A method for predicting energy-saving effect of a building envelope

Abstract

The present application relates to the technical field of energy consumption prediction, and particularly relates to a building envelope energy saving effect prediction method, which specifically comprises the following steps: deploying sensors on a building maintenance structure to collect relevant data and constructing a data set; labeling the collected data to form a training set; using a physical constraint decoupling normalization method to generate decoupled and normalized features from the data in the training set; using a physical prior guided attention mechanism to generate a weighted feature vector; constructing an energy saving effect prediction network, inputting the weighted feature vector into the network for prediction; optimizing the network to obtain a trained network; inputting newly collected data into the trained network after processing, outputting final prediction probabilities of each energy saving level, and taking the level with the maximum probability as the prediction result. The present application can overcome the shortcomings of inaccurate data processing and unreasonable feature selection by preprocessing the collected data before inputting it into the network, thereby improving the accuracy of the prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy consumption prediction, and particularly relates to a building envelope energy-saving effect prediction method. BACKGROUND

[0002] With the continuous improvement of building energy efficiency requirements, how to accurately predict the energy-saving effect of building envelope has become an important issue in building design and energy-saving optimization. The envelope plays a crucial role in building thermal performance, and its thermal conduction, convection and radiation directly affect the energy consumption of the building. In order to improve the energy efficiency of buildings and meet the energy-saving standards, it is necessary to accurately monitor and evaluate the thermal performance of the envelope. However, the existing building energy-saving effect prediction methods generally face some challenges, such as inaccurate data processing, unreasonable feature selection, inability of the model network to fully capture the complexity of the physical mechanism, and the lack of uncertainty quantification.

[0003] The Chinese invention patent with publication number CN120181334A proposes a building construction site energy prediction and supervision method and system, which belongs to the technical field of building construction site. By combining construction plans, mechanical shift quantities and weather prediction data, the energy consumption of the building construction site can be accurately predicted, providing a scientific basis for construction management. The energy consumption data and weather information are also statistically analyzed to dynamically adjust the energy distribution scheme, ensure the priority supply of energy for key equipment, and improve energy utilization efficiency. The historical energy consumption data is analyzed in combination with weather prediction to accurately predict future energy consumption and improve the intelligent level of energy management. This invention can adapt to different weather conditions and construction stages, has strong universality and applicability, and can ensure the normal operation of key equipment by reasonably allocating energy, improve construction efficiency, and shorten the project duration.

[0004] However, the prior art still has the following problems to be further solved: the prior art usually adopts a simple linear normalization method, which may cause distortion of the feature distribution, destroy the physical correlation of the data, and affect the accurate prediction of the energy saving effect when dealing with the dimensional and scale differences of different physical quantities; the prior art often uses traditional feature selection methods, which directly discard weakly correlated features or use equal weight processing for all features, which easily loses weak physical signals and cannot effectively highlight key physical features closely related to energy saving effect, resulting in reduced model network performance; traditional neural networks usually rely on single-channel architecture and cannot simultaneously capture local physical patterns and global physical characteristics, thereby limiting the comprehensive learning and prediction ability of the model network on the complex thermal characteristics of building envelopes; the prior art mostly ignores uncertainty quantification in the prediction process, and the conventional Bayesian method fails to consider physical constraints such as thermodynamics, which may lead to prediction results that violate the laws of thermodynamics, especially in the case of large sensor noise, the prediction of traditional methods is unreliable and cannot meet the needs of practical applications.

[0005] Therefore, the present application proposes a building envelope energy saving effect prediction method to solve the above problems. SUMMARY

[0006] The present application is directed to the deficiencies of the prior art, and a building envelope energy saving effect prediction method is developed. The present application can overcome the problems of inaccurate data processing, unreasonable feature selection, inability to fully capture the complexity of physical mechanisms, and deficiencies in uncertainty quantification by preprocessing the collected data before inputting it into the network, thereby improving the accuracy of the prediction results.

[0007] The technical scheme for solving the technical problems of the present application is a building envelope energy saving effect prediction method, comprising the following steps:

[0008] S1, deploying multiple types of sensors at key positions of the building envelope, collecting physical quantity data related to the thermal performance of the building in real time, and constructing a data set according to the collected data;

[0009] S2, based on the principles of building thermotics and energy saving evaluation standards, double-labeling the collected sensor data, including physical index labeling and energy saving level labeling, mapping the labeled labels and collected data one by one to form a training set;

[0010] S3, using a physical constraint decoupling normalization method for the multi-dimensional monitoring data in the training set, constructing a decoupling matrix through covariance characteristic decomposition to eliminate dimensional differences and decouple nonlinear coupling between physical features, and generating decoupled and normalized features;

[0011] 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.

[0012] 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.

[0013] 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;

[0014] 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.

[0015] The data collection process in S1 is as follows:

[0016] 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.

[0017] The double annotation in S2 is as follows:

[0018] Physical indicator labeling: divide data dimension labels according to sensor types;

[0019] 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;

[0020] Divide into five levels from 1 to 5, 1 represents the lowest level and 5 represents the highest level;

[0021] 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;

[0022] 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.

