Load simulation calculation method and system for building energy efficiency

Abstract

The invention relates to the technical field of building energy conservation and intelligent buildings, in particular to a load simulation calculation method and system for building energy efficiency, and the method comprises the steps: collecting multi-source data, and constructing a multi-dimensional data set; building load data is deconstructed into three kinds of heterogeneous modalities including a class sequence, a class image and a class video; calculating a dynamic correction coefficient based on the heat storage effect of the enclosure structure; respectively processing three types of heterogeneous modes through a BiGRU network, an STNN network and a 3DCNN network, and fusing and outputting an initial load by adopting Stacking ensemble learning; optimizing and outputting a final load prediction value through a multivariate feature recurrent neural network; the system integrates a multi-source data acquisition module, a multi-modal feature processing module, a physical correction module, a hybrid prediction module and a time sequence optimization module, and is deployed at an edge node through a knowledge distillation compression model. According to the method, a physical model and a data driving method are fused, and the problems that in the prior art, data processing is insufficient, feature construction is simple, a model fusion shallow layer is insufficient, and calculation architecture and timeliness are insufficient are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical fields of building energy conservation and intelligent buildings, and particularly to a method and system for load simulation calculation of building energy efficiency. Background Art

[0002] In the early stage, building load prediction mainly relied on traditional physical models, calculating theoretical loads through building envelope structure parameters and meteorological data. Although such methods have physical interpretability, they have poor adaptability to complex building forms and user behaviors, and rely on a large number of accurate input parameters, resulting in large errors in practical applications.

[0003] With the development of Internet of Things technology and machine learning, data-driven methods have gradually become the mainstream. By modeling the correlation between historical load data and meteorological characteristics, the prediction flexibility has been improved. Some studies have tried to combine physical models and data-driven models, but only achieved collaboration through simple superposition or parameter adjustment, without deeply exploring the internal relationship between building physical mechanisms and data characteristics.

[0004] Currently, building load prediction technology faces the following core bottlenecks: existing methods mostly model based on one-dimensional time series data, ignoring the spatial distribution of loads and the spatio-temporal dynamic evolution law, resulting in insufficient prediction accuracy for load mutation or spatial coupling scenarios; physical models are difficult to adapt to parameter uncertainties, data models lack physical interpretability, and traditional hybrid methods do not achieve explicit guidance of physical constraints on data modeling; the cloud centralized computing architecture has data transmission delays, and it is difficult to deploy complex deep learning models on edge nodes lightly, making it difficult to meet the real-time regulation requirements of building energy systems; there is no effective processing mechanism for abnormal load data, and simple elimination strategies are often used, resulting in limited model generalization ability.

[0005] Chinese invention patent CN115659459A discloses a method and system for calculating the dynamic heat load of a solar heating building. To solve the problem of inaccurate calculation of the heat load fluctuation of a solar heating building, a dynamic heat load calculation method based on the Z transfer function method is proposed. This invention lacks multi-dimensional feature engineering of load data and cannot explain the spatial heterogeneity and dynamic fluctuation of loads.

[0006] Therefore, the present invention discloses a method and system for load simulation calculation of building energy efficiency. Summary of the Invention

[0007] The purpose of the present invention is to propose a method and system for load simulation calculation of building energy efficiency to solve the problems of single data modality, insufficient collaboration between physical models and data-driven models, poor real-time performance, and weak robustness under extreme working conditions in the prior art.

[0008] To achieve the above purpose, the present invention adopts the following technical solutions: A method for load simulation calculation of building energy efficiency, comprising: Step S1, multi-source data preprocessing: Obtain building physical parameters, meteorological data, and building load data through the Internet of Things sensor network, and construct a multi-dimensional data set after removing outliers using the Isolation Forest algorithm; Step S2, multi-modal feature construction: Decompose the building load data into three heterogeneous modalities: class sequence, class image, and class video; Step S3, dynamic physical correction: Calculate the dynamic correction coefficient based on the heat storage effect of the building envelope , and quantify the real-time impact of wall heat storage and release on the load; Step S4, hybrid model prediction: Process the three heterogeneous modalities through the BiGRU network, STNN network, and 3DCNN network respectively, and use Stacking ensemble learning to fuse and output the initial load; Step S5, time series optimization output: Input the initial load, dynamic correction coefficient , building load data sequence, and meteorological data sequence into a multi-feature recurrent neural network, and output the final load prediction value to achieve the collaborative optimization of the physical model and data-driven.

