Method for predicting energy-saving effect of building envelope

Abstract

The invention relates to the technical field of energy consumption prediction, in particular to a building envelope energy-saving effect prediction method, which specifically comprises the following steps: deploying a sensor in a building maintenance structure to collect related data, and constructing a data set; marking the collected data to form a training set; a physical constraint decoupling normalization method is adopted for the data in the training set to generate features after decoupling normalization; a weighted feature vector is generated by adopting an attention mechanism guided by physical prior; constructing an energy-saving effect prediction network, and inputting the weighted feature vectors into the network for prediction; optimizing the network to obtain a trained network; newly-collected data is processed and then input into the trained network, the final prediction probability of each energy-saving grade is output, and the grade with the maximum probability is taken as a prediction result. According to the method, the collected data is preprocessed and then input into the network, so that the defects of inaccurate data processing, unreasonable feature selection and the like can be overcome, and the accuracy of a prediction result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy consumption prediction, and in particular to a method for predicting energy-saving effects of building envelope structures. Background Art

[0002] With the increasing demand for building energy efficiency, accurately predicting the energy-saving effects of building envelopes has become an important issue in building design and energy-saving optimization. Building envelopes play a vital role in the thermal performance of buildings, and their thermal physical processes such as heat conduction, convection, and radiation directly affect the building's energy consumption. In order to improve building energy efficiency and meet energy-saving standards, it is necessary to accurately monitor and evaluate the thermal performance of building envelopes. However, existing building energy-saving effect prediction methods generally face challenges, such as inaccurate data processing, unreasonable feature selection, the inability of model networks to fully capture the complexity of physical mechanisms, and deficiencies in uncertainty quantification.

[0003] The Chinese invention patent with publication number CN120181334A proposes a method and system for predicting and supervising energy consumption at construction sites. It belongs to the field of construction site technology. By combining construction plans, machine hours and weather forecast data, it can accurately predict the energy consumption of construction sites, provide a scientific basis for construction management, and collect energy consumption data and weather information to dynamically adjust the energy allocation plan to ensure priority energy supply for key equipment and improve energy utilization efficiency; analyze historical energy consumption data and combine it with weather forecasts to achieve accurate prediction of future energy consumption and enhance the intelligent level of energy management; this invention can adapt to different weather conditions and construction stages, has strong versatility and applicability, and ensures the normal operation of key equipment by rationally allocating energy, improves construction efficiency and shortens project construction period.

[0004] However, the existing technology still has the following problems that need to be further solved: the existing technology mostly adopts a simple linear normalization method, which may lead to the distortion of feature distribution when dealing with the dimensional and scale differences of different physical quantities, destroy the physical correlation of the data, and affect the accurate prediction of energy-saving effects; the existing technology often uses traditional feature selection methods, directly discards features with weak correlation, or adopts an equal-weighted method to process all features, which is easy to lose physical weak signals and cannot effectively highlight the key physical features closely related to energy-saving effects, resulting in reduced model network performance; traditional neural networks usually rely on a single-channel architecture and cannot simultaneously capture local physical patterns and global physical properties, thereby limiting the model network's comprehensive learning and prediction capabilities for the complex thermal characteristics of building envelope structures; the existing technology mostly ignores uncertainty quantification in the prediction process, and the conventional Bayesian method fails to consider physical constraints such as thermodynamics, which may cause the prediction results to violate the laws of thermodynamics. Especially when the sensor noise is large, the prediction of the traditional method is unreliable and cannot meet the needs of practical applications.

[0005] Therefore, the present invention proposes a method for predicting the energy-saving effect of building envelope structures to solve the above problems. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention develops a method for predicting the energy-saving effect of building envelope structures. By preprocessing the collected data and then inputting it into the network, the present invention can overcome the shortcomings of inaccurate data processing, unreasonable feature selection, inability to fully capture the complexity of physical mechanisms, and deficiencies in uncertainty quantification, thereby improving the accuracy of the prediction results.

[0007] The technical solution to the technical problem of the present invention is a method for predicting the energy-saving effect of a building envelope structure, comprising the following steps: S1. Deploy various types of sensors at key locations of building maintenance structures to collect real-time data on physical quantities directly related to building thermal performance, and construct a data set based on the collected data; S2. Based on building thermal engineering principles and energy-saving evaluation standards, the collected sensor data is double-labeled, including physical indicator labeling and energy-saving level labeling. The labeled labels are mapped one-to-one with the collected data to form a training set. S3. Apply the physical constraint decoupling normalization method to the multi-dimensional monitoring data in the training set. Construct a decoupling matrix through covariance feature decomposition to eliminate dimensional differences and release the nonlinear coupling between physical features, generating decoupled and normalized features. S4. Adopting the attention mechanism guided by physical priors, we decouple the correlation between normalized features and labels through mutual information quantification, and generate adaptive weights based on physical contribution priors to generate weighted feature vectors. S5. Construct an energy-saving effect prediction network, which includes a local-global dual-channel gating fusion network module, a physically constrained Dirichlet evidence uncertainty quantification module, a dual-driven classification decision-making with differentiable physical rule gating, and a multi-objective loss function calculation. The weighted feature vector is input into the network for prediction. S6. Optimize the trainable parameters of the energy-saving effect prediction network through iterative optimization, use Xavier normal distribution to initialize the trainable parameters, use an adaptive moment estimation optimizer to minimize the total loss of the network, and set the convergence condition for stopping iterative training, ultimately obtaining a trained energy-saving effect prediction model network; S7. After physical decoupling normalization and feature adaptive weighting, the newly collected data is input into the trained energy-saving effect prediction network, and the final prediction probability of each energy-saving level is output. The level with the highest probability is taken as the prediction result.

[0008] The data collection process in S1 is as follows: Data collection adopts a distributed Internet of Things architecture. Sensors continuously upload data to the central server in a cycle of 1-10 minutes, and synchronize the multi-dimensional data through timestamp alignment to form a full-dimensional original monitoring data set of multi-physical mechanisms.

[0009] The double annotation in S2 is as follows: Physical indicator labeling: divide data dimension labels according to sensor types; Energy-saving grade marking: Based on the national standard "Energy-saving Design Standard for Public Buildings" or based on measured energy consumption data, the energy-saving grade of the enclosure structure performance in each monitoring period is assessed by manual marking; Divided into five levels from 1 to 5, 1 represents the lowest level and 5 represents the highest level; The energy-saving grade is labeled based on the heat loss per unit area, thermal inertia index, and energy consumption deviation from the baseline model network; The physical indicator labels and energy-saving grade labels are stored in a structured form and mapped one-to-one with the collected data to form a training set.

[0010] The physical constraint decoupling and normalization method in S3 is as follows: S3.1. Calculate the physical empirical mean and dimensionless standard deviation: Based on all samples in the historical data set of the building maintenance structure, the physical empirical mean and dimensional standard deviation of the physical quantity in each dimension are calculated. The reference center position of each physical quantity is determined based on the physical empirical mean vector, and the typical range of the numerical fluctuation of each physical quantity itself is determined based on the dimensional standard deviation vector. S3.2. Covariance matrix construction and eigendecomposition: The historical data set is centralized using the physical empirical mean vector. The covariance matrix between physical quantities in each dimension is then calculated based on the centralized data. The degree of correlation between different physical quantities is then quantified through the covariance matrix. S3.3. Solve the physical decoupling matrix: Perform eigendecomposition on the covariance matrix to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The column vectors of the eigenvector matrix represent the direction of change of the principal components, and the eigenvalue diagonal matrix represents the size of the variance contribution in the direction of change of each principal component. Based on the eigenvector matrix and the eigenvalue diagonal matrix, a physical decoupling matrix is ​​constructed to eliminate the coupling relationship between different physical quantities caused by dimensional differences and nonlinear interaction effects, so that the transformed features are statistically independent of each other, and then the physical decoupling matrix is ​​obtained; S3.4. Perform decoupled normalization calculation: The data in the historical data set are normalized by physical dimensions to eliminate the influence of the dimensions and scale differences of different physical quantities. Then, the normalized eigenvector is multiplied by the physical decoupling matrix to obtain the decoupled eigenvector.