[0023] The physical constraint decoupling and normalization method in S3 is as follows:

[0024] S3.1, calculate the physical empirical mean and dimensional standard deviation:

[0025] Based on all samples in the building maintenance structure historical data set, for each dimension of the physical quantity, calculate its physical empirical mean and dimensional standard deviation, determine the reference center position of each physical quantity according to the physical empirical mean vector, and determine the typical range of the numerical fluctuation of each physical quantity itself according to the dimensional standard deviation vector;

[0026] S3.2, covariance matrix construction and eigenvalue decomposition:

[0027] Using the physical empirical mean vector, the historical data set is centered, and then the covariance matrix between the physical quantities of each dimension is calculated based on the centered data, and then the degree of correlation between different physical quantities is quantified through the covariance matrix;

[0028] S3.3, solve the physical decoupling matrix:

[0029] The eigenvalue decomposition operation is performed on the covariance matrix to obtain the eigenvector matrix and the eigenvalue diagonal matrix, and the column vectors of the eigenvector matrix represent the change direction of the principal component, and the eigenvalue diagonal matrix represents the size of the variance contribution in the change direction of each principal component;

[0030] 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 due to the dimensional difference and nonlinear interaction effect, so that the transformed features are statistically independent, and then the physical decoupling matrix is obtained;

[0031] S3.4, perform decoupling normalization calculation:

[0032] The data in the historical data set is subjected to physical dimensional normalization processing to eliminate the influence of the dimensional and scale differences of different physical quantities, and then the standardized eigenvector is multiplied by the physical decoupling matrix to obtain the decoupling eigenvector.

[0033] The adaptive feature weighting process guided by the physical prior in S4 is as follows:

[0034] S4.1, feature-label mutual information calculation:

[0035] For each decoupled feature vector dimension after decoupling normalization, the mutual information value between the feature and the building energy saving grade label is calculated;

[0036] S4.2, construct physical attention weight:

[0037] The mutual information value of each feature dimension and label is fused with a preset physical contribution degree prior value, a sharpening factor is used to control the concentration degree of weight distribution, and an attention weight value between 0 and 1 is generated for each feature dimension through an exponential function and normalization operation of the fusion result;

[0038] S4.3, performing feature adaptive weighting:

[0039] The attention weight vector is multiplied element by element with the decoupled feature vector, and according to the attention weight value of each feature, the importance of the physical mechanism and the statistically strongly correlated features is amplified, and the influence of the weakly correlated features is weakened, to obtain a weighted feature vector.

[0040] The operation of the energy saving effect prediction network in S5 is as follows:

[0041] The weighted feature vector is input into the network for energy saving effect prediction, and the weighted feature vector is input into the local-global dual-channel gated fusion network module to extract local features and global features, respectively, and then the extracted features are adaptively fused through gating.

[0042] The fused features are input into the physical constraint Dirichlet evidence uncertainty quantification module to generate physical constraint residuals that collect various thermal constraints, and map the fused features to Dirichlet concentration parameters of each energy saving level, and then calculate the uncertainty loss of the network according to the physical constraint residuals and the Dirichlet concentration parameters of each energy saving level.

[0043] A differentiable physical decision gating strategy is adopted, physical rule confidence is defined for each energy saving level based on the laws of thermodynamics and engineering experience, then the predicted Dirichlet concentration parameters and the physical rule confidence are dynamically reconciled to generate the prediction probability of the energy saving level, and finally the loss between the predicted Dirichlet concentration parameters and the physical rule confidence is calculated.

[0044] The total loss of the network is calculated by a multi-objective loss function, including the cross-entropy loss of the prediction result and the true label, and the loss calculated by the two modules in the network, and a weight coefficient is set for each loss.

[0045] 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 operation, and the other branch learns global characteristics through fully connected layer, and the two are fused through adaptive gating mechanism, the specific steps are as follows:

[0046] (1) Physical feature grouping and input distribution:

[0047] 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-spatially correlated physical quantities, and the dimension is ;

[0048] 2) Extract local spatial features convolution:

[0049] Perform one-dimensional convolution operation on the local feature subset, use the convolution kernel to slide in the spatial dimension to calculate, extract the spatially correlated physical patterns contained in the local features, and the convolution calculation result is processed by the activation function to obtain the local feature vector containing the local physical pattern information;

[0050] 3) Global physical property fully connected extraction

[0051] Perform a fully connected transformation operation on the global feature subset, linearly combine and nonlinearly activate the global features through the weight matrix and bias vector, learn and generate feature representations reflecting non-spatial physical properties, and obtain the global feature vector;

[0052] 4) Gating adaptive feature fusion:

[0053] Concatenate the local feature vector obtained by convolution and the global feature vector obtained by fully connected processing, based on the concatenated feature vector, generate a gating vector through a gating weight matrix and an activation function, then use the gating vector to weight and fuse the local feature vector and the global feature vector, realize feature fusion of physical property perception, and obtain a fused feature vector.

