Multi-modal fusion exhibition building tall atrium natural ventilation evaluation method

Abstract

The invention discloses a multi-modal fusion exhibition building tall atrium natural ventilation evaluation method. The method comprises the steps that multi-modal data are obtained, wherein the multi-modal data comprise meteorological time sequence parameters, building three-dimensional point cloud and opening geometric parameters and air conditioner cooling parameters; inputting the multi-modal data into a multi-modal fusion neural network model to obtain a natural ventilation evaluation result, the natural ventilation evaluation result comprising vertical temperature distribution of the atrium and the ventilation quantity of each opening of the atrium; wherein the multi-modal fusion neural network model comprises a meteorological time sequence feature extraction module, a building geometric coding module, an air conditioning system parameter coding module, a cross-modal feature fusion module, a vertical temperature field prediction module and a ventilation quantity prediction module. According to the method, a space-time cooperative sensing mechanism of meteorological time sequence dynamic, building space topology and air conditioner parameter features is established, and the technical bottleneck of a traditional method in the aspects of cross-modal feature space alignment and interaction efficiency is broken through.

Description

Technical Field

[0001] The present invention belongs to the field of building ventilation prediction, and particularly relates to a method for evaluating the natural ventilation of a large atrium in a convention and exhibition building by multi-modal fusion. Background Art

[0002] Currently, the building ventilation volume prediction technology mainly adopts a hybrid method of mechanism modeling and data-driven based on computational fluid dynamics (CFD) numerical simulation, but there are still significant limitations in multi-physical field coupling modeling and engineering practicability.

[0003] Traditional hybrid models mostly adopt static parametric mechanism equations and global average turbulence correction schemes, which are difficult to effectively handle the spatial heterogeneity of flow coefficients caused by the geometric diversity of building openings. Especially in the modeling of the unsteady characteristics of vortex motion in complex atrium spaces, the existing methods generally adopt the Reynolds-averaged Navier-Stokes equation combined with the k-ε turbulence model, and its calculation takes up to several hours and is difficult to meet the real-time control requirements. Although the data-driven method has improved the ability to extract temporal sequence features through networks such as LSTM and CNN, most studies have not paid attention to the explicit association between meteorological temporal dynamics and three-dimensional spatial topology, resulting in insufficient spatio-temporal coupling modeling accuracy of key physical phenomena (such as the chimney effect driven by thermal pressure and the cross-ventilation dominated by wind pressure). Most existing fusion models adopt simple splicing or weighted averaging in cross-modal feature alignment, and fail to effectively solve the problem of spatial scale mismatch between building point cloud data and air conditioning parameter features, especially with significant errors in capturing the sudden change characteristics of thermal gradients in the area near the air supply outlet. These limitations severely restrict the accurate ventilation volume regulation ability of intelligent building control systems in complex environments, and there is an urgent need to develop a new fusion prediction framework that combines physical interpretability and data-driven advantages. Summary of the Invention

[0004] In order to at least solve one of the problems existing in the prior art, the present invention provides a method for evaluating the natural ventilation of a large atrium in a convention and exhibition building by multi-modal fusion, providing a new direction and technical solution for the optimal design of the atrium space.