[0009] A load simulation calculation system for building energy efficiency, comprising: A multi-source data acquisition module, a multi-modal feature processing module, a physical correction module, a hybrid prediction module, a time series optimization module, and an edge computing module; The multi-source data acquisition module is used to obtain building physical parameters, meteorological data, and building load data; The multi-modal feature processing module is used to construct class sequence, class image, and class video features; The physical correction module is used to calculate the thickness of the violently fluctuating layer of the building envelope and the dynamic correction coefficient ; The hybrid prediction module includes BiGRU, STNN, 3DCNN sub-models and a Stacking integrator; The time series optimization module runs a multi-feature recurrent neural network and outputs the final load prediction value; The edge computing module compresses the multi-feature recurrent neural network through knowledge distillation and deploys it on a local edge node to improve the real-time prediction ability of the edge node.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention captures time series dependencies through the class sequence modality, models spatial distributions through the class image modality, and mines spatio-temporal coupling features through the class video modality. The present invention can reduce the load prediction error and improve the prediction accuracy of spatial load mutation scenarios.

[0011] Through dynamic physical correction, the present invention calculates the dynamic correction coefficient based on the heat storage effect of the envelope structure, quantifies the impact of the heat storage and release of the wall on the load, and realizes the deep integration of the physical model and the data-driven model. The present invention can make up for the defects that the physical model in the traditional method is difficult to adapt to parameter uncertainty and the data-driven model lacks physical interpretability.

[0012] At the bottom layer, the present invention independently processes three types of heterogeneous modalities through BiGRU, STNN, and 3DCNN to extract differential features. At the top layer, the initial load is integrated and output through Stacking ridge regression, and then optimized in time series by integrating with the historical sequence through LSTM. The present invention can improve the generalization ability of the model and reduce the prediction error fluctuation of cross-building types.

[0013] The present invention compresses the multi-feature recurrent neural network through knowledge distillation and deploys it on the local edge node. The present invention can reduce the model inference latency and realize the real-time linkage between load prediction and energy regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1 It is a schematic flowchart of a method for load simulation calculation of building energy efficiency provided by an embodiment of the present invention; Figure 2 It is a schematic flowchart of the dynamic correction coefficient calculation of a method for load simulation calculation of building energy efficiency provided by an embodiment of the present invention; Figure 3 It is a schematic flowchart of the hybrid model prediction of a method for load simulation calculation of building energy efficiency provided by an embodiment of the present invention; Figure 4 It is a schematic diagram of the system architecture of a load simulation calculation system for building energy efficiency provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in combination with the accompanying drawings and preferred embodiments, details the specific implementation manners, structures, features, and effects of a method and system for load simulation calculation of building energy efficiency proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs.

[0018] The following embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0019] The following specifically describes the specific solutions of a building energy efficiency load simulation calculation method and system provided by the present invention with reference to the accompanying drawings.

[0020] Embodiment

[0021] Please refer to Figure 1 which is a schematic flowchart of the method provided by an embodiment of the present invention.

[0022] In an embodiment of the present invention, a building energy efficiency load simulation calculation method is provided. The method includes: Step S1, multi-source data preprocessing. Obtain building physical parameters, meteorological data, and building load data through the Internet of Things sensor network, and construct a multi-dimensional data set after removing outliers using the Isolation Forest algorithm; Among them, in step S1, the following sub-steps are further included: S1-1, the building physical parameters include thermal conductivity, material density, material specific heat capacity, envelope heat transfer coefficient, and envelope area; S1-2, the meteorological data includes outdoor temperature and indoor set temperature; S1-3, the building load data includes historical total load data and user unit load data; S1-4, detect and remove abnormal data points through the Isolation Forest algorithm, and fill in missing values by linear interpolation; S1-5, partition and standardize the cleaned data according to the feature type, including: Adopt Z-score standardization for meteorological data and building load data; Adopt Min-Max normalization for building physical parameters; S1-6, construct a multi-dimensional data set including building load data, meteorological data, and time series features. The time series features include sine or cosine encoding of hours, day-of-week type, and holiday flag.