[0054] The Dirichlet evidence uncertainty quantification module with physical constraints adopts a Dirichlet evidence quantification method with physical constraints to generate a confidence interval that meets both the statistical rules of machine learning and the laws of thermodynamics. The specific process is as follows:

[0055] 1) Constructing a thermodynamic residual energy penalty term:

[0056] Define a physical constraint violation metric function to quantify the degree to which the predicted results obtained by the energy-saving effect prediction network violate physical laws, where the physical laws refer to the laws of thermodynamics. Calculate whether the predicted energy values exceed the maximum allowed energy values based on the material thermodynamic limits. If the predicted values exceed the allowed values, a penalty term is applied, and the physical constraint residual is finally obtained. The physical constraint residual is the sum of all energy penalty terms that violate the constraints;

[0057] 2) Dirichlet concentration parameter mapping:

[0058] The fusion feature vector is input to a fully connected layer and is mapped to a concentration parameter vector of Dirichlet distribution through an exponential activation function and a plus 1 operation, each concentration parameter corresponds to an energy saving level, the uncertainty of the network to the probability distribution of the sample belonging to each energy saving level is described through the Dirichlet distribution, and the confidence degree of the network to each level is reflected through the size of the concentration parameter, and then the parameterization and quantization of the classification uncertainty are realized;

[0059] 3) Calculate the physical regularization uncertainty loss:

[0060] The uncertainty loss function of network training is constructed, including the evidence loss part based on Dirichlet distribution and the physical constraint residual term, the evidence loss part based on Dirichlet distribution measures the difference between the probability distribution predicted by the network and the true label probability distribution, and quantifies the uncertainty of the network itself, and the physical constraint residual term punishes the prediction that violates the physical law;

[0061] The differentiable physical rule gated double driving classification decision specifically adopts a differentiable physical decision gating strategy to dynamically reconcile data prediction and physical rules, and the specific operation is as follows:

[0062] 1) Build a thermodynamic rule confidence engine:

[0063] Based on the laws of thermodynamics and engineering experience, define the physical rule confidence for each energy saving level, the physical rule confidence is calculated through a differentiable function, considering the measured value of the key physical indicators, for each indicator, compare its measured value with the preset threshold value for each level, and through The function calculates the degree to which the index meets the rules of each level to obtain 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 meet the corresponding level rules, and the physical rule confidence value close to 1 indicates that the physical indicators of the sample strongly support it to belong to this level, and close to 0 indicates strong violation;

[0064] 2) Synthesize data-rule gating probability:

[0065] Fuse the Dirichlet distribution expected probability predicted by the network based on the fusion feature vector with the physical rule confidence, specifically through a learnable gating weight scalar for fusion, and the gating weight scalar is calculated by the fusion feature vector through a gating weight vector and The function calculates the degree to which the index meets the rules of each level to obtain 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 meet the corresponding level rules, and the physical rule confidence value close to 1 indicates that the physical indicators of the sample strongly support it to belong to this level, and close to 0 indicates strong violation;

[0064] 2) Synthesize data-rule gating probability:

[0065] Fuse the Dirichlet distribution expected probability predicted by the network based on the fusion feature vector with the physical rule confidence, specifically through a learnable gating weight scalar for fusion, and the gating weight scalar is calculated by the fusion feature vector through a gating weight vector and The function calculates the degree to which the index meets the rules of each level to obtain 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 meet the corresponding level rules, and the physical rule confidence value close to 1 indicates that the physical indicators of the sample strongly support it to belong to this level, and close to 0 indicates strong violation;

[0064] 2) Synthesize data-rule gating probability:

[0065] Fuse the Dirichlet distribution expected probability predicted by the network based on the fusion feature vector with the physical rule confidence, specifically through a learnable gating weight scalar for fusion, and the gating weight scalar is calculated by the fusion feature vector through a gating weight vector and The function calculates the degree to which the index meets the rules of each level to obtain 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 meet the corresponding level rules, and the physical rule confidence value close to 1 indicates that the physical indicators of the sample strongly support it to belong to this level, and close to 0 indicates strong violation;

[0066] 3) Calculate the physical decision boundary loss constraint:

[0067] The physical decision boundary loss is used to act on the expected probability of the Dirichlet distribution of network prediction. When the confidence of a certain level of physical rule is lower than the set threshold, it indicates that the physical index of the sample strongly supports it to belong to this level. The physical decision boundary loss function will punish the high probability value of the network prediction of this level, forcing the network to increase its uncertainty, that is, to reduce the probability value of this level, and thus to ensure that the network will not make overconfident wrong predictions when violating the physical rules.

[0068] The total loss function of the energy saving effect prediction network is calculated by the multi-objective loss function, and the specific operation is as follows:

[0069] The total loss function of network training is calculated, and the total loss function is a multi-objective loss function, which is composed of three parts, including the cross-entropy loss function between the prediction result and the actual label, 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;

[0070] The uncertainty loss weight coefficient is set to a larger value close to 1 in the early stage of training, and is smoothly reduced to 0 according to the cosine decay function as the training round progresses;

[0071] The physical decision boundary loss weight coefficient is set to 0 in the early stage of training, and is smoothly increased to 1 according to the Sigmoid function when the training round approaches a certain round.