[0005] A method for evaluating the natural ventilation of a large atrium in a convention and exhibition building by multi-modal fusion provided by the present invention includes the steps of: Obtaining multi-modal data, where the multi-modal data includes meteorological temporal parameters, building three-dimensional point cloud and opening geometric parameters, and air conditioning cooling parameters; Inputting the multi-modal data into a multi-modal fusion neural network model to obtain a natural ventilation evaluation result, where the natural ventilation evaluation result includes the vertical temperature distribution in the atrium and the ventilation volume of each opening in the atrium; Among them, the multi-modal fusion neural network model includes a meteorological time-series feature extraction module, a building geometry encoding module, an air-conditioning system parameter encoding module, a cross-modal feature fusion module, a vertical temperature field prediction module, and a ventilation volume prediction module. The meteorological time-series feature extraction module is used to extract the temporal dynamic features of meteorological time-series parameters through a bidirectional long short-term memory network and a multi-head self-attention mechanism to obtain a meteorological feature vector; the building geometry encoding module is used to generate building geometry features by fusing the opening position and area parameters based on voxelization preprocessing, feature extraction, and dynamic graph convolution; the air-conditioning system parameter encoding module is used to map the air-conditioning operation parameters into a three-dimensional parameter field to obtain air-conditioning parameter features; the cross-modal feature fusion module is used to perform spatial alignment and multi-modal interaction modeling on the meteorological feature vector, building geometry features, and air-conditioning parameter features, perform feature compression and semantic enhancement to obtain fused features, and make the fused features consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to jointly realize the vertical temperature field modeling through the superposition of basis functions and residual correction to obtain the vertical temperature distribution in the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening.

[0006] The present invention also provides a natural ventilation evaluation system for a large atrium of an exhibition building with multi-modal fusion, including: A data acquisition unit for acquiring multi-modal data, where the multi-modal data includes meteorological time-series parameters, building three-dimensional point clouds, opening geometric parameters, and air-conditioning cooling parameters; An evaluation unit for inputting the multi-modal data into the multi-modal fusion neural network model to obtain a natural ventilation evaluation result, where the natural ventilation evaluation result includes the vertical temperature distribution in the atrium and the ventilation volume of each opening in the atrium.

[0007] The present invention also provides a computer device, and when its processor executes a computer program, it implements the natural ventilation evaluation method for a large atrium of an exhibition building with multi-modal fusion.

[0008] The present invention also provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the natural ventilation evaluation method for a large atrium of an exhibition building with multi-modal fusion.

[0009] Compared with the prior art, the present invention has achieved multi-dimensional technological innovations in the field of building ventilation prediction, and its core advantages are reflected in the following aspects: (1) By constructing a multi-modal fusion architecture of bidirectional LSTM-attention collaborative coding and three-dimensional point cloud geometry learning, an innovative spatio-temporal collaborative perception mechanism of meteorological time-series dynamics, building spatial topology, and air-conditioning parameter features is established, breaking through the technical bottlenecks of traditional methods in cross-modal feature space alignment and interaction efficiency; (2) A geometric topology-sensitive graph neural network is proposed, which adopts a dynamic neighborhood construction multi-head attention aggregation strategy to achieve refined modeling of the spatial layout of building openings and the turbulent interference effect, and improves the analytical ability of complex cavity connectivity features compared with the conventional CNN encoding method; (3) A Bayesian optimization-based dynamic fusion framework is constructed to generate the adaptive weights of the mechanism model and the data-driven branch through variational inference, which significantly enhances the prediction robustness of the model under critical pressure difference conditions while retaining the physical core of the Darcy-Weisbach equation; (4) Through the collaborative calculation of spatio-temporal dilated convolution and sparse attention, and through the adjustable receptive field and k-nearest neighbor optimization, while maintaining the accuracy of capturing turbulent pulsation characteristics, the computational complexity is reduced, providing technical support for intelligent control. Description of the Drawings

[0010] Figure 1 It is a flowchart of the steps of a method for evaluating the natural ventilation of a tall atrium in an exhibition building with multi-modal fusion provided by an embodiment of the present invention.

[0011] Figure 2 It is a schematic diagram of the composition of a multi-modal fusion neural network model in an embodiment of the present invention.

[0012] Figure 3 It is a schematic diagram of the composition of a system for evaluating the natural ventilation of a tall atrium in an exhibition building with multi-modal fusion in an embodiment of the present invention. Detailed Embodiments

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts are within the scope of the present invention.