[0023] It should be noted that meteorological data is collected by temperature and humidity sensors; historical total load is collected by smart meters; user unit load is collected by sub-metering meters; all sensor data is synchronously collected at 1-hour intervals, and the timestamps are aligned to the ISO 8601 standard format.

[0024] For building physical parameters, thermal conductivity, density, and specific heat capacity are collected by material detectors.

[0025] The heat transfer coefficient of the building envelope is calculated by the following formula: , where R represents the total thermal resistance; K represents the heat transfer coefficient of the building envelope.

[0026] The area F of the building envelope is obtained from the point cloud data acquired by a 3D laser scanner and generated after being processed by BIM software.

[0027] For the outlier detection of the Isolation Forest algorithm, the outlier ratio э = 1% is set, and outliers are identified by calculating the average path length of data points in the random forest; if the path length of a certain point , it is determined as an outlier and removed; where, represents the mean value, represents the standard deviation.

[0028] In the processing of missing values, for data points with a missing duration ≤ 2 hours in the continuous time series, linear interpolation is used for filling; if the missing duration exceeds 2 hours, the average value of the same period in the previous and next 24 hours is used for filling.

[0029] The time features capture periodicity through sine / cosine encoding, and the week / holiday flags distinguish the load patterns. All data is aligned at the hourly granularity to construct a multi-dimensional data set.

[0030] Step S2, multi-modal feature construction, decomposes the building load data into three heterogeneous modalities: class sequence, class image, and class video; Among them, in step S2, the following sub-steps are also included: S2-1, class sequence modality construction, concatenates the historical total load data, meteorological data, and time series features along the time axis into a one-dimensional feature vector, which is used as the input of the BiGRU network to generate the first prediction value reflecting the time series evolution law ; the vector dimension is T×m, where T is the time step and m is the feature dimension; S2-2, class image modality construction, includes: Mapping the load value of the user unit to spatial pixel points, and each pixel value represents the load size of the corresponding unit; Constructing a spatial matrix representing the load distribution of the building functional areas, where the units in the same functional area are spatially adjacent in the matrix; the functional areas include the office area, commercial area, and public area; Based on the statistical correlation analysis of the load time series data, optimizing the spatial arrangement position of the units in the matrix to make the units with high correlation close to each other; Performing Min-Max normalization processing on the matrix to make the element value range be [0,1]; S2-3, class video modality construction, stacks the class image modality matrices at nine equally spaced consecutive moments along the time dimension into a three-dimensional tensor to capture the spatio-temporal dynamic changes of the load.

[0031] It should be noted that for the feature splicing method, the historical total load data, meteorological data, and time series features are spliced along the time axis into a T×6-dimensional feature vector; the time step T is taken as 24, covering the load change in a complete daily cycle.

[0032] The pixel mapping rule includes: The building floor plan is divided into an N×N grid (N is dynamically adjusted according to the total number of units, preferably 11×11); The units in the same functional area are four-connected and adjacent in the matrix; Optimized arrangement based on the Spearman correlation coefficient.

[0033] For tensor construction, take the class image matrices for 9 consecutive hours (from t - 8 to t), stack them into a three-dimensional tensor of [11, 11, 9], and set the number of input channels of the 3DCNN to 9.

[0034] The data of the three types of modalities are synchronized based on the whole hour to ensure temporal consistency.

[0035] Please refer to Figure 2 It is a schematic diagram of the dynamic correction coefficient calculation process provided by an embodiment of the present invention.