[0072] The effects provided in the summary of the invention are only the effects of the embodiments, not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0073] The building sensor monitoring data collected by the present application is multi-dimensional data with multi-physical dimension coupling characteristics and different dimensions. Therefore, the physical constraint decoupling and normalization method is adopted, the non-linear interaction effect between different physical quantities is eliminated through covariance characteristic decomposition, so that the multi-physical quantity data can be processed unbiasedly in the same feature space, and the traditional linear normalization method can be avoided. Linear scaling, ignoring the non-linear interaction effect between physical quantities, resulting in feature distribution distortion and physical correlation damage;

[0074] Since the decoupled and normalized features still have a significant problem of correlation difference with the energy saving level, the present application adopts a physical prior guided adaptive feature weighting mechanism, calculates the mutual information between the features and the energy saving level, and combines the physical contribution degree prior to automatically assign weights to each feature. It can ensure that the model can better highlight the key physical features related to the energy saving effect, and overcome the problems caused by ignoring physical weak signals or equal weight processing in traditional methods;

[0075] The application innovatively adopts a double-channel gated fusion network structure, one branch extracts local features through convolution operation, the other branch extracts global features through a fully connected layer, and then the two are fused through an adaptive gating mechanism, so that the complex thermal performance of building envelope can be fully captured, the feature loss problem of single-channel model can be avoided, and the prediction accuracy is improved.

[0076] In order to solve the problems of sensor noise and model prediction uncertainty, the application adopts a Dirichlet evidence uncertainty quantification method with physical constraints, maps the thermodynamic residual energy penalty term and the Dirichlet concentration parameter, generates a confidence interval conforming to the thermodynamic law, and effectively quantifies the uncertainty of prediction, so that the problem that the traditional method may violate the law of thermodynamics can be avoided.

[0077] In summary, the application can overcome the problems of inaccurate data processing, unreasonable feature selection, inability to fully capture the complexity of physical mechanism, and insufficient uncertainty quantification, and improve the accuracy of the prediction result. BRIEF DESCRIPTION OF DRAWINGS

[0078] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application.

[0079] Figure 1 The figure is a method flowchart of the application.

[0080] Figure 2 The figure is a prediction accuracy comparison chart of different normalization methods.

[0081] Figure 3 The figure is an effect comparison chart of different normalization methods according to heat flux density.

[0082] Figure 4 The figure is a feature distribution comparison chart after normalization of different methods.

[0083] Figure 5 The figure is a performance comparison chart of different feature weighting strategies.

[0084] Figure 6 The figure is a comparison chart of the influence of physical constraints on prediction uncertainty.

[0085] Figure 7 The figure is a prediction accuracy bar chart of different methods in each energy-saving grade. DETAILED DESCRIPTION

[0086] For the purpose of clearly illustrating the technical features of the present application, the present application will be described in detail below with reference to specific embodiments and drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present application. In order to simplify the disclosure of the present application, the components and settings of specific examples are described below.

[0087] Embodiment 1

[0088] An energy-saving effect prediction method for a building envelope is as follows:

[0089] S1, deploying multiple types of sensors at key positions of the building envelope to collect physical quantity data directly related to the thermal performance of the building in real time, and constructing a data set according to the collected data;

[0090] The physical quantity data directly related to the thermal performance of the building includes heat flow data, temperature data, material physical property data, and environmental parameters.

[0091] Heat flow data: using heat flow sensors to continuously monitor the dynamic changes of heat flow density on the inner and outer surfaces of the building envelope, recording the heat transfer value per unit area per unit time, with the unit being "W / m²";

[0092] Temperature data: arranging high-precision temperature sensors at the inner and outer surfaces of the building envelope, interlayers, and material interfaces to synchronously collect time series data of surface temperature difference, material internal temperature gradient, and environmental temperature, with the unit being "℃";

[0093] Material physical property data: obtaining the real-time thermal conductivity of the wall and thermal insulation material through embedded thermal conductivity sensors or laboratory calibration devices, with the unit being "W / m·K", and recording the thickness of the material structure layer using a thickness measuring instrument, with the unit being "m";

[0094] Environmental parameters: integrating humidity, wind speed, and solar radiation sensors to supplement the collection of environmental relative humidity (unit: "%"), near-surface wind speed (unit: "m / s"), and solar radiation intensity (unit: "W / m²") to quantify the influence of external climate on heat transfer;

[0095] In the specific embodiment, the data collection adopts a distributed Internet of Things architecture, and the sensors continuously upload to the central server at a cycle of 1-10 minutes, and the multi-dimensional data are synchronously fused through time stamp alignment to form a full-dimensional original monitoring data set of multiple physical mechanisms;

[0096] The multiple physical mechanisms include heat conduction, convection, and radiation.

[0097] S2, based on the principle of building thermal engineering and energy-saving evaluation standard, the collected sensor data is double-labeled, including physical index labeling and energy-saving level labeling, the labeled labels are one-to-one mapped with the collected data, and a training set is formed;

[0098] In the specific implementation, the double labeling is as follows:

[0099] Physical index labeling: according to the type of sensor, the data dimension label is divided;

[0100] The data includes heat flow density, temperature difference sequence, thermal conductivity, thickness and other multi-dimensional physical quantities;

[0101] Energy-saving level labeling: according to the national standard of "Public Building Energy-saving Design Standard" (GB50189-2015), or according to the measured energy consumption data, the energy-saving level of the building envelope performance of each monitoring period is evaluated by manual labeling;

[0102] Divide into 1-5 levels, 1 level represents the lowest level, and 5 represents the highest level;

[0103] The labeling basis of energy-saving level labeling includes unit area heat loss, thermal inertia index, and energy consumption deviation from the benchmark model network;

[0104] The physical index labeling and energy-saving level labeling are stored in a structured form, and are one-to-one mapped with the collected data to form a training set.