[0014] In this embodiment, aiming at the existing defects in multi-physical-field coupling modeling and the bottleneck of cross-modal feature fusion in the field of building ventilation volume prediction, the key technical problems of heterogeneous data collaborative calculation and physical constraint embedding are mainly solved. It mainly addresses the deficiencies of existing hybrid models in dynamic environment adaptability and spatial topology perception: First, the static parametric equations in traditional mechanism-data fusion architectures are difficult to represent the spatial heterogeneity of flow coefficients caused by the geometric diversity of openings, and lack graph structure modeling of turbulent interference between adjacent openings; Second, the spatial-temporal-physical dimension mismatch of multi-source heterogeneous data (meteorological time series, building point clouds, air-conditioning parameters) leads to low cross-modal feature interaction efficiency, and existing methods cannot effectively establish the non-linear correlation between hourly meteorological fluctuations and local physical fields in three-dimensional space; Third, conventional fusion strategies rely on empirical weights in adjusting the contribution degrees of mechanism models and neural networks, and cannot achieve adaptive allocation according to geometric features such as the aspect ratio and azimuth angle of openings, resulting in insufficient stability of flow direction prediction in critical pressure difference regions; Fourth, the weak constraint forms of existing deep learning models for the laws of thermodynamics conservation and dimensional consistency are prone to generating non-physical oscillation solutions under conditions of small openings or sudden boundary conditions. The embodiment of the present invention realizes the coordinated improvement of the physical rationality and working condition generalization ability of ventilation volume prediction in complex building environments by constructing a geometric topology-sensitive graph attention network, a spatio-temporal collaborative cross-modal interaction mechanism, a dynamically fused framework optimized by Bayesian, and a feature space constrained by differential equations.

[0015] Please refer to Figure 1 , a method for evaluating natural ventilation of a tall atrium in a convention and exhibition building with multi-modal fusion provided in this embodiment, includes the following steps: Step 1: Obtain input data, where the input data includes meteorological time series parameters, building three-dimensional point clouds, opening geometric parameters, and air-conditioning cooling parameters.

[0016] In one embodiment of the present invention, the three types of core input data required include: (1) Meteorological time series parameters: total solar radiation intensity, outdoor temperature, outdoor wind speed and direction; (2) Building three-dimensional point clouds and opening geometric parameters: the form of the convention and exhibition building (used to obtain the wind pressure on the outer surface of the building through CFD simulation), the position of the atrium opening, the area of the atrium opening, the position of the atrium skylight, the area of the atrium skylight, the internal form of the atrium (three-dimensional form), and collect building three-dimensional point cloud data as the input of the building form; (3) Air-conditioning cooling parameters: stratified air-conditioning cooling capacity.

[0017] Step 2: Input the multi-modal data into the multi-modal fusion neural network model to obtain the natural ventilation evaluation result, where the natural ventilation evaluation result includes the vertical temperature distribution in the atrium and the ventilation volume of each opening in the atrium.

[0018] The multi-modal fusion neural network model constructed in this embodiment is a deep neural network architecture for multi-modal spatio-temporal feature fusion. Please refer to Figure 2 , and through modules such as the meteorological time-series feature extraction module, building geometric coding module, cross-modal feature fusion module, vertical temperature field prediction module, and ventilation volume prediction module, the joint prediction of the atrium opening ventilation volume and vertical temperature distribution is realized. Based on meteorological time-series feature extraction, building geometric coding, and air-conditioning parameter fusion, a dual-task decoder is used to generate high-precision prediction results respectively, providing the intelligent building control system with the ability to predict key environmental parameters.

[0019] The meteorological time-series feature extraction module is used to extract the temporal dynamic features of meteorological time-series parameters through a bidirectional long short-term memory network (BiLSTM) and a multi-head self-attention mechanism, obtain meteorological feature vectors, and realize the temporal feature modeling of meteorological time-series parameters.

[0020] The input data of the meteorological time-series feature extraction module, that is, the meteorological time-series parameters, is a four-dimensional time-series matrix containing outdoor temperature (°C), wind speed (m / s), wind direction (°), and solar radiation intensity (W / m²) , representing a matrix with 24 rows and 4 columns, containing 4 kinds of meteorological parameter data for 24 hours.