[0036] Step S3, dynamic physical correction, calculate the dynamic correction coefficient based on the heat storage effect of the envelope structure , and quantify the real-time impact of the heat storage and release of the wall on the load; Among them, in step S3, the following sub-steps are further included: S3-1, based on the Fourier heat conduction theory, calculate the thickness δ of the violently fluctuating layer of the envelope structure, and the specific formula is: , where α represents the thermal diffusivity of the wall material; T represents the temperature fluctuation period, with a value of 24 hours; represents the thickness of the fluctuating layer, which is used to define the effective area of the wall heat storage; π represents the pi, which is a mathematical constant; The calculation formula for the thermal diffusivity α of the wall material is: , where, represents the thermal conductivity; represents the material density; represents the specific heat capacity of the material; S3-2, calculate the heat storage of the violently fluctuating layer at τ moment, and the specific formula is: , where ρ represents the material density; c represents the specific heat capacity of the material; represents the thickness of the fluctuating layer; F represents the area of the envelope structure; represents the temperature difference between both sides of the fluctuating layer; represents the heat storage of the violently fluctuating layer at τ moment; Through the outdoor temperature at τ moment And the indoor set temperature Calculate the absolute value of the difference between the two; S3-3. Calculate the heat transfer quantity on the inner surface of the building envelope at time τ , and the specific formula is: , where K represents the heat transfer coefficient of the building envelope; F represents the area of the building envelope; represents the indoor set temperature; is the outdoor temperature; represents the heat transfer quantity on the inner surface of the building envelope at time τ; S3-4. Calculate the dynamic correction coefficient , and the specific formula is: , where represents the heat transfer quantity on the inner surface of the building envelope at time τ; represents the heat storage quantity in the violently fluctuating layer at time τ; represents the dynamic correction coefficient.

[0037] It should be noted that, based on Fourier's heat conduction theory, the thickness δ of the fluctuating layer is calculated by a formula, where the thermal diffusivity α is determined by the thermal physical properties of the material (λ, ρ, c), and the temperature fluctuation period T is taken as 24 hours to reflect the diurnal variation.

[0038] The heat storage quantity in the violently fluctuating layer at time τ is calculated through the heat capacity of the material, the volume of the fluctuating layer, and the temperature difference between indoors and outdoors, where the temperature difference takes the absolute value of the indoor set temperature and the real-time outdoor temperature, reflecting the driving effect of the environmental temperature difference on heat storage.

[0039] The heat transfer quantity on the inner surface of the building envelope at time τ is calculated based on the heat transfer coefficient K of the building envelope and the temperature difference between indoors and outdoors, representing the contribution of the direct heat transfer through the wall to the load.

[0040] The dynamic correction coefficient is the ratio of the heat transfer quantity on the inner surface to the heat storage quantity, quantifying the real-time impact of the heat storage effect, with a value range of 0-1, and is used to directly input the subsequent time series optimization module to correct the load prediction.

[0041] Please refer to Figure 3 which is the schematic diagram of the hybrid model prediction process provided by the embodiment of the present invention.

[0042] Step S4. Hybrid model prediction: Process three types of heterogeneous modalities through the BiGRU network, STNN network, and 3DCNN network respectively, and adopt Stacking ensemble learning to fuse and output the initial load; Among them, in step S4, the following sub-steps are also included: S4-1. Process the class sequence modality: Process the class sequence modality features through the BiGRU network and output the first prediction value ; S4-2, Image-like modality processing, processing the image-like modality features through the STNN network and outputting the second prediction value ; S4-3, Video-like modality processing, processing the video-like modality features through the 3DCNN network and outputting the third prediction value ; S4-4, Using the Stacking ensemble learning method to fuse the first prediction value , the second prediction value and the third prediction value , generating the initial load ; The meta-learner of the ensemble learning is the ridge regression model, trained through 5-fold cross-validation, and the ensemble formula is: , where represents the ensemble mapping function based on the ridge regression model; represents the initial load.

[0043] It should be noted that BiGRU represents a bidirectional gated recurrent unit, capturing the bidirectional temporal dependencies of the sequence-like modality. The BiGRU network contains 2 hidden layers, with 64 neurons in each layer, the activation function is tanh, the input dimension is T×m (T = 24 hours, m = 6), and the first prediction value is output .

[0044] STNN represents a spatio-temporal neural network, extracting the spatial distribution features of the image-like modality. The STNN network contains 3 convolutional layers, and the convolutional kernel sizes are 5×5, 3×3, and 3×3 in sequence. The input is an 11×11 functional area load matrix, and the second prediction value is output .