[0105] S3, the multi-dimensional monitoring data in the training set is normalized by using a physical constraint decoupling method, a decoupling matrix is constructed by covariance characteristic decomposition, dimensional differences are eliminated and nonlinear coupling between physical characteristics is eliminated, and decoupled normalized features are generated;

[0106] S3.1, calculate the physical empirical mean and dimensional standard deviation:

[0107] Based on all samples in the historical data set of the building maintenance structure, for each dimension of physical quantity, the physical empirical mean and dimensional standard deviation are calculated, the reference center position of each physical quantity is determined according to the physical empirical mean vector, and the typical range of the numerical value fluctuation of each physical quantity is determined according to the dimensional standard deviation vector;

[0108] The calculation formula is as follows:

[0109] ,

[0110] ,

[0111] In the formula, N represents the total number of samples in the historical data set; The first sample in the 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;

[0112] 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;

[0113] 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.

[0114] S3.2. Covariance matrix construction and eigendecomposition:

[0115] 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.

[0116] The calculation formula is as follows:

[0117] ,

[0118] 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;

[0119] It should be noted that Term characterization performs a central processing operation on the raw sensor data and eliminates the baseline offset of each physical quantity by subtracting the mean of the corresponding physical quantity;

[0120] S3.3. Solve the physical decoupling matrix:

[0121] The covariance matrix is subjected to eigen decomposition operation to obtain an eigenvector matrix and an eigenvalue diagonal matrix, column vectors of the eigenvector matrix represent directions of principal component changes, and the eigenvalue diagonal matrix represents sizes of variance contributions in directions of the principal components;

[0122] A physical decoupling matrix is constructed based on the eigenvector matrix and the eigenvalue diagonal matrix, coupling relationships between different physical quantities due to dimensional differences and nonlinear interaction effects are eliminated, transformed features are statistically independent of each other, and then the physical decoupling matrix is obtained.

[0123] The calculation formula is as follows:

[0124] ,

[0125] wherein, the physical decoupling matrix is represented; the eigenvector matrix is represented, and column vectors correspond to principal component directions; the eigenvalue diagonal matrix is represented, and is obtained by subjecting the covariance matrix to eigen decomposition, , diagonal elements of the eigenvalue diagonal matrix are eigenvalues of ; is a transpose of ;

[0126] It should be noted that physical quantities such as heat flux density and temperature difference have fixed correlations, eigen decomposition of the covariance matrix projects original features to an orthogonal principal component space, and diagonalization of the covariance matrix makes transformed features linearly independent, therefore, column vectors of the eigenvector matrix are principal component directions, representing independent physical modes in data, such as pure heat conduction and convection, and realizing mathematical decoupling of physical mechanisms; The term is not a normal standardization method, and the negative square root operation of the eigenvalue diagonal matrix is essentially a variance normalization of the principal component, which can eliminate dimensional differences in different directions, and the eigenvector matrix ensures that transformed feature directions are aligned with original physical mechanisms, for example, the first principal component may correspond to a dominant heat transfer mode.

[0127] S3.4, perform decoupling and normalization calculation:

[0128] The data in the historical data set is subjected to physical dimensional standardization processing to eliminate the influence of dimensional differences and scale differences of different physical quantities, then the standardized eigenvector is multiplied by the physical decoupling matrix to obtain a decoupled eigenvector, which can not only eliminate dimensional differences, but also eliminate nonlinear coupling relationships between original features.

[0129] The calculation formula is as follows:

[0130] ,

[0131] Where, Indicates the The decoupled feature vector of samples.

[0132] 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.

[0133] S4.1. Calculation of feature-label mutual information:

[0134] For each decoupled feature vector dimension after decoupling normalization, calculate the mutual information value between the feature and the building energy efficiency grade label;

[0135] The calculation formula is as follows:

[0136] ,

[0137] 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;

[0138] 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.

[0139] S4.2. Constructing physical attention weights:

[0140] 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.

[0141] The calculation formula is as follows:

[0142] ,

[0143] 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 ;

[0144] S4.3. Perform feature adaptive weighting:

[0145] 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.

[0146] The calculation formula is as follows:

[0147] ,

[0148] Where, Indicates the The weighted feature vector of samples; Represents the Hadamard product of element-wise multiplication; Represents vector transpose.

[0149] S5, constructing an energy-saving effect prediction network, the network comprising a local-global dual-channel gated fusion network module, a physically constrained Dirichlet evidence uncertainty quantification module, a differentiable physically rule gated dual-driven classification decision and a multi-objective loss function calculation, and the weighted feature vector is input into the network for prediction;

[0150] The operation of the energy-saving effect prediction network is as follows:

[0151] The weighted feature vector is input into the network for energy-saving effect prediction, and the weighted feature vector is extracted by the local-global dual-channel gated fusion network module to extract local features and global features, and then the extracted features are adaptively fused by gating;

[0152] The fused features are input into the physically constrained Dirichlet evidence uncertainty quantification module to generate physical constraint residuals that collect various thermal constraints, and the fused features are mapped to Dirichlet concentration parameters of each energy-saving level, and then the uncertainty loss of the network is calculated according to the physical constraint residuals and the Dirichlet concentration parameters of each energy-saving level;

[0153] A differentiable physical decision gating strategy is adopted, physical rule confidence is defined for each energy-saving level based on the laws of thermodynamics and engineering experience, then the predicted Dirichlet concentration parameters and the physical rule confidence are dynamically reconciled to generate the prediction probability of the energy-saving level, and finally the loss between the predicted Dirichlet concentration parameters and the physical rule confidence is calculated;

[0154] The total loss of the network is calculated by a multi-objective loss function, including the cross-entropy loss of the prediction result and the true label, and the loss calculated in the network, and a weight coefficient is set for each loss.