[0011] The adaptive feature weighting process guided by physical priors in S4 is as follows: S4.1. Calculation of feature-label mutual information: For each decoupled feature vector dimension after decoupling normalization, calculate the mutual information value between the feature and the building energy efficiency grade label; S4.2. Constructing physical attention weights: The mutual information value of each feature dimension and the label is integrated, as well as the preset physical contribution prior value. The concentration of the weight distribution is controlled by using a sharpening factor. The fusion result is passed through an exponential function and normalized to generate an attention weight value between 0 and 1 for each feature dimension. S4.3. Perform feature adaptive weighting: The attention weight vector is multiplied element-by-element by the decoupled feature vector. According to the attention weight value of each feature, the importance of physical mechanisms and statistically strongly correlated features is amplified, while the influence of weakly correlated features is weakened to obtain a weighted feature vector.

[0012] The operation of the energy-saving effect prediction network in S5 is as follows: The weighted feature vector is input into the network to predict the energy saving effect. The weighted feature vector is passed through the local-global dual-channel gated fusion network module to extract local features and global features respectively, and then the extracted features are subjected to gated adaptive feature fusion; The fused features are input into the Dirichlet evidence uncertainty quantification module of physical constraints to generate physical constraint residuals that aggregate various thermal constraints. The fused features are then mapped to Dirichlet concentration parameters for each energy-saving level. The uncertainty loss of the network is then calculated based on the physical constraint residuals and the Dirichlet concentration parameters for each energy-saving level. A differentiable physical decision gating strategy is adopted. First, the physical rule confidence level is defined for each energy conservation level based on thermodynamic laws and engineering experience. Then, the predicted Dirichlet concentration parameter is dynamically reconciled with the physical rule confidence level to generate the predicted probability of the energy conservation level. Finally, the loss between the predicted Dirichlet concentration parameter and the physical rule confidence level is calculated. The multi-objective loss function is calculated to obtain the total loss of the network, including the cross entropy loss between the prediction result and the true label, as well as the loss calculated by the two modules in the network, and the weight coefficient is set for each loss.

[0013] The local-global dual-channel gated fusion network module specifically adopts a dual-branch gated fusion architecture. One branch extracts local physical pattern features through convolution operations, and the other branch learns global features through a fully connected layer. The two are then fused through an adaptive gating mechanism. The specific steps are as follows: (1) Physical feature grouping and input allocation: According to the physical characteristics, the weighted feature vector of each sample is divided into a local feature subset and a global feature subset. The local feature subset contains physical quantities with spatial correlation and the dimension is ; The global feature subset contains non-space-related physical quantities, with dimensions of ; 2) Extract local spatial feature convolution: Perform a one-dimensional convolution operation on the local feature subset, using the convolution kernel to slide calculations in the spatial dimension to extract the spatially related physical patterns contained in the local features. The convolution calculation results are processed by an activation function to obtain a local feature vector containing local physical pattern information; 3) Global physical characteristics full connection extraction Perform a fully connected transformation operation on the global feature subset, perform linear combination and nonlinear activation on the global features through the weight matrix and bias vector, learn and generate feature representations that reflect non-spatial physical characteristics, and obtain a global feature vector; 4) Gated Adaptive Feature Fusion: The local feature vector obtained by convolution processing and the global feature vector obtained by full connection processing are spliced ​​together. Based on the spliced ​​feature vector, a gating vector is generated through the gating weight matrix and activation function. Then, the gating vector is used to perform weighted fusion of the local feature vector and the global feature vector to achieve feature fusion of physical property perception and obtain a fused feature vector.

[0014] The physically constrained Dirichlet evidence uncertainty quantification module uses the physically constrained Dirichlet evidence quantification method to generate confidence intervals that conform to both the statistical laws of machine learning and the laws of thermodynamics. The specific process is as follows: 1) Constructing thermodynamic residual energy penalty term: A physical constraint violation metric function is defined to quantify the extent to which the prediction results obtained by the energy-saving effect prediction network violate physical laws. Physical laws specifically refer to the laws of thermodynamics. The function calculates whether the predicted energy values ​​exceed the maximum allowable energy values ​​set based on the thermodynamic limits of the materials. If the predicted value exceeds the allowable value, a penalty term is applied, ultimately resulting in a physical constraint residual, which is the sum of the energy penalty terms for all constraint violations. 2) Dirichlet concentration parameter mapping: The fused feature vector is input into the fully connected layer and mapped into a concentration parameter vector of the Dirichlet distribution through an exponential activation function and an addition operation. Each concentration parameter corresponds to an energy-saving level. The Dirichlet distribution is used to describe the network's uncertainty in the probability distribution of samples belonging to each energy-saving level. The size of the concentration parameter reflects the network's confidence in the prediction of each level, thereby achieving parameterized quantification of classification uncertainty. 3) Calculate physical regularization uncertainty loss: Construct an uncertainty loss function for network training, including a Dirichlet-based evidence loss and a physical constraint residual term. The Dirichlet-based evidence loss measures the difference between the probability distribution of network predictions and the probability distribution of true labels, while quantifying the uncertainty of the network itself. The physical constraint residual term penalizes predictions that violate physical laws. The dual-driven classification decision-making with differentiable physical rule gating specifically adopts a differentiable physical decision gating strategy to dynamically reconcile data prediction and physical rules. The specific operations are as follows: 1) Build a thermodynamic rule confidence engine: Based on the laws of thermodynamics and engineering experience, the physical rule confidence is defined for each energy saving level. The physical rule confidence is calculated by differentiable functions, taking into account the measured values ​​of key physical indicators. For each indicator, the measured value is compared with the threshold value preset for each level, and the confidence is calculated by The function calculates the degree to which the indicator satisfies the rules of each level and obtains the physical rule confidence of each level. The final physical rule confidence is the product of the degree to which all key physical indicators of the sample satisfy the corresponding level rules. A physical rule confidence value close to 1 indicates that the physical indicators of the sample strongly support its belonging to that level, while a value close to 0 indicates a strong violation. 2) Synthetic Data - Rule-gated Probability: The network is fused with the expected probability of the Dirichlet distribution predicted by the fusion feature vector and the confidence of the physical rule, specifically through a learnable gating weight scalar, and the gating weight scalar is obtained by the fusion feature vector through the gating weight vector and The function is calculated, and the relative importance of data prediction and physical rules in the final decision is dynamically controlled through the gating mechanism. The data prediction probability and the physical rule confidence are weighted and summed according to the gating weight scalar to obtain the final prediction probability, and then the prediction result is obtained; 3) Calculate the physical decision boundary loss constraint: The physical decision boundary loss is applied to the expected probability of the Dirichlet distribution predicted by the network. When the confidence of the physical rules of a certain level is lower than the set threshold, indicating that the physical indicators of the sample strongly do not support it belonging to this level, the physical decision boundary loss function will punish the high probability value predicted by the network for this level, forcing the network to increase the uncertainty of its prediction, that is, to reduce the probability value of this level, thereby ensuring that the network will not make overconfident and erroneous predictions when violating the physical rules.