[0045] 3DCNN represents a three-dimensional convolutional neural network, capturing the spatio-temporal dynamic features of the video-like modality. The 3DCNN network includes 3 3D convolutional layers, and the convolutional kernel sizes are 5×5×3, 3×3×3, and 3×3×3 in sequence, and the third prediction value is output .

[0046] Using Stacking ensemble learning to fuse the prediction values of the three modalities, using the ridge regression model as the meta-learner, the regularization parameter λ = 1.0, trained through 5-fold cross-validation, generating the initial load , realizing the complementary enhancement of temporal, spatial, and spatio-temporal features.

[0047] Among them, the input and output dimensions of each modality network match the load prediction target, the training data is divided into 7:3, and the model performance is evaluated through the overall error and robustness, ensuring the generalization ability of the initial load prediction.

[0048] Step S5, timing optimization output, input the initial load, dynamic correction coefficient , the building load data sequence and the meteorological data sequence into a multi-feature recurrent neural network, and output the final load prediction value to achieve the collaborative optimization of the physical model and data-driven; Among them, in step S5, the following sub-steps are also included: S5-1, use a multi-feature recurrent neural network to process the input data, and the update formula of the LSTM memory unit is: , where represents the state of the memory unit at the current moment; represents the cell state at the previous moment τ-1, that is, the historical memory information; , are the outputs of the forget gate and the input gate respectively; represents the dynamic correction coefficient; represents concatenating the previous moment hidden state, the current input and the dynamic correction coefficient; represents the weight matrix; represents the bias term; represents element-wise multiplication; tanh represents the hyperbolic tangent activation function; S5-2, calculate the final load prediction value through the output layer, and the specific formula is: , where, represents the LSTM hidden state at the current moment; represents the weight matrix of the output layer; represents the bias vector, represents the final load prediction value. [[ID=##]]

[0049] It should be noted that the LSTM memory unit combines the dynamic correction coefficient through the forget gate and the input gate , and selectively retains the historical load pattern and the current physical correction information. As an external control signal, it participates in the gating calculation through the concatenation operation, enabling the model to quickly adjust the memory weight during load mutation and improving the prediction robustness. [[ID=##]]

[0050] The output layer maps the LSTM hidden state to the final prediction value , and this state integrates the timing feature (initial load), physical correction ( ) and environmental variables (meteorological data). Through linear transformation, the weighted fusion of multi-source information is realized, making the prediction result conform to both the data trend and the thermodynamic law. [[ID=##]]

[0051] All input data are standardized. The initial load, the building load data sequence and the meteorological data are normalized by Z-score, and the dynamic correction coefficient Keep within the range of [0, 1], align all sequences at the hourly level to ensure the temporal consistency of the LSTM input.

[0052] The model training uses a weighted loss function to balance the overall error, robustness, and relative error. It is iteratively trained through the Adam optimizer with a learning rate set to 0.001 to ensure that the model can capture the details of load fluctuations while maintaining generalization ability.

[0053] Please refer to Figure 4 which is the schematic diagram of the system architecture provided by the embodiments of the present invention.

[0054] A load simulation calculation system for building energy efficiency, comprising: a multi-source data acquisition module, a multi-modal feature processing module, a physical correction module, a hybrid prediction module, a temporal optimization module, and an edge computing module; The multi-source data acquisition module is used to obtain building physical parameters, meteorological data, and building load data; The multi-modal feature processing module is used to construct sequence-like, image-like, and video-like features; The physical correction module is used to calculate the thickness of the violently fluctuating layer of the envelope structure and the dynamic correction coefficient ; The hybrid prediction module includes BiGRU, STNN, 3DCNN sub-models, and a Stacking integrator; The temporal optimization module runs a multi-feature recurrent neural network and outputs the final load prediction value; The edge computing module compresses the multi-feature recurrent neural network through knowledge distillation and deploys it on local edge nodes to improve the real-time prediction ability of edge nodes.

[0055] It should be noted that the multi-source data acquisition module obtains data through BIM models, Internet of Things sensors, and smart meters, and outputs a multi-dimensional data set after outlier removal, missing value filling, and standardization processing.