[0155] S6, the trainable parameters of the energy-saving effect prediction network are optimized by iterative optimization, the trainable parameters are initialized by Xavier normal distribution, the adaptive moment estimation optimizer is used to minimize the total loss of the network, and the convergence condition for stopping iterative training is set, and finally the trained energy-saving effect prediction model network is obtained;

[0156] S7, after the newly collected data is subjected to physical decoupling normalization and feature adaptive weighting, it is input into the trained energy-saving effect prediction network, and the final prediction probability of each energy-saving level is output, and the level with the maximum probability is taken as the prediction result.

[0157] 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:

[0158] (1) Physical feature grouping and input allocation:

[0159] 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 ;

[0160] 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;

[0161] 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;

[0162] is the local feature dimension, is the total dimension of the feature, and ;

[0163] 2) Extract local spatial feature convolution:

[0164] 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;

[0165] The calculation formula is as follows:

[0166] ,

[0167] 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;

[0168] 3) Global physical characteristics full connection extraction

[0169] 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;

[0170] The calculation formula is as follows:

[0171] ,

[0172] 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;

[0173] 4) Gated Adaptive Feature Fusion:

[0174] 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.

[0175] The calculation formula is as follows:

[0176] ,

[0177] ,

[0178] Where, represents the gate vector; Represents feature concatenation operation; represents the gating weight matrix; express Activation function; represents the fused feature vector;

[0179] 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.

[0180] 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:

[0181] 1) Constructing thermodynamic residual energy penalty term:

[0182] 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.

[0183] The calculation formula is as follows:

[0184] ,

[0185] 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;

[0186] 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 The energy conversion vector is linearly transformed by a material heat capacity coefficient; The energy conversion vector is linearly transformed by a material heat capacity coefficient; The energy conversion vector is linearly transformed by a material heat capacity coefficient;

[0187] It should be noted that, The term is added to the loss function as an L2 norm penalty term, which amplifies the exceeding amount twice and forces the predicted value to return to a reasonable range. The term represents a physical residual energy term, which is realized based on a differentiable approximation function to maintain differentiability during model network training and ensure gradient propagation, The term is added to the loss function as an L2 norm penalty term, which amplifies the exceeding amount twice and forces the predicted value to return to a reasonable range. The term is an input term of the function, and when The physical residual energy term is close to 0 when the term is less than 0. In actual implementation, if the energy conversion vector cannot be correctly constructed, it can be set as a learnable parameter vector to fit the energy conversion relationship through training data, thereby reducing the implementation experience threshold of the present technology.

[0188] 2) Dirichlet concentration parameter mapping:

[0189] The fusion feature vector is input into a fully connected layer and mapped into a Dirichlet distribution concentration parameter vector through an exponential activation function and a plus 1 operation. Each concentration parameter corresponds to an energy saving level. The Dirichlet distribution describes the uncertainty of the network's probability distribution of the sample belonging to each energy saving level, and the size of the concentration parameter reflects the network's confidence in predicting each level, thereby realizing parameterized quantization of classification uncertainty.

[0190] The calculation formula is as follows:

[0191] ,

[0192] In the formula, represents the Dirichlet concentration parameter of the i-th energy saving level, which reflects the "evidence" of the model network for the i-th class, The larger the value, the more sufficient the evidence and the lower the uncertainty. represents the Dirichlet concentration parameter of the first energy saving level; represents the Dirichlet concentration parameter of the i-th energy saving level; represents the Dirichlet concentration parameter of the first energy saving level; represents the Dirichlet concentration parameter of the i-th energy saving level; represents the Dirichlet concentration parameter of the i-th 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 being 1, and the 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;

[0193] 3) Calculate physical regularization uncertainty loss:

[0194] 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.

[0195] The calculation formula is as follows:

[0196] ,

[0197] 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 ;

[0198] 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.

[0199] 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.

[0200] 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:

[0201] 1) Build a thermodynamic rule confidence engine:

[0202] 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.

[0203] The calculation method of physical rule confidence is expressed as:

[0204] ,

[0205] 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 ;

[0206] It should be noted that in the physical rule confidence During the calculation process, Item as a multiplication operation, ensures that all physical indicators meet the threshold at the same time , and the form of multiplication requires all physical indicators to meet the threshold at the same time, such as thermal inertia, thermal conductivity, and thickness, in addition, Item uses the design of the activation function, converts the hard threshold into a soft decision, maintains the differentiability in the model network training process, ensures that the gradient can be propagated, and if a step function is used, such as , , the model network may produce gradient vanishing phenomenon in the training process, and cannot learn the boundary samples close to the threshold.