[0015] The total loss function of the energy-saving effect prediction network is calculated through a multi-objective loss function. The specific operations are as follows: Calculate the total loss function of network training. The total loss function is a multi-objective loss function, which consists of three parts: the cross entropy loss function between the predicted results and the actual labels, the uncertainty loss function, and the physical decision boundary loss function. The uncertainty loss function and the physical decision boundary loss function are weighted by weight coefficients respectively. The uncertainty loss weight coefficient is set to a large value close to 1 at the beginning of training, and decreases smoothly to 0 according to the cosine decay function as the training rounds progress; The physical decision boundary loss weight coefficient is set to 0 at the beginning of training, and increases smoothly to 1 according to the Sigmoid function as the training round approaches a specific round.

[0016] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects: The building sensor monitoring data collected by the present invention is multi-dimensional data with multi-dimensional coupling characteristics and different dimensions. Therefore, a physical constraint decoupling normalization method is adopted to eliminate the nonlinear interaction effects between different physical quantities through covariance eigendecomposition, thereby ensuring that the multi-physical quantity data can be processed unbiasedly in the same feature space. This can avoid the traditional linear normalization method that only performs linear scaling and ignores the nonlinear interaction effects between physical quantities, resulting in distorted feature distribution and destruction of physical correlation. Since the decoupled and normalized features still have significant correlation differences with the energy-saving level, the present invention adopts an adaptive feature weighting mechanism guided by physical priors. By calculating the mutual information between the features and the energy-saving level and combining the physical contribution priors, it automatically assigns weights to each feature. This ensures that the model can better highlight the key physical features related to the energy-saving effect, overcoming the problems caused by traditional methods that ignore weak physical signals or perform equal weighting. This paper innovatively adopts a dual-channel gated fusion network structure. One branch extracts local features through convolution operations, and the other branch extracts global features through a fully connected layer. The two are then fused through an adaptive gating mechanism. This can fully capture the complex thermal performance of building envelopes, avoid the feature loss problem of single-channel models, and improve prediction accuracy. In order to solve the problems of sensor noise and model prediction uncertainty, the present invention adopts a physically constrained Dirichlet evidence uncertainty quantification method. By mapping the thermodynamic residual energy penalty term with the Dirichlet concentration parameter, a confidence interval that conforms to the laws of thermodynamics is generated, thereby effectively quantifying the uncertainty of the prediction and avoiding the problem that traditional methods may violate the laws of thermodynamics.

[0017] In summary, the present invention can overcome the shortcomings of inaccurate data processing, unreasonable feature selection, inability to fully capture the complexity of physical mechanisms, and deficiencies in uncertainty quantification, thereby improving the accuracy of prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0019] Figure 1 Schematic diagram of the method of the present invention.

[0020] Figure 2 Comparison chart of prediction accuracy of different normalization methods.

[0021] Figure 3 A comparison chart of the effects of different normalization methods based on heat flux density.

[0022] Figure 4 Comparison of feature distribution after normalization using different methods.

[0023] Figure 5 Performance comparison chart of different feature weighting strategies.

[0024] Figure 6 A comparison chart of the effects of physical constraints on prediction uncertainty.

[0025] Figure 7The bar chart shows the prediction accuracy of different methods at each energy saving level. DETAILED DESCRIPTION

[0026] To clearly illustrate the technical features of this solution, the present invention is described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and configurations of specific examples are described below.

[0027] Example 1 A method for predicting energy-saving effects of building envelope structures is as follows: S1. Deploy various types of sensors at key locations of building maintenance structures to collect real-time data on physical quantities directly related to building thermal performance, and construct a data set based on the collected data; Physical quantity data directly related to building thermal performance include heat flow data, temperature data, material property data, environmental parameters, etc. Heat flow data: Use heat flow sensors to continuously monitor the dynamic changes in heat flux density on the inner and outer surfaces of the enclosure structure, and record the heat transfer value per unit area per unit time in "W / m²"; Temperature data: High-precision temperature sensors are placed on the inner and outer surfaces of the enclosure structure, interlayers, and material interfaces to simultaneously collect surface temperature differences, material internal temperature gradients, and ambient temperature time series data in degrees Celsius. Material property data: Use embedded thermal conductivity sensors or laboratory calibration devices to obtain the real-time thermal conductivity of wall and insulation materials in "W / m·K", and use thickness gauges to record the thickness of material structure layers in "m"; Environmental parameters: Integrated humidity, wind speed, and solar radiation sensors collect ambient relative humidity (in %), near-surface wind speed (in m / s), and solar radiation intensity (in W / m²) to quantify the impact of external climate on heat transfer; In a specific implementation, data collection adopts a distributed Internet of Things architecture, where sensors continuously upload data to a central server in a cycle of 1-10 minutes, and synchronize the multi-dimensional data through timestamp alignment to form a full-dimensional raw monitoring data set of multi-physical mechanisms; Multi-physics mechanisms include heat conduction, convection, radiation, etc.

[0028] S2. Based on building thermal engineering principles and energy-saving evaluation standards, the collected sensor data is double-labeled, including physical indicator labeling and energy-saving level labeling. The labeled labels are mapped one-to-one with the collected data to form a training set. In a specific implementation, the double marking is as follows: Physical indicator labeling: divide data dimension labels according to sensor types; The data includes multi-dimensional physical quantities such as heat flux, temperature difference series, thermal conductivity, thickness, etc. Energy-saving grade marking: Based on the national standard "Energy-saving Design Standard for Public Buildings" (GB50189-2015), or based on measured energy consumption data, the energy-saving grade of the enclosure structure performance in each monitoring period is assessed by manual marking; Divided into five levels from 1 to 5, 1 represents the lowest level and 5 represents the highest level; The energy-saving grade is labeled based on the heat loss per unit area, thermal inertia index, and energy consumption deviation from the baseline model network; The physical indicator labels and energy-saving grade labels are stored in a structured form and mapped one-to-one with the collected data to form a training set.