[0056] The multi-modal feature processing module deconstructs the data into sequence-like (temporal), image-like (spatial), and video-like (spatiotemporal) features, and inputs them into the BiGRU, STNN, and 3DCNN sub-models respectively.

[0057] The physical correction module calculates the thickness of the fluctuating layer and the dynamic correction coefficient based on Fourier heat conduction theory to quantify the heat storage effect of the envelope structure.

[0058] The hybrid prediction module generates the initial load by fusing the three-modal prediction values through Stacking ensemble learning.

[0059] The temporal optimization module uses the LSTM network combined with the dynamic correction coefficient and temporal data to output the final load prediction value.

[0060] The edge computing module compresses the model through knowledge distillation and deploys it on local nodes to improve the real-time prediction ability of edge nodes and support the dynamic regulation of the building energy system.

[0061] For the load simulation calculation method and system of building energy efficiency in this embodiment, after the operation is completed, each module and device shall be shut down one by one according to the specified procedures to ensure safe power-off, and the monitoring data and operation records during the operation shall be sorted out and saved to provide a basis for subsequent clinical research and quality control.

[0062] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A load simulation calculation method for building energy efficiency, characterized in that, The method includes: Step S1, multi-source data preprocessing, obtaining building physical parameters, meteorological data, and building load data through the Internet of Things sensor network, and constructing a multi-dimensional data set after removing outliers using the Isolation Forest algorithm; Step S2, multi-modal feature construction, decomposing the building load data into three heterogeneous modalities: class sequence, class image, and class video; Step S3, dynamic physical correction, calculating a dynamic correction coefficient based on the heat storage effect of the building envelope , quantifying the real-time impact of heat storage and release in the wall on the load; Step S4, hybrid model prediction, processing the three heterogeneous modalities through the BiGRU network, STNN network, and 3DCNN network respectively, and using Stacking ensemble learning to fuse and output the initial load; Step S5, timing optimization output, input the initial load, dynamic correction coefficient , building load data sequence and meteorological data sequence into the multi-feature recurrent neural network, and output the final load prediction value to achieve the collaborative optimization of the physical model and data-driven.

2. A load simulation calculation method for building energy efficiency according to claim 1, characterized in that: Wherein in step S1, the following sub-steps are further included: S1-1, the building physical parameters include thermal conductivity, material density, material specific heat capacity, envelope heat transfer coefficient, and envelope area; S1-2, the meteorological data includes outdoor temperature and indoor set temperature; S1-3, the building load data includes historical total load data and user unit load data; S1-4, detecting and removing abnormal data points through the Isolation Forest algorithm, and filling in missing values through linear interpolation; S1-5, partitioning and normalizing the cleaned data according to feature types, including: Using Z-score normalization for the meteorological data and building load data; Using Min-Max normalization for the building physical parameters; S1-6, constructing a multi-dimensional data set including building load data, meteorological data, and time series features, and the time series features include sine or cosine encoding of hours, day-of-week type, and holiday flag.

3. A load simulation calculation method for building energy efficiency according to claim 1, characterized in that: Wherein in step S2, the following sub-steps are further included: S2-1. Class sequence mode construction: The historical total load data, meteorological data, and time series features are concatenated into a one-dimensional feature vector along the time axis as the input of the BiGRU network, which is used to generate the first prediction value reflecting the time series evolution law. The dimension of the vector is T×m, where T is the time step and m is the feature dimension. S2-2, class image modality construction, including: Mapping the user unit load value to spatial pixel points, and each pixel value represents the load size of the corresponding unit; Constructing a spatial matrix representing the load distribution of building functional areas, and units in the same functional area are spatially adjacent in the matrix; the functional areas include office areas, commercial areas, and public areas; Based on the statistical correlation analysis of load time series data, optimizing the spatial arrangement position of units in the matrix to make units with high correlation close to each other; Performing Min-Max normalization processing on the matrix to make the element value range [0,1]; S2-3, class video modality construction, stacking the class image modality matrices at nine consecutive equally spaced moments along the time dimension into a three-dimensional tensor for capturing the spatio-temporal dynamic changes of the load.

