard product and global average pooling calculations with the original node load sequence to obtain the decoupled temperature control energy consumption value and the base energy consumption value. Finally, these are integrated into energy consumption decoupling assessment data and output to the logistics intelligent management platform.

[0103] 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 deep learning-based method for decoupling multiple building energy consumption indicators, characterized in that, include: Based on edge IoT sensing networks, environmental thermal field sequences, supply and return water temperature sequences, node load sequences, and business flow maps are obtained. Sliding window translation calculations are performed on the environmental thermal field sequence and the nodal load sequence to output the period alignment characteristics; Construct a spatial adjacency matrix based on the business flow graph; The spatial adjacency matrix and periodic alignment features are input into the energy efficiency management analysis engine for spatiotemporal aggregation, and the thermoelectric feature tensor is output. The thermoelectric feature tensor is input into a dual-branch feature decoupling network. The baseline expectation matrix is ​​output through the baseline fitting branch, the bias matrix and mean matrix are output through the residual peaking branch, and the net residual matrix is ​​established through broadcast subtraction. The supply and return water temperature sequences are truncated from the preset water temperature threshold, and the thermal boundary mask matrix is ​​output. The thermoelectric feature tensor and the net residual matrix are linearly projected to output the key tensor, value tensor and query tensor. The thermal boundary mask matrix is ​​used for filtering, and dot product scaling and dot product calculation are performed to output the boundary constraint tensor. The baseline expectation matrix and boundary constraint tensor are spliced ​​together to construct the decoupling weight matrix; the decoupling weight matrix and the node load sequence are multiplied by Hadamard to output the temperature control energy consumption value and the lighting energy consumption value, thus generating energy consumption decoupling assessment data.

2. The deep learning-based multi-index decoupling method for building energy consumption according to claim 1, characterized in that, The specific generation process of the periodic alignment feature includes: applying a fast Fourier transform to the environmental thermal field sequence and the node load sequence to extract the periodic thermal field evaluation quantity and the periodic electrical work evaluation quantity; calculating the cross-correlation evaluation product of the periodic thermal field evaluation quantity and the periodic electrical work evaluation quantity to obtain the cross-power spectral density tensor; processing the cross-power spectral density tensor using an inverse fast Fourier transform to generate a time-domain cross-correlation function sequence; extracting the peak index of the time-domain cross-correlation function sequence corresponding to each node to establish the response delay management vector; shifting the environmental thermal field sequence according to the response delay management vector to obtain the hysteresis-compensated thermal field sequence; performing constant-length sliding slice extraction calculations on the hysteresis-compensated thermal field sequence and the node load sequence respectively to obtain the windowed thermal field tensor and the windowed electrical work tensor respectively; and concatenating the windowed thermal field tensor and the windowed electrical work tensor in the channel dimension to construct the periodic alignment feature.

3. The deep learning-based multi-index decoupling method for building energy consumption according to claim 1, characterized in that, The specific process of constructing the spatial adjacency matrix includes: based on the business flow graph, analyzing and extracting the spatial business flow probability and connection frequency between building space entity nodes; calculating the product of the spatial business flow probability and the connection frequency as the node collaborative management weight, and constructing an initial graph matrix accordingly; extracting the symmetric connection attribute of the initial graph matrix and establishing a symmetric connection tensor; filling the main diagonal elements of the symmetric connection tensor with preset constants to construct a self-collaborative evaluation tensor; performing numerical summation and aggregation operations on the self-collaborative evaluation tensor along the row direction to obtain a degree matrix tensor, and normalizing it to obtain a normalized degree matrix tensor; performing continuous matrix multiplication operations between the normalized degree matrix tensor and the self-collaborative evaluation tensor to generate a spatial adjacency matrix.