[0029] S3. Apply the physical constraint decoupling normalization method to the multi-dimensional monitoring data in the training set. Construct a decoupling matrix through covariance feature decomposition to eliminate dimensional differences and release the nonlinear coupling between physical features, generating decoupled and normalized features. S3.1. Calculate the physical empirical mean and dimensionless standard deviation: Based on all samples in the historical data set of the building maintenance structure, the physical empirical mean and dimensional standard deviation of the physical quantity in each dimension are calculated. The reference center position of each physical quantity is determined based on the physical empirical mean vector, and the typical range of the numerical fluctuation of each physical quantity itself is determined based on the dimensional standard deviation vector. The calculation formula is as follows: , , Where, Represents the total number of samples in the historical data set; Indicates the The sample in The original eigenvalues ​​of dimensional physical quantities; Indicates the The empirical mean of a dimensional physical quantity; Indicates the dimensional standard deviation of a dimensional physical quantity; and are all positive integers; Physical empirical mean Expressed as , dimensionless standard deviation Expressed as , The number of data types in the historical dataset corresponds to the total dimension of the physical quantity. yes Dimensional column vectors can be used to centralize data, eliminate the baseline offset of each physical quantity, and make the covariance matrix only reflect the correlation between features; It should be noted that the dimensionality of the sample depends on the type of building sensor monitoring data, such as heat flux density, temperature difference between internal and external surfaces, thermal conductivity of materials, etc. S3.2. Covariance matrix construction and eigendecomposition: The historical data set is centralized using the physical empirical mean vector. The covariance matrix between physical quantities in each dimension is then calculated based on the centralized data. The degree of correlation between different physical quantities is then quantified through the covariance matrix. The calculation formula is as follows: , Where, Represents the covariance matrix containing the covariance relationship between multi-dimensional sensor data features; Indicates the The original feature vector of samples; Represents the matrix transpose operation; It should be noted that The term representation performs a central processing operation on the raw sensor data and eliminates the reference offset of each physical quantity by subtracting the mean of the corresponding physical quantity; S3.3. Solve the physical decoupling matrix: Perform eigendecomposition on the covariance matrix to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The column vectors of the eigenvector matrix represent the direction of change of the principal components, and the eigenvalue diagonal matrix represents the size of the variance contribution in the direction of change of each principal component. Based on the eigenvector matrix and the eigenvalue diagonal matrix, a physical decoupling matrix is ​​constructed to eliminate the coupling relationship between different physical quantities caused by dimensional differences and nonlinear interaction effects, so that the transformed features are statistically independent of each other, and then the physical decoupling matrix is ​​obtained; The calculation formula is as follows: , in, represents the physical decoupling matrix; Represents the eigenvector matrix, where the column vectors correspond to the principal component directions; It is a diagonal matrix of eigenvalues, and the covariance matrix Perform eigendecomposition to obtain, , The diagonal elements of The characteristic value of for The transpose of It should be noted that there is an inherent correlation between physical quantities such as heat flux and temperature difference. The eigendecomposition of the covariance matrix projects the original features into the orthogonal principal component space. The diagonalization of the covariance matrix makes the transformed features linearly independent. Therefore, the eigenvector matrix The column vector of is the principal component direction, representing the independent physical modes inherent in the data, such as pure heat conduction, convection, etc., to achieve mathematical decoupling of the physical mechanism; The term is not a conventional normalization method. The negative square root operation of the eigenvalue diagonal matrix is ​​actually to normalize the variance of the principal component, which can eliminate the scale differences in different directions. Combined with the eigenvector matrix Ensure that the transformed feature directions are aligned with the original physical mechanisms, e.g. the first principal component may correspond to the dominant heat transfer mode.

[0030] S3.4. Perform decoupled normalization calculation: The data in the historical data set are normalized in terms of physical dimensions to eliminate the influence of the dimension and scale differences of different physical quantities. Then, the normalized eigenvector is multiplied by the physical decoupling matrix to obtain the decoupled eigenvector. This not only eliminates the dimensional differences, but also releases the nonlinear coupling relationship between the original features.

[0031] The calculation formula is as follows: , Where, Indicates the The decoupled feature vector of samples.

[0032] S4. Adopting the attention mechanism guided by physical priors, we decouple the correlation between normalized features and labels through mutual information quantification, and generate adaptive weights based on physical contribution priors to generate weighted feature vectors. S4.1. Calculation of feature-label mutual information: For each decoupled feature vector dimension after decoupling normalization, calculate the mutual information value between the feature and the building energy efficiency grade label; The calculation formula is as follows: , Where, Indicates the of samples The value space of the dimensional decoupling spatial feature, Indicates the of samples dimensional decoupling spatial features, ; Represents a set of energy-saving level labels. is the specific value of the energy-saving grade label. ; express and The mutual information value between them can quantify the statistical correlation between features and labels and avoid subjective bias; express The value is and The value is The joint probability of express The value is The marginal probability of express The value is The marginal probability of For logarithmic functions, the default base is a natural constant; It should be noted that mutual information is used to quantify the contribution of the information contained in each feature dimension to the prediction of the energy-saving grade label. The larger the mutual information value, the higher the contribution of the feature dimension to the classification task. S4.2. Constructing physical attention weights: The mutual information value of each feature dimension and the label is integrated with the preset physical contribution prior value. By using a sharpening factor to control the concentration of the weight distribution, the fusion result is passed through an exponential function and normalized to generate an attention weight value between 0 and 1 for each feature dimension. This can reflect the comprehensive contribution of the feature under the dual importance of physical mechanism and statistical data. The calculation formula is as follows: , Where, Indicates the of samples The attention weight value of the dimension feature has a range of ; Indicates the of samples Mutual information between dimensional features and energy-saving grade labels; Indicates the of samples The physical contribution prior of dimensional features can be preset based on physical knowledge or expert experience, and domain knowledge can be used to correct data deviations; Indicates the of samples The physical contribution prior of dimensional features; Indicates the sharpening factor that controls the steepness of the weight distribution. ; is a natural exponential function; The attention weight vector of a sample is expressed as ; S4.3. Perform feature adaptive weighting: The attention weight vector is multiplied element-by-element by the decoupled feature vector. According to the attention weight value of each feature, the importance of physical mechanisms and statistically strongly correlated features is amplified, while the influence of weakly correlated features is weakened to obtain a weighted feature vector.

[0033] The calculation formula is as follows: , Where, Indicates the The weighted feature vector of samples; Represents the Hadamard product of element-wise multiplication; Represents vector transpose.

[0034] S5. Construct an energy-saving effect prediction network, which includes a local-global dual-channel gating fusion network module, a physically constrained Dirichlet evidence uncertainty quantification module, a dual-driven classification decision-making with differentiable physical rule gating, and a multi-objective loss function calculation. The weighted feature vector is input into the network for prediction. The operation of the energy-saving effect prediction network is as follows: The weighted feature vector is input into the network to predict the energy saving effect. The weighted feature vector is passed through the local-global dual-channel gated fusion network module to extract local features and global features respectively, and then the extracted features are subjected to gated adaptive feature fusion; The fused features are input into the Dirichlet evidence uncertainty quantification module of physical constraints to generate physical constraint residuals that aggregate various thermal constraints. The fused features are then mapped to Dirichlet concentration parameters for each energy-saving level. The uncertainty loss of the network is then calculated based on the physical constraint residuals and the Dirichlet concentration parameters for each energy-saving level. A differentiable physical decision gating strategy is adopted. First, the physical rule confidence level is defined for each energy conservation level based on thermodynamic laws and engineering experience. Then, the predicted Dirichlet concentration parameter is dynamically reconciled with the physical rule confidence level to generate the predicted probability of the energy conservation level. Finally, the loss between the predicted Dirichlet concentration parameter and the physical rule confidence level is calculated. The multi-objective loss function is calculated to obtain the total loss of the network, including the cross entropy loss between the prediction result and the true label, as well as the loss calculated by the two modules in the network, and the weight coefficient is set for each loss.

[0035] S6. Optimize the trainable parameters of the energy-saving effect prediction network through iterative optimization, use Xavier normal distribution to initialize the trainable parameters, use an adaptive moment estimation optimizer to minimize the total loss of the network, and set the convergence condition for stopping iterative training, ultimately obtaining a trained energy-saving effect prediction model network; S7. After physical decoupling normalization and feature adaptive weighting, the newly collected data is input into the trained energy-saving effect prediction network, and the final prediction probability of each energy-saving level is output. The level with the highest probability is taken as the prediction result.