4. A load simulation calculation method for building energy efficiency according to claim 1, characterized in that: Wherein in step S3, the following sub-steps are further included: S3-1. Based on Fourier's heat conduction theory, calculate the thickness δ of the violently fluctuating layer of the envelope structure. The specific formula is as follows: , where α represents the thermal diffusivity of the wall material; T represents the temperature fluctuation period, with a value of 24 hours; represents the thickness of the fluctuating layer, which is used to define the effective heat storage area of the wall; π represents the pi, which is a mathematical constant; The calculation formula for the thermal diffusivity α of the wall material is as follows: , where represents the thermal conductivity; represents the material density; represents the specific heat capacity of the material; S3-2, calculate the heat storage of the violently fluctuating layer at time τ , and the specific formula is as follows: , where ρ represents the material density; c represents the specific heat capacity of the material; represents the thickness of the fluctuating layer; F represents the area of the enclosure structure; represents the temperature difference between both sides of the fluctuating layer; represents the heat storage of the violently fluctuating layer at time τ; The absolute value of the difference between the outdoor temperature at time τ and the indoor set temperature is calculated. S3-3, calculate the heat transfer quantity on the inner surface of the building envelope at time τ , and the specific formula is as follows: , where, K represents the heat transfer coefficient of the building envelope; F represents the area of the building envelope; represents the indoor set temperature; is the outdoor temperature; represents the heat transfer quantity on the inner surface of the building envelope at time τ; S3-4, calculate the dynamic correction coefficient , and the specific formula is: , where represents the heat transfer amount on the inner surface of the enclosure structure at time τ; represents the heat storage amount in the violently fluctuating layer at time τ; represents the dynamic correction coefficient.

5. A load simulation calculation method for building energy efficiency according to claim 1, characterized in that : Wherein in step S4, the following sub-steps are further included: S4-1, class sequence modality processing, processing the class sequence modality features through a BiGRU network and outputting a first prediction value ; S4-2, class image modality processing, processing class image modality features through an STNN network and outputting a second predicted value ; S4-3, Video-like modality processing, processes video-like modality features through a 3D CNN network and outputs a third prediction value ; S4-4, using the Stacking ensemble learning method to fuse the first predicted value , the second predicted value and the third predicted value to generate an initial load ; The meta-learner of the ensemble learning is a ridge regression model, trained by 5-fold cross-validation, and the ensemble formula is: , where represents the ensemble mapping function based on the ridge regression model; represents the initial load.

6. A load simulation calculation method for building energy efficiency according to claim 1, characterized in that: Wherein in step S5, the following sub-steps are further included: S5-1. Process the input data using a multi - feature recurrent neural network, where the update formula for the LSTM memory cell is: , where represents the state of the memory cell at the current time; represents the cell state at the previous time τ - 1, that is, the historical memory information; and are the outputs of the forget gate and the input gate respectively; represents the dynamic correction coefficient; represents concatenating the previous hidden state, the current input, and the dynamic correction coefficient; represents the weight matrix; represents the bias term; represents element - wise multiplication; tanh represents the hyperbolic tangent activation function; S5-2. Calculate the final load prediction value through the output layer. The specific formula is: , where represents the LSTM hidden state at the current moment; represents the output layer weight matrix; represents the bias vector, represents the final load prediction value.

7. A load simulation calculation system for building energy efficiency, which is used to implement the load simulation calculation method for building energy efficiency described in any one of claims 1-6, characterized in that, Including: Multi-source data acquisition module, multi-modal feature processing module, physical correction module, hybrid prediction module, time series optimization module, and edge computing module; The multi-source data acquisition module is used to obtain building physical parameters, meteorological data, and building load data; The multi-modal feature processing module is used to construct class sequence, class image, and class video features; The physical correction module is used to calculate the thickness of the violently fluctuating layer of the enclosure structure and the dynamic correction coefficient ; The hybrid prediction module includes BiGRU, STNN, 3DCNN sub-models, and a Stacking integrator; The time series optimization module runs a multi-variate feature recurrent neural network and outputs the final load prediction value; The edge computing module compresses the multi-variate feature recurrent neural network through knowledge distillation and is deployed on a local edge node to enhance the real-time prediction ability of the edge node.

Images