4. The deep learning-based multi-index decoupling method for building energy consumption according to claim 1, characterized in that, The energy efficiency management analysis engine includes a spatial energy efficiency map supervision layer and a time-series energy consumption trend prediction layer. In the spatial energy efficiency map supervision layer, the period alignment features are linearly mapped and multiplied using a preset spatial management evaluation weight tensor to obtain a dimension-reduced node feature matrix. The spatial adjacency matrix and the dimension-reduced node feature matrix are then aggregated using standard matrix multiplication to extract a primary aggregated feature matrix. The primary aggregated feature matrix is ​​processed by nonlinear mapping with leakage correction linear units to generate a primary spatial topology tensor. The primary spatial topology tensor is fed into a preset multi-head attention allocation mechanism to complete node distance allocation and obtain a weighted spatial feature tensor. Layer normalization numerical constraint calculation and residual addition operation are performed on the weighted spatial feature tensor in sequence to establish the spatial topology tensor. The spatial topology tensor is forward-transferred to the time series energy consumption trend prediction layer as a single-step topology input tensor.

5. The deep learning-based multi-index decoupling method for building energy consumption according to claim 4, characterized in that, The specific process of generating the thermoelectric feature tensor includes: in the time-series energy consumption trend prediction layer, extracting the hidden state tensor and the core state evaluation tensor of the previous time step; concatenating the single-step topological input tensor and the hidden state tensor along the channel dimension to construct a joint spatiotemporal tensor; performing a nonlinear mapping on the joint spatiotemporal tensor using a preset set of gating weight matrices to extract the redundancy removal evaluation tensor, the trend introduction evaluation tensor, and the output gating tensor respectively; performing state update calculation by combining the trend introduction evaluation tensor, the redundancy removal evaluation tensor, and the core state evaluation tensor to obtain the current step core state evaluation tensor; calculating the activation mapping of the current step core state evaluation tensor and the Hadamard product of the output gating tensor to generate the thermoelectric feature tensor.

6. The deep learning-based multi-index decoupling method for building energy consumption according to claim 1, characterized in that, The specific generation process of the baseline expectation matrix and the net residual matrix includes: processing the thermoelectric feature tensor using a multilayer perceptron for dimensionality reduction and fusion calculation, and extracting the shared hidden layer feature tensor; guiding the shared hidden layer feature tensor into the baseline fitting branch, and generating the baseline expectation matrix using matrix multiplication with a preset baseline fitting fully connected weight tensor; passing the shared hidden layer feature tensor through the residual peaking branch, and obtaining the bias matrix by matrix multiplication with a preset peaking fully connected weight tensor; performing mean summation calculation on the bias matrix along the feature channel dimension to obtain the residual mean scalar matrix; expanding the residual mean scalar matrix to the scale of the bias matrix along the channel dimension to construct the mean matrix; calculating the element-wise subtraction cancellation result between the bias matrix and the mean matrix to establish the net residual matrix.

7. The deep learning-based multi-index decoupling method for building energy consumption according to claim 1, characterized in that, The specific generation process of the boundary constraint tensor includes: linearly mapping the thermoelectric feature tensor and the net residual matrix using a preset set of projection weight matrices to extract the key tensor, the value tensor, and the query tensor; calculating the inner product multiplication result of the query tensor and the key tensor to establish an unscaled attention score matrix; applying an additive masking process to the unscaled attention score matrix using the thermal boundary mask matrix to filter illegal boundary association positions and obtain a masking score matrix; performing dot product scaling and exponential normalization calculations on the masking score matrix to construct an attention weight normalization matrix; and generating the boundary constraint tensor by performing matrix multiplication superposition calculations between the attention weight normalization matrix and the value tensor.

8. The deep learning-based multi-index decoupling method for building energy consumption according to claim 1, characterized in that, The specific process for generating energy consumption decoupling assessment data includes: concatenating the baseline expectation matrix and the boundary constraint tensor along the feature depth channel to construct a prediction weight matrix; feeding the prediction weight matrix into a preset multilayer perceptron feature upscaling network to perform nonlinear mapping upscaling calculation to extract the decoupling weight matrix; applying channel extraction and segmentation operations to the decoupling weight matrix to separate the temperature control weight channel matrix and the lighting weight channel matrix; calculating the Hadamard product of the temperature control weight channel matrix and the node load sequence, and performing global average pooling calculation to obtain the temperature control energy consumption value; calculating the Hadamard product of the lighting weight channel matrix and the node load sequence, and performing global average pooling calculation to obtain the lighting energy consumption value; and integrating the data structures of the temperature control energy consumption value and the lighting energy consumption value through serialization combination operations to generate energy consumption decoupling assessment data.

Images