[0036] In a specific implementation, the local-global dual-channel gated fusion network module specifically adopts a dual-branch gated fusion architecture. One branch extracts local physical pattern features through convolution operations, and the other branch learns global features through a fully connected layer. The two are then fused through an adaptive gating mechanism. The specific steps are as follows: (1) Physical feature grouping and input allocation: According to the physical characteristics, the weighted feature vector of each sample is divided into a local feature subset and a global feature subset. The local feature subset contains physical quantities with spatial correlation and the dimension is ; The global feature subset contains non-space-related physical quantities, with dimensions of ; No. The local feature subset of a sample is expressed as , including physical quantities with spatial correlation, such as temperature series, heat flux density distribution, etc. These features usually contain local physical pattern information; No. The global feature subset of samples is , including non-space-related physical quantities, such as material thermal conductivity and heat capacity, which usually reflect the overall thermal properties of the material; is the local feature dimension, is the total dimension of the feature, and ; 2) Extract local spatial feature convolution: Perform a one-dimensional convolution operation on the local feature subset, using the convolution kernel to slide calculations in the spatial dimension to extract the spatially related physical patterns contained in the local features. The convolution calculation results are processed by an activation function to obtain a local feature vector containing local physical pattern information; The calculation formula is as follows: , Where, Represents the local feature vector output by the convolution layer, which represents the feature vector obtained by convolution and contains local physical pattern information; Represents a one-dimensional convolution operation; Represents the local feature convolution kernel parameter matrix; represents the rectified linear unit activation function; 3) Global physical characteristics full connection extraction Perform a fully connected transformation operation on the global feature subset, perform linear combination and nonlinear activation on the global features through the weight matrix and bias vector, learn and generate feature representations that reflect non-spatial physical characteristics, and obtain a global feature vector; The calculation formula is as follows: , Where, The global feature vector output by the fully connected layer represents the feature vector containing global physical mode information; represents the global feature weight matrix; is the global feature weight bias vector; 4) Gated Adaptive Feature Fusion: The local feature vector obtained by convolution processing and the global feature vector obtained by full connection processing are spliced ​​together. Based on the spliced ​​feature vector, a gating vector is generated through the gating weight matrix and activation function. Then, the gating vector is used to perform weighted fusion of the local feature vector and the global feature vector to achieve feature fusion of physical property perception and obtain a fused feature vector.

[0037] The calculation formula is as follows: , , Where, represents the gate vector; Represents feature concatenation operation; represents the gating weight matrix; express Activation function; represents the fused feature vector; It should be noted that when the gating vector is used to perform weighted fusion of the local feature vector and the global feature vector, the position with a large gating vector value represents a large contribution of the local feature, and the position with a small gating vector value represents a large contribution of the global feature, thereby achieving adaptive perception of the degree of feature contribution; it should be noted that the calculation process of the gating vector is implemented using the Sigmoid activation function, which constrains the gate value to (0,1) and directly represents the contribution ratio of the local / global features. For example, It means that local features account for 80% of the weight. It should also be noted that local features are extracted using one-dimensional convolution operations, and global features are extracted using fully connected layers. Different feature types have different extraction methods, and the two extraction methods cannot be interchanged at will. This is because local features have spatial correlation, such as temperature series have spatial correlation, and one-dimensional convolution can capture the patterns of neighboring sensors, while global features have no spatial structure, such as material properties have no spatial structure, and fully connected layers are more suitable for learning abstract thermal properties.

[0038] In a specific implementation, the physically constrained Dirichlet evidence uncertainty quantification module uses a physically constrained Dirichlet evidence quantification method to generate confidence intervals that conform to both the statistical laws of machine learning and the laws of thermodynamics. The specific process is as follows: 1) Constructing thermodynamic residual energy penalty term: A physical constraint violation metric function is defined to quantify the degree to which the prediction results obtained by the energy-saving effect prediction network violate physical laws. Physical laws specifically refer to the laws of thermodynamics. The predicted energy values, such as sensible heat and latent heat, are calculated to see whether they exceed the maximum allowable energy values ​​set based on the thermodynamic limits of the material. If the predicted value exceeds the allowable value, a penalty term is applied, ultimately resulting in a physical constraint residual. The physical constraint residual is the sum of all energy penalty terms that violate the constraint and can be used to penalize predictions that do not conform to physical laws during model network training. The calculation formula is as follows: , Where, represents the physical constraint residual; express function; Indicates the The maximum allowable value of the class energy is set based on the thermodynamic limits of the material; represents the L2 norm; represents the number of thermodynamic constraint categories; Indicates the Predicted values ​​of class energies, for example, predicted energy values ​​of sensible or latent heat, The calculation method is expressed as , Indicates the Class energy conversion vector, Indicates the The bias term for class energy prediction, Indicates the Similar energy conversion vector, specifically constructing a linear transformation vector through the material heat capacity coefficient; Indicates the The bias term for class energy prediction is fitted from the training data; It should be noted that The term is added to the loss function as an L2 norm penalty term to amplify the excess amount twice, forcing the predicted value to return to a reasonable range; it should also be noted that The term represents the physical residual energy term, which is implemented based on a differentiable approximate function. It maintains its derivative during the model network training process and ensures that the gradient can be propagated. Item as The input of the function, when When the term is greater than 0, it grows approximately linearly. When the term is less than 0, the physical residual energy term is close to 0; it should also be noted that in actual implementation, if the energy conversion vector cannot be constructed correctly, it can be set as a learnable parameter vector, and the energy conversion relationship can be fitted by training data, thereby reducing the implementation experience threshold of the technology of the present invention.

[0039] 2) Dirichlet concentration parameter mapping: The fused feature vector is input into the fully connected layer and mapped into a concentration parameter vector of the Dirichlet distribution through an exponential activation function and an addition operation. Each concentration parameter corresponds to an energy-saving level. The Dirichlet distribution is used to describe the network's uncertainty in the probability distribution of samples belonging to each energy-saving level. The size of the concentration parameter reflects the network's confidence in the prediction of each level, thereby achieving parameterized quantification of classification uncertainty. The calculation formula is as follows: , Where, Indicates the The Dirichlet concentration parameter of the energy-saving level reflects the model network's The amount of evidence for each category, The larger the value, the stronger the evidence and the lower the uncertainty; represents the Dirichlet concentration parameter of the first energy-saving level; Indicates the Dirichlet concentration parameter for each energy saving level; represents the evidence weight matrix, which is a trainable parameter; represents the evidence bias vector; Represents a column vector with all elements set to 1, dimension , used to ensure that the Dirichlet distribution parameter is greater than 0, to prevent the Dirichlet distribution from degenerating into extreme determinism; Indicates the total number of energy-saving levels, set , corresponding to 5 levels; 3) Calculate the physical regularization uncertainty loss: Construct an uncertainty loss function for network training, including a Dirichlet-based evidence loss and a physical constraint residual term. The Dirichlet-based evidence loss measures the difference between the probability distribution of network predictions and the probability distribution of true labels, while quantifying the uncertainty of the network itself. The physical constraint residual term penalizes predictions that violate physical laws, thereby jointly optimizing the statistical accuracy and physical rationality of network predictions. The calculation formula is as follows: , Where, represents the uncertainty loss function; The first one-hot encoding of the energy-saving level label elements, when the sample belongs to When level, , otherwise 0; Represents the physical residual weight, set ; express Function in The value at is calculated as ; express Function in The value at is the sum of concentration parameters, and the calculation method is expressed as ; is a logarithmic function with a natural constant as its base; Express The partial derivative of for function.

[0040] It should be noted that the uncertainty loss function can effectively quantify the violation of physical constraints. When the limit is exceeded, the physical constraint residual Increase, resulting in uncertainty loss function Increase, further leading to the model network automatically increasing the total concentration parameter , to improve the overall uncertainty rather than blindly adjust the probability distribution.