[0021] In one embodiment of the present invention, it is also necessary to perform data preprocessing on the meteorological time-series parameters. In the data preprocessing stage, the dimensional difference is eliminated through dynamic standardization, and the diurnal cycle characteristics of the meteorological time-series parameters are retained: ; where represents the th meteorological time-series parameter after standardization processing, represents the original value of the th meteorological time-series parameter at the th moment, represents the mean value of the th meteorological time-series parameter in the historical data at the th moment, is the corresponding standard deviation, represents the th moment in 24 hours, , is a constant used to ensure numerical stability, is the th meteorological time-series parameter. In one embodiment of the present invention, = 1, 2, 3, 4, corresponding to 4 meteorological parameters of outdoor temperature, wind speed, wind direction and solar radiation intensity.

[0022] Data preprocessing makes parameters with different dimensions comparable, and at the same time can avoid eliminating real physical fluctuations during the standardization process.

[0023] The bidirectional long short-term memory network (BiLSTM) includes two independent LSTM units, namely the forward LSTM unit and the backward LSTM unit. The forward LSTM unit and the backward LSTM unit process meteorological time series parameters from the forward ( ) and backward ( ) respectively. Let be the hidden state of the previous time step, indicating that the dimension of the hidden state vector is 128. The cell state update process is implemented through a gating mechanism: ; ; ; ; Among them, is the output of the forget gate, controlling the influence degree of the previous cell state on the current cell state ; is the output of the input gate, determining the influence of the current input on the current cell state; is the current candidate cell state, a combination of the current input and the previous hidden state processed by the tanh activation function, used to generate the current cell state, is the Sigmoid activation function operation, are the weight matrices of the forget gate, input gate and candidate state respectively, represents the dimension of the weight matrix, are the corresponding bias terms respectively, represents concatenating the hidden state of the previous time step and the current input , represents the Hadamard product, is the current input. In one embodiment of the present invention, , that is, the 4 meteorological parameter data after data preprocessing at time represents a 4D vector.

[0024] Forward propagation And backward propagation The splicing forms 256-dimensional temporal features, and this structure simultaneously captures the forward cumulative effect and backward causality of meteorological temporal parameters. It is the final temporal feature output of the bidirectional long short-term memory network (BiLSTM), which is formed by splicing the outputs of LSTM units in two directions, namely the outputs of the forward LSTM unit and the backward LSTM unit, and its dimension is , indicating that at 24 time steps, the feature dimension at each time step is 256.

[0025] To enhance the feature extraction ability of key time nodes, at the final output end of the bidirectional long short-term memory network (BiLSTM), that is, at the temporal feature output , a multi-head self-attention mechanism is introduced (in one embodiment of the present invention, it is a 4-head self-attention mechanism). The query vector , key vector and value vector of each attention head in the multi-head self-attention mechanism are all generated through independent fully connected layers; the attention weight calculation adopts the scaled dot product method to generate the feature importance distribution in the time dimension; the outputs of the multi-head attention heads are processed by layer normalization after splicing, and a time-aware gating mechanism is introduced to finally generate a 128-dimensional meteorological feature vector . This design enables the multi-modal fusion neural network model to autonomously identify the impacts of key meteorological events such as strong radiation at noon and nocturnal inversion. The multi-head self-attention mechanism enhances the feature expression ability through multiple groups of parallel calculations.

[0026] Project the temporal feature output of the bidirectional long short-term memory network (BiLSTM) to a 64-dimensional subspace: ; Among them, , , are the query matrix, key matrix and value matrix respectively, is a learnable parameter matrix, represents its dimension. This operation means compressing the 256-dimensional LSTM features to a 64-dimensional subspace to achieve multi-head parallel calculations. The single-head attention calculation adopts the scaled dot product form: ; Among them, is the output matrix of the th attention head, with a dimension of , indicating the temporal feature pattern extracted by this attention head through dot product operations; , , are the query matrix, key matrix and value matrix of the th attention head respectively, Features for matching key time nodes; in is the matrix transpose operator, and the whole represents the transpose of the key matrix of the -th attention head, which is used to calculate the similarity with the query vector; carries the temporal feature information that needs to be aggregated. The scaling factor is used to control the order of magnitude of the dot product result and prevent gradient disappearance.