[0207] 2) Synthetic data - rule gated probability:

[0208] Fuse the expected probability of the Dirichlet distribution predicted by the network based on the fusion feature vector with the physical rule confidence. Specifically, the fusion is performed through a learnable gating weight scalar, and the gating weight scalar is calculated by the fusion feature vector through the gating weight vector and function, dynamically control the relative importance of data prediction and physical rules in the final decision through the gating mechanism, weight the data prediction probability and the physical rule confidence according to the gating weight scalar, and obtain the final prediction probability, and then obtain the prediction result;

[0209] The calculation formula is as follows:

[0210] ,

[0211] ,

[0212] In the formula, represents the gating weight vector, the dimension is the same as the fusion feature vector ; represents the final prediction probability of the energy saving level; represents the gating weight scalar, the value range is , controls the fusion ratio of data prediction and rule confidence, and is set to ; is the transpose of ;

[0213] It should be noted that the item represents the data preliminary prediction probability, and the gating weight scalar dynamically balances the data prediction probability and the physical rule confidence , when the sensor noise is large, it indicates that the data is unreliable, , depends on the physical rule prediction classification, when the data is reliable, , the model network prediction classification depends on the model network;

[0214] 3) Calculate the physical decision boundary loss constraint:

[0215] The physical decision boundary loss acts on the Dirichlet distribution expected probability of network prediction. When the physical rule confidence of a certain level is lower than the set threshold, it indicates that the physical index of the sample strongly supports it belonging to this level. The physical decision boundary loss function will punish the high probability value of the network prediction of this level, forcing the network to increase its uncertainty, that is, to reduce the probability value of this level, and thus ensure that the network will not make overconfident wrong predictions when violating the physical rules.

[0216] The calculation formula is as follows:

[0217] ,

[0218] In the formula, The physical decision boundary loss function is represented by.

[0219] In the specific implementation, the total loss function of the energy saving effect prediction network is calculated by a multi-objective loss function, and the specific operation is as follows:

[0220] The total loss function of network training is calculated, and the total loss function is a multi-objective loss function, which is specifically composed of three parts, including the cross-entropy loss function between the prediction result and the actual label, the uncertainty loss function and the physical decision boundary loss function. The uncertainty loss function and the physical decision boundary loss function are respectively weighted by a weight coefficient;

[0221] The uncertainty loss weight coefficient is set to a larger value close to 1 at the beginning of training, and is smoothly reduced to 0 according to the cosine decay function as the training round progresses, which can realize the gradual exit of uncertainty regularization;

[0222] The physical decision boundary loss weight coefficient is set to 0 at the beginning of training, and is smoothly increased to 1 according to the Sigmoid function when the training round approaches a certain round, which can realize the delay of the physical rule constraint, and thus realize the coordination of the role of different loss terms in different training stages.

[0223] The calculation formula of the total loss function is as follows:

[0224] ,

[0225] In the formula, The total loss function is represented by. The cross-entropy loss function is represented by.

[0226] The uncertainty loss weight coefficient is represented by, and the value range is , control uncertainty loss function at the beginning of training smoothly decays to , characterizes the beginning of training, when is always , and thus realizes the gradual exit of uncertainty regularization, the calculation method is represented as ; represents the minimum value of and , that is, when is , otherwise ; represents the current training round, an integer and ; represents the uncertainty loss starting round, set , or set to 1 / 5 of the total training rounds;

[0227] represents the physical rule loss weight coefficient, the value range is , controls the physical decision boundary loss function, and when the training round approaches , it is smoothly increased from to , realizing the delay of physical rule constraint, the calculation method is in the form of Sigmoid activation function, and the calculation method is represented as ; represents the physical rule loss starting round, set , or set to 1 / 2 of the total training rounds; represents the Sigmoid smoothing coefficient, set .

[0228] It should be noted that the calculation method of the uncertainty loss weight coefficient is the design method of the cosine decay function, and the calculation method of the physical rule loss weight coefficient is the design method of the Sigmoid activation function. The calculation design methods of the two are considered for the stability of the model network training. The switching of cosine decay and Sigmoid activation can ensure smooth training and solve the conflict problem in multi-objective optimization.

[0229] In the specific embodiment, the training batch size of the energy-saving effect prediction network during training is set to 64; the initial learning rate is set to 0.001 when minimizing the total loss function, and the learning rate is decayed 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 iteration number, such as 1000 times, is reached, then the iteration is stopped.

[0230] In the detailed implementation, the trained network is used to predict the energy saving level of newly collected building monitoring data. Specifically, the heat flux density, temperature difference, thermal conductivity and other physical quantities are collected in the same way as in step S1 to form 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 sequentially performed. The weighted feature vector of the sample is input into the trained energy saving effect prediction network. The data is processed by the network through forward propagation, and the final prediction probability of each energy saving level corresponding to the sample is output. The level with the maximum probability is taken as the prediction result. For example, when the final prediction probability of the fourth energy saving level is level 4 is output.

[0231] The Dirichlet concentration parameters of each energy saving level are synchronously provided As the uncertainty quantification result, when the total concentration parameter a low confidence warning is triggered, and manual review is suggested.

[0232] Embodiment 2

[0233] As shown in Figure 2 , the influence of different feature normalization methods on the prediction accuracy of building energy saving effect is evaluated. The traditional Min-Max normalization, Z-Score standardization and the physical constraint decoupling normalization method proposed in the present application are compared. Figure 2 The box plot shows the accuracy range of the three methods, and the scatter points represent the results of a single experiment. The green box corresponds to the method of the present application. The experimental results show that the prediction accuracy of the method of the present application is significantly higher than that of the traditional method, and the result fluctuation range is smaller and the stability is stronger, which verifies the advantages of the physical constraint decoupling normalization in eliminating dimension difference and nonlinear coupling through covariance feature decomposition, effectively solving the feature distribution distortion problem caused by the traditional method, and enabling the network to more accurately capture the essential features of building thermal performance. It should be noted that the horizontal and vertical coordinates of the experimental graph are dimensionless values without physical units.