[0041] In a specific implementation, the dual-driven classification decision-making of differentiable physical rule gating specifically adopts a differentiable physical decision gating strategy to dynamically reconcile data prediction and physical rules. The specific operations are as follows: 1) Build a thermodynamic rule confidence engine: Based on the laws of thermodynamics and engineering experience, a physical rule confidence level is defined for each energy-saving level. The physical rule confidence level is calculated by a differentiable function, taking into account the measured values ​​of key physical indicators, including thermal inertia index, thermal conductivity, thickness, etc. For each indicator, the measured value is compared with the threshold value preset for each level, and the confidence level is calculated by The function calculates the degree to which the indicator satisfies the rules of each level and obtains the physical rule confidence of each level. The final physical rule confidence is the product of the degree to which all key physical indicators of the sample satisfy the corresponding level rules. A physical rule confidence value close to 1 indicates that the physical indicators of the sample strongly support its belonging to that level, while a value close to 0 indicates a strong violation. The calculation method of physical rule confidence is expressed as: , Where, Indicates the The confidence level of physical rules is ; Indicates the quantity of physical indicators; Indicates the The measured value of a physical indicator, for example, Measured values ​​of physical indicators is the measured value of thermal inertia index. According to the calculation method of thermal inertia index, thermal inertia index = material density × specific heat capacity × thickness, then ; Indicates the Level For example, if the first indicator is thermal inertia index, the threshold value of the first level for the first indicator is , the threshold value of the first level for the second indicator ; for Steepness coefficient, setting ; It should be noted that in the physical rule confidence During the calculation process, The term is used as a cumulative multiplication operation to ensure that all physical indicators meet the threshold at the same time , and the cumulative form requires that all physical indicators meet the threshold at the same time, such as thermal inertia, thermal conductivity, and thickness. In addition, Item adopted The activation function is designed in such a way that the hard threshold Converted into soft decision, maintains the differentiability during the model network training process, and ensures that the gradient can be propagated. If a step function is used, such as , is an indicator function, which is 1 when the condition is met and 0 otherwise. The model network may produce gradient disappearance during training and fail to learn boundary samples close to the threshold.

[0042] 2) Synthetic Data - Rule-gated Probability: The network is fused with the expected probability of the Dirichlet distribution predicted by the fusion feature vector and the confidence of the physical rule, specifically through a learnable gating weight scalar, and the gating weight scalar is obtained by the fusion feature vector through the gating weight vector and The function is calculated, and the relative importance of data prediction and physical rules in the final decision is dynamically controlled through the gating mechanism. The data prediction probability and the physical rule confidence are weighted and summed according to the gating weight scalar to obtain the final prediction probability, and then the prediction result is obtained; The calculation formula is as follows: , , Where, Represents the gate weight vector, dimension and fusion feature vector same; Indicates the The final predicted probability of each energy saving level; Represents the gate weight scalar, with a value range of , control the fusion ratio of data prediction and rule confidence, set ; for The transpose of It should be noted that Item represents the initial prediction probability of the data, and the gate weight scalar Dynamically balance data prediction probability and physical rule confidence ,When the sensor noise is large, it indicates that the data is unreliable, , relying on physical rules to predict classification, when the data is reliable, ,rely on model network to predict classification; 3) Calculate the physical decision boundary loss constraint: The physical decision boundary loss is applied to the expected probability of the Dirichlet distribution predicted by the network. When the confidence of the physical rules of a certain level is lower than the set threshold, indicating that the physical indicators of the sample strongly do not support it belonging to this level, the physical decision boundary loss function will punish the high probability value predicted by the network for this level, forcing the network to increase the uncertainty of its prediction, that is, to reduce the probability value of this level, thereby ensuring that the network will not make overconfident and erroneous predictions when violating the physical rules.

[0043] The calculation formula is as follows: , Where, represents the physical decision boundary loss function.

[0044] In a specific implementation, the total loss function of the energy-saving effect prediction network is calculated using a multi-objective loss function. The specific operations are as follows: Calculate the total loss function of network training. The total loss function is a multi-objective loss function, which consists of three parts: the cross entropy loss function between the predicted results and the actual labels, the uncertainty loss function, and the physical decision boundary loss function. The uncertainty loss function and the physical decision boundary loss function are weighted by weight coefficients respectively. The uncertainty loss weight coefficient is set to a large value close to 1 at the beginning of training. As the training rounds progress, it decreases smoothly to 0 according to the cosine decay function, which can achieve a gradual exit from uncertainty regularization. The physical decision boundary loss weight coefficient is set to 0 at the beginning of training. As the training round approaches a specific round, it smoothly increases to 1 according to the Sigmoid function. This can achieve the delay of physical rule constraints and thus coordinate the effects of different loss terms at different training stages.

[0045] The calculation formula of the total loss function is as follows: , Where, represents the total loss function; represents the cross entropy loss function; Represents the uncertainty loss weight coefficient, and its value range is , control uncertainty loss function In the early stages of training Smooth decay to , In the early stage of representation training, Shi Hengwei , thereby realizing the gradual exit of uncertainty regularization, the calculation method is expressed as ; Indicates taking and The minimum value of At that time , otherwise ; Indicates the current training round, an integer and ; Indicates the starting round of uncertainty loss, set , or set 1 / 5 of the total training rounds; Represents the physical rule loss weight coefficient, with a value range of , controlling the physical decision boundary loss function, close to Shi Cong Smooth growth to , realizing the delay constrained by physical rules, the calculation method is the Sigmoid activation function form, and the calculation method is expressed as ; Indicates the starting round of physical rule loss, set , or set 1 / 2 of the total training rounds; Represents the Sigmoid smoothing coefficient, set .

[0046] It should be noted that the uncertainty loss weight coefficient The calculation method is the design method of the cosine attenuation function, and the physical rule loss weight coefficient The calculation method of is the design method of Sigmoid activation function. The calculation design methods of the two are based on the stability of model network training. The switching between cosine attenuation and Sigmoid activation can ensure smooth training and solve the conflict problem in multi-objective optimization.

[0047] In a specific implementation, the training batch size during energy-saving effect prediction network training is set to 64; the initial learning rate is set to 0.001 when minimizing the total loss function, and the learning rate decays by 10% every 50 rounds; the convergence condition is that the total loss function value changes by less than 0.05 for 20 consecutive rounds, or the maximum number of iterations, such as 1000, is reached, then the iteration is stopped.

[0048] In a specific implementation, the trained network is used to predict the energy-saving level of newly collected building monitoring data. Specifically, physical quantities such as heat flux, temperature difference, thermal conductivity, etc. are collected in the same way as step S1, and constructed into data samples with the same format and dimension as the training samples. The physical decoupling normalization of step S3 and the feature adaptive weighting of step S4 are performed in sequence. The weighted feature vector of the sample is input into the trained energy-saving effect prediction network. After the data is forward propagated through the network, the final prediction probability of the sample corresponding to each energy-saving level is output, and the level with the largest probability is taken as the prediction result. For example, the final prediction probability of the 4th energy-saving level is When , the output level is 4; Simultaneously provide Dirichlet concentration parameters for each energy saving level As a result of uncertainty quantification, when the total concentration parameter A low confidence warning is triggered and manual review is recommended.

[0049] Example 2 like Figure 2 As shown in the figure, the influence of different feature normalization methods on the prediction accuracy of building energy-saving effect is evaluated, and the traditional Min-Max normalization, Z-Score normalization and the physical constraint decoupling normalization method proposed in this invention are compared. Figure 2The middle box shows the accuracy distribution range of the three methods, with the scattered points representing the results of a single experiment. The green box corresponds to the method of our invention. The experimental results show that our method not only achieves significantly higher prediction accuracy than traditional methods, but also exhibits smaller fluctuations and greater stability. This demonstrates the advantage of physical constraint decoupling normalization in eliminating dimensional differences and nonlinear coupling through covariance feature decomposition, effectively resolving the distorted feature distribution caused by traditional methods and enabling the network to more accurately capture the essential characteristics of building thermal performance. It should be noted that the horizontal and vertical coordinates of this experimental graph are dimensionless values ​​without physical units.