[0027] The multi-head output is concatenated and then undergoes linear transformation and residual connection: ; ; where represents the result of the linear transformation after the concatenation of the multi-head attention output. In one embodiment of the present invention, the dimension after concatenation is (each head is , and the four-head concatenation is ); is the weight matrix of the linear transformation, which is used to map the concatenated multi-head output back to the 256-dimensional space, represents 's dimension. Both the input and output are 256-dimensional to ensure consistent dimensions; is the output after residual connection and layer normalization. Add to the original , and then perform layer normalization to keep the dimension unchanged. Among them, the layer normalization operation accelerates the training convergence by standardizing the hidden layer activation values, and its calculation includes the learnable scaling parameter and the translation parameter (trainable parameters automatically learned during the model training process and automatically optimized through backpropagation).

[0028] The meteorological temporal feature extraction module dynamically adjusts the feature weights of key meteorological events by introducing a time-aware gating mechanism. Among them, the key meteorological events are the mutation points of meteorological states that have a greater impact on building environment parameters and are automatically identified through attention weights.

[0029] Define the time difference gate: ; ; where is the time difference gate coefficient, which generates channel attention weights through the Sigmoid function to dynamically adjust the importance of hidden state features; is the hidden state of the current time step output by the bidirectional long short-term memory network (BiLSTM), which represents the temporal dynamic features of the meteorological temporal parameters at the current moment; is the time difference between the current moment and the moment of peak solar radiation intensity, which is used to enhance the cycle sensitivity; is the gating weight matrix, represents an input dimension of 257 ( and The concatenation operation forms a 257-dimensional joint input vector) and an output dimension of 128, corresponding to the number of channels of the gating coefficient, which is aligned with the hidden layer dimension of the bidirectional long short-term memory network (BiLSTM) to achieve gating adjustment at the feature level; is the moment of peak solar radiation intensity.

[0030] In one embodiment of the present invention, the time-aware gating mechanism enables the multi-modal fusion neural network model to automatically enhance the radiation-related feature weights during the noon period ( ). The attention weight calculation of the key meteorological event is: ; where is the projection matrix, is the learnable parameter vector, is the temporal position (corresponding to the 24-hour system), is the temporal position of the hidden state, is the hidden state of the current time step.

[0031] Through this time-aware gating mechanism, the attention weights of key meteorological events such as morning and evening temperature inversions can be improved.

[0032] The time convolutional layer is a one-dimensional convolutional dimension , and a causal convolution with a time span of kernel_size = 24 is used to output a 24×128 temporal feature matrix, that is, a meteorological feature vector: ; The building geometry coding module is used to generate building geometry features by fusing the opening position and area parameters based on voxelization preprocessing, feature extraction, and dynamic graph convolution.

[0033] The building geometry coding module is based on an improved PointNet++ architecture to achieve multi-level geometric feature learning of building three-dimensional point cloud data. The building geometry coding module includes a voxelization preprocessing layer, a feature extraction network (including three-dimensional convolution), a dynamic graph convolution layer, an opening feature fusion layer, and a multi-scale feature fusion layer.

[0034] The input data of the building geometry coding module includes the coordinates of the original point cloud of the building atrium and the opening parameters , and the opening parameters include position coordinates and area parameters. Among them, is the number in the atrium space The coordinates of the position, is the opening area, represents the three-dimensional point cloud data matrix, represents the total number of points in the building atrium point cloud, is the th eigenvector of the opening, , is the number of all openings in the building atrium, is the opening parameter matrix.