[0234] As shown in Table 1, the time series data of heat flux density, surface temperature difference and thermal conductivity collected by actual building envelope structure sensors are used as input, and the three normalization methods of Min-Max normalization, Z-Score normalization and the method of the present application are compared and analyzed.

[0235] Table 1 Experimental configuration table

[0236]

[0237] Three analysis dimensions are set in the experiment, including time series shape retention ability, correlation retention degree between physical quantities and feature distribution consistency, as shown in Figure 3As shown in the experimental results, according to the comparison of the time series, it can be found that the original heat flow density (blue curve) presents typical diurnal periodic fluctuation, the peak value appears in the afternoon, and the valley value appears in the early morning. The minimum-maximum normalization (red dotted line) retains the basic shape, but amplifies the small fluctuation at night, and the distortion is obvious. The Z-Score normalization (green dot-dash line) is overall downward, resulting in the appearance of negative values, which destroys the physical meaning (the heat flow density should be positive). The method (purple solid line) perfectly maintains the phase and shape characteristics of the original curve, while suppressing noise interference. It should be noted that the vertical coordinate of the experimental graph is a dimensionless value, and the horizontal coordinate is in units of "hours".

[0238] In the correlation analysis experiment, the correlation coefficient of the heat flow density and the surface temperature difference was calculated, and the specific experimental data are shown in Table 2.

[0239] Table 2 Comparison of correlation coefficients of different methods

[0240]

[0241] The experimental results shown in Table 2 show that the correlation retention rate of the method is close to 100%, which proves that the method can effectively retain the thermodynamic correlation between physical quantities.

[0242] In addition, according to the comparison of the normalized feature distribution Figure 4 It is shown that the traditional method causes distortion of the distribution shape, the minimum-maximum method produces double peaks, and the Z-Score normalization method causes the distribution to widen. The distribution curve of the method is highly consistent with the original data, and the kurtosis and skewness are best maintained, which proves that the scale retention ability has an advantage.

[0243] Example 3

[0244] As Figure 5 shown, the performance response of the three feature weighting strategies is analyzed when the sharpening factor changes, and the curves of the equal weight, mutual information weighting and the physical attention mechanism of the application are compared. Figure 5 The curve shows that the method of the application (green) maintains the best and stable performance under different sharpening factors, especially in the key area (sharpening factor =1.5-2.5), which is the most prominent, which verifies the double advantages of the physical prior guided attention mechanism. On the one hand, the mutual information quantifies the statistical correlation between the features and the labels, and on the other hand, it integrates the physical contribution degree prior knowledge such as thermal inertia index, which overcomes the defects of the traditional method of discarding weakly correlated features or equal weight processing, so that the network can adaptively strengthen the feature expression of the key physical mechanism such as the thermal conductivity coefficient. It should be noted that the horizontal and vertical coordinates of the experimental graph are dimensionless values without physical units.

[0245] Example 4

[0246] AsFigure 6 As shown in the figure, the analysis of the regulatory effect of the physical constraint mechanism on the prediction uncertainty, the performance of the model without physical constraints and the method of the present application on 100 test samples, Figure 6 The broken line reflects the prediction uncertainty index (Dirichlet total concentration parameter), and the green area shows the uncertainty reduction range. The uncertainty curve of the method of the present application is always below that of the traditional method, and the difference is more significant in the sample segment with large sensor noise (such as No. 30-50), which proves that the Dirichlet evidence module of the physical constraint can effectively constrain the prediction results within the physically feasible domain through the thermodynamic residual energy penalty term, solving the problem that the traditional Bayesian method may violate the laws of thermodynamics, and enabling the model to still maintain reasonable uncertainty quantification in the data noise environment. It should be noted that the horizontal and vertical coordinates of the experimental graph are dimensionless values without physical units.

[0247] Example 5

[0248] As Figure 7 shown, the prediction performance of the traditional neural network, the basic physical constraint model and the complete method of the present application on 5 energy-saving levels is compared, Figure 7 The column height represents the average accuracy of each level, and the scatter points represent the results of different test samples. The method of the present application (green column) maintains the highest accuracy in all energy-saving levels, especially in the key high-level (4-5 level) area, which reflects the advantage of the dual-channel gating fusion network of the present application. Through local feature convolution to extract spatial patterns such as temperature field distribution, combined with global feature full connection to learn overall characteristics such as material thermal inertia, and then through the adaptive gating mechanism to realize the feature fusion of physical perception, the complex thermal coupling mechanism of building energy saving is fully captured, and the feature loss problem of single-channel model is overcome. It should be noted that the horizontal and vertical coordinates of the experimental graph are dimensionless values without physical units.

[0249] Although the specific embodiments of the application have been described above with reference to the accompanying drawings, the description is not a limitation on the scope of protection of the present application. Various modifications or variations made by those skilled in the art without creative labor on the basis of the technical solutions of the present application are still within the scope of protection of the present application.

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 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