[0050] As shown in Table 1, the heat flux, surface temperature difference, and thermal conductivity time series data collected by actual building envelope sensors are used as input to compare and analyze the three normalization methods: minimum-maximum normalization, Z-Score normalization, and the method of the present invention. Table 1 Experimental configuration table The experiment sets three analysis dimensions, including the ability to preserve the time series morphology, the degree of preservation of correlation between physical quantities, and the consistency of feature distribution, such as Figure 3 As shown in the experimental results, based on time series comparison, it can be found that the original heat flux density (blue curve) exhibits typical diurnal periodic fluctuations, with a peak in the afternoon and a trough in the early morning. Although the minimum-maximum normalization (red dashed line) retains the basic shape, it amplifies the subtle fluctuations at night, resulting in significant distortion. Z-score normalization (green dotted line) shifts the overall value downward, resulting in negative values, which destroys the physical meaning (heat flux density should be positive). However, this method (purple solid line) perfectly preserves the phase and morphological characteristics of the original curve while suppressing noise interference. It should be noted that the vertical axis of this experimental figure is dimensionless and has no physical units, while the horizontal axis is in hours.

[0051] In the correlation analysis experiment, the correlation coefficient between heat flux and surface temperature difference was calculated. The specific experimental data are shown in Table 2. Table 2 Comparison of correlation coefficients of different methods The experimental results shown in Table 2 indicate that the correlation retention rate of this method is close to 100%, proving that the thermodynamic correlation between physical quantities can be effectively preserved.

[0052] In addition, the normalized feature distribution comparison experiment Figure 4 The results show that the traditional method leads to the distortion of distribution shape, the minimum-maximum method produces a double peak, and the Z-Score normalization method leads to the broadening of the distribution. The distribution curve of this method is highly consistent with the original data, and the kurtosis and skewness are kept optimal, proving that the scale retention ability has advantages.

[0053] Example 3 like Figure 5 As shown in Figure 2, the performance responses of the three feature weighting strategies when the sharpening factor changes are analyzed, and the equal weight, mutual information weighting and the physical attention mechanism of the present invention are compared. Figure 5 The middle curve shows that the proposed method (green) maintains the best and stable performance under different sharpening factors, especially in the key area (sharpening factor =1.5-2.5) is the most prominent, validating the dual advantages of the attention mechanism guided by physical priors. On the one hand, it quantifies the statistical correlation between features and labels through mutual information, and on the other hand, it incorporates prior knowledge of physical contributions such as thermal inertia metrics. This overcomes the shortcomings of traditional methods that discard weakly correlated features or treat them equally, enabling the network to adaptively enhance the representation of key physical mechanisms such as thermal conductivity. It should be noted that the horizontal and vertical coordinates in this experimental figure are dimensionless values ​​without physical units.

[0054] Example 4 like Figure 6 As shown in Figure 2, the regulation effect of physical constraint mechanism on prediction uncertainty is analyzed, and the performance of the model without physical constraint and the method of the present invention are compared on 100 test samples. Figure 6 The middle line reflects the prediction uncertainty indicator (Dirichlet total concentration parameter), and the green area shows the extent of uncertainty reduction. The uncertainty curve of the proposed method is consistently lower than that of the traditional method, and the difference is more significant in the sample segment with high sensor noise (such as samples 30-50). This proves that the physically constrained Dirichlet evidence module effectively constrains the prediction results within the physically feasible domain through the thermodynamic residual energy penalty term, resolving the potential violation of thermodynamic laws by traditional Bayesian methods and enabling the model to maintain reasonable uncertainty quantification even in a noisy data environment. It should be noted that the horizontal and vertical coordinates in this experimental figure are dimensionless values ​​without physical units.

[0055] Example 5 like Figure 7 As shown in Figure 2, the prediction performance of the traditional neural network, the basic physical constraint model and the complete method of the present invention at five energy-saving levels is compared. Figure 7 The height of the middle column represents the average accuracy at each level, and the scattered points represent the results of different test samples. The proposed method (green column) maintains the highest accuracy across all energy-saving levels, with a particularly significant advantage in the critical high-level region (levels 4-5). This demonstrates the advantages of the proposed dual-channel gated fusion network. Local feature convolution extracts spatial patterns such as temperature field distribution, while global features are fully connected to learn overall properties such as material thermal inertia. Physically-aware feature fusion is then achieved through an adaptive gating mechanism, fully capturing the complex thermal-engineering coupling mechanisms of building energy conservation and overcoming the feature loss problem of single-channel models. It should be noted that the horizontal and vertical coordinates in this experimental figure are dimensionless values ​​without physical units.

[0056] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting energy-saving effects of building envelope structures, characterized by: The following steps are involved: S1. Deploy various types of sensors at key locations of building maintenance structures to collect real-time data on physical quantities directly related to building thermal performance, and construct a data set based on the collected data; S2. Based on building thermal engineering principles and energy-saving evaluation standards, the collected sensor data is double-labeled, including physical indicator labeling and energy-saving level labeling. The labeled labels are mapped one-to-one with the collected data to form a training set. S3. Apply the physical constraint decoupling normalization method to the multi-dimensional monitoring data in the training set, construct a decoupling matrix through covariance feature decomposition, and generate decoupled normalized features; S4. Adopting the attention mechanism guided by physical priors, we decouple the correlation between normalized features and labels through mutual information quantification, and generate adaptive weights based on physical contribution priors to generate weighted feature vectors. S5. Construct an energy-saving effect prediction network, which includes a local-global dual-channel gating fusion network module, a physically constrained Dirichlet evidence uncertainty quantification module, a dual-driven classification decision-making with differentiable physical rule gating, and a multi-objective loss function calculation. The weighted feature vector is input into the network for prediction. S6. Optimizing the trainable parameters of the energy-saving effect prediction network through iterative optimization, and setting a convergence condition for stopping iterative training to obtain a trained energy-saving effect prediction network; S7. After physical decoupling normalization and feature adaptive weighting, the newly collected data is input into the trained energy-saving effect prediction network, and the final prediction probability of each energy-saving level is output. The level with the highest probability is taken as the prediction result.

2. A method for predicting energy-saving effects of building envelope structures according to claim 1, characterized in that: The data collection process in S1 is as follows: Data collection adopts a distributed Internet of Things architecture. Sensors continuously upload data to the central server in a cycle of 1-10 minutes, and synchronize the multi-dimensional data through timestamp alignment to form a full-dimensional original monitoring data set of multi-physical mechanisms.

3. The method for predicting energy-saving effects of building envelope structures according to claim 2, wherein the double marking in S2 is as follows: Physical indicator labeling: divide data dimension labels according to sensor types; Energy-saving grade marking: manual marking is used to assess the energy-saving grade of the enclosure structure performance in each monitoring period; The physical indicator labels and energy-saving grade labels are stored in a structured form and mapped one-to-one with the collected data to form a training set.

4. A method for predicting energy-saving effects of building envelope structures according to claim 3, characterized in that: The physical constraint decoupling and normalization method in S3 is as follows: S3.

1. Calculate the physical empirical mean and dimensionless standard deviation: Based on all samples in the historical data set of the building maintenance structure, the physical empirical mean and dimensional standard deviation of the physical quantity in each dimension are calculated. The reference center position of each physical quantity is determined based on the physical empirical mean vector, and the typical range of the numerical fluctuation of each physical quantity itself is determined based on the dimensional standard deviation vector. S3.