[0035] The voxelization preprocessing layer is used to perform voxelization preprocessing on the original point cloud of the building atrium to convert the unordered point cloud into a structured feature field. The specific processing steps include: The input original point cloud is preprocessed by voxelization, dividing the atrium space into cubic grids with a preset voxel size (in one embodiment of the present invention, the preset voxel size is 1 m³). The center coordinates of each voxel are used as the initial features, and a multi-dimensional feature vector is output :

[0036] where is the voxel center coordinate, is the principal component of the normal vector, is the standard deviation of the point density, is the opening quantity statistic.

[0037] This voxelization preprocessing converts the unordered point cloud into a structured feature field , is the dimension of the structured feature field, , , respectively represent the number of divisions of the voxel grid in the length, width, and height directions, is the feature dimension of each voxel, corresponding to 's 7 feature components (3 + 3 + 1 + 1).

[0038] The feature extraction network of the first level is used to extract features from the structured feature field to generate primary geometric features.

[0039] The feature extraction network is implemented by a 1×1 convolutional kernel with 64 channels, and the primary geometric features are generated through batch normalization and the ReLU activation function : ; where, represents a three-dimensional convolution operation, BN is the batch normalization layer, is the convolutional kernel parameter of 64 channels. The feature extraction network outputs 64-dimensional primary geometric features . In is the spatial size of the three-dimensional convolution kernel (single-point convolution), is the number of input channels (corresponding to the voxel feature dimension), is the number of output channels (feature enhancement dimension), which is the primary geometric feature dimension extracted by convolution, is the spatial resolution.

[0040] The dynamic graph convolution layer is used for local geometric feature extraction and feature aggregation. Through dynamic graph convolution, the original point cloud of the building atrium is converted into 128-dimensional intermediate geometric features .

[0041] In the second level, a dynamic graph convolution operation is introduced. A local spherical domain with a radius of 2m is constructed centered at each position of the atrium to aggregate neighborhood features using max pooling.

[0042]

[0043] Where represents the neighborhood point set of is a network with two fully connected layers of 64 and 128, is the aggregated feature vector (128-dimensional) at the position, obtained by max pooling of neighborhood point features; is the neighborhood point set belonging to the -th neighborhood point of is the first 64-dimensional feature of the neighborhood point , extracted from .

[0044] For feature aggregation is performed to generate 128-dimensional intermediate geometric features .

[0045] The opening feature fusion layer is used to fuse the three-dimensional coordinates O in the opening position parameters with the area information, enhance the opening features, and splice them with the intermediate geometric features to obtain the channel-spliced features.

[0046] The influence range of each opening is modeled by a Gaussian kernel function: ; Where is the atrium position coordinate, is used to control the influence radius. In one embodiment of the present invention, .

[0047] After the weight distribution of all openings is normalized by Softmax, it is weighted and fused with the main network features (i.e., intermediate geometric features ), enabling the multi-modal fusion neural network model to accurately perceive the impact of the opening layout on the ventilation path.

[0048] Weighted fusion with the opening area after normalization: ; Among them, is the opening feature, is the opening index variable, traversing all openings, is the Gaussian weight at the position of the -th opening in the atrium coordinates, Encoding the opening coordinates into a 64-dimensional feature, i.e., the opening feature . This 64-dimensional feature is concatenated with the main network feature to obtain the channel concatenation feature : ; Among them, is the channel concatenation feature of the intermediate geometric feature and the opening feature , is to keep the spatial resolution unchanged, indicating that each voxel point in the three-dimensional space grid has a 192-dimensional feature (128-dimensional intermediate geometric feature plus 64-dimensional opening feature).

[0049] The multi-scale feature fusion layer is used to integrate the primary geometric feature, intermediate geometric feature, and channel concatenation feature to obtain the building geometric feature.

[0050] Among them, the multi-scale feature fusion layer integrates features of different scales through skip connections to obtain the spatial topological feature, i.e., the building geometric feature : ; Among