2. Covariance matrix construction and eigendecomposition: The historical data set is centralized using the physical empirical mean vector. The covariance matrix between physical quantities in each dimension is then calculated based on the centralized data. The degree of correlation between different physical quantities is then quantified through the covariance matrix. S3.

3. Solve the physical decoupling matrix: Perform eigendecomposition on the covariance matrix to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The column vectors of the eigenvector matrix represent the direction of change of the principal components, and the eigenvalue diagonal matrix represents the size of the variance contribution in the direction of change of each principal component. Based on the eigenvector matrix and the eigenvalue diagonal matrix, a physical decoupling matrix is ​​constructed to eliminate the coupling relationship between different physical quantities caused by dimensional differences and nonlinear interaction effects, and then the physical decoupling matrix is ​​obtained; S3.

4. Perform decoupled normalization calculation: The data in the historical data set are normalized by physical dimensions to eliminate the influence of the dimensions and scale differences of different physical quantities. Then, the normalized eigenvector is multiplied by the physical decoupling matrix to obtain the decoupled eigenvector.

5. A method for predicting energy-saving effects of building envelope structures according to claim 4, characterized in that: The adaptive feature weighting process guided by physical priors in S4 is as follows: S4.

1. Calculation of feature-label mutual information: For each decoupled feature vector dimension after decoupling normalization, calculate the mutual information value between the feature and the building energy efficiency grade label; S4.

2. Constructing physical attention weights: The mutual information value of each feature dimension and the label is integrated, as well as the preset physical contribution prior value. The concentration of the weight distribution is controlled by the sharpening factor. The fusion result is passed through the exponential function and normalized to generate an attention weight value between 0 and 1 for each feature dimension. S4.

3. Perform feature adaptive weighting: The attention weight vector is multiplied element-by-element by the decoupled feature vector. According to the attention weight value of each feature, the importance of physical mechanisms and statistically strongly correlated features is amplified, while the influence of weakly correlated features is weakened to obtain a weighted feature vector.

6. A method for predicting energy-saving effects of building envelope structures according to claim 5, characterized in that: The operation of the energy-saving effect prediction network in S5 is as follows: The weighted feature vector is input into the network to predict the energy saving effect. The weighted feature vector is passed through the local-global dual-channel gated fusion network module to extract local features and global features respectively, and then the extracted features are subjected to gated adaptive feature fusion; The fused features are input into the Dirichlet evidence uncertainty quantification module of physical constraints to generate physical constraint residuals that aggregate various thermal constraints. The fused features are then mapped to Dirichlet concentration parameters for each energy-saving level. The uncertainty loss of the network is then calculated based on the physical constraint residuals and the Dirichlet concentration parameters for each energy-saving level. A differentiable physical decision gating strategy is adopted. First, the physical rule confidence level is defined for each energy conservation level based on thermodynamic laws and engineering experience. Then, the predicted Dirichlet concentration parameter is dynamically reconciled with the physical rule confidence level to generate the predicted probability of the energy conservation level. Finally, the loss between the predicted Dirichlet concentration parameter and the physical rule confidence level is calculated. The multi-objective loss function is calculated to obtain the total loss of the network, including the cross entropy loss between the prediction result and the true label, as well as the loss calculated by the two modules in the network, and the weight coefficient is set for each loss.

7. A method for predicting energy-saving effects of building envelope structures according to claim 6, characterized in that: The global dual-channel gated fusion network module adopts a dual-branch gated fusion architecture. One branch extracts local physical pattern features through convolution operations, while the other branch learns global features through a fully connected layer. The two are then fused through an adaptive gating mechanism. The specific steps are as follows: (1) Physical feature grouping and input allocation: According to the physical characteristics, the weighted feature vector of each sample is divided into a local feature subset and a global feature subset. The local feature subset contains physical quantities with spatial correlation and the dimension is ; The global feature subset contains non-space-related physical quantities, with dimensions of ; 2) Extract local spatial feature convolution: Perform a one-dimensional convolution operation on the local feature subset, using the convolution kernel to slide calculations in the spatial dimension to extract the spatially related physical patterns contained in the local features. The convolution calculation results are processed by an activation function to obtain a local feature vector containing local physical pattern information; 3) Fully connected extraction of global physical characteristics Perform a fully connected transformation operation on the global feature subset, perform linear combination and nonlinear activation on the global features through the weight matrix and bias vector to obtain the global feature vector; 4) Gated Adaptive Feature Fusion: The local feature vector and the global feature vector are concatenated, and a gating vector is generated through a gating weight matrix and an activation function. Then, the gating vector is used to perform weighted fusion of the local feature vector and the global feature vector to obtain a fused feature vector.

8. A method for predicting energy-saving effects of building envelope structures according to claim 7, characterized in that: The specific operations of the physical constraint Dirichlet evidence uncertainty quantification module are as follows: 1) Constructing thermodynamic residual energy penalty term: A physical constraint violation metric function is defined to quantify the extent to which the prediction results obtained by the energy-saving effect prediction network violate physical laws. Physical laws specifically refer to the laws of thermodynamics. The function calculates whether the predicted energy values ​​exceed the maximum allowable energy values ​​set based on the thermodynamic limits of the materials. If the predicted value exceeds the allowable value, the physical constraint residual is obtained by summing up all the energy penalty terms that violate the constraints. 2) Dirichlet concentration parameter mapping: The fused feature vector is input into the fully connected layer and mapped into a concentration parameter vector of the Dirichlet distribution through an exponential activation function and an addition operation. Each concentration parameter corresponds to an energy-saving level. The Dirichlet distribution describes the network's uncertainty in the probability distribution of the sample belonging to each energy-saving level, and the size of the concentration parameter reflects the network's confidence in the prediction of each level. 3) Calculate the physical regularization uncertainty loss: The uncertainty loss function for network training is constructed based on the evidence loss part of Dirichlet distribution and the physical constraint residual term.

9. A method for predicting energy-saving effects of building envelope structures according to claim 8, characterized in that: The specific operations of the dual-drive classification decision-making with differentiable physical rule gating are as follows: 1) Build a thermodynamic rule confidence engine: Based on the laws of thermodynamics and engineering experience, the physical rule confidence is defined for each energy-saving level, and the physical rule confidence is calculated by differentiable functions. Specifically, for each indicator, the measured value is compared with the threshold value preset for each level, and the confidence is calculated by The function calculates the degree to which the indicator satisfies the rules of each level and obtains the confidence of the physical rules of each level; 2) Synthetic Data - Rule-gated Probability: The network is fused with the expected probability of the Dirichlet distribution predicted by the fusion feature vector and the confidence of the physical rule, specifically through a learnable gating weight scalar, and the gating weight scalar is obtained by the fusion feature vector through the gating weight vector and The function is calculated, and the relative importance of data prediction and physical rules in the final decision is dynamically controlled through the gating mechanism. The data prediction probability and the physical rule confidence are weighted and summed according to the gating weight scalar to obtain the final prediction probability, and then the prediction result is obtained; 3) Calculate the physical decision boundary loss constraint: A physical decision boundary loss is applied to the Dirichlet distribution expected probability predicted by the network.

10. A method for predicting energy-saving effects of building envelope structures according to claim 9, characterized in that: The total loss function of the energy-saving effect prediction network is calculated through a multi-objective loss function. The specific operations are as follows: Calculate the total loss function of network training. The total loss function is a multi-objective loss function, which consists of three parts: the cross entropy loss function between the predicted results and the actual labels, the uncertainty loss function, and the physical decision boundary loss function. The uncertainty loss function and the physical decision boundary loss function are weighted separately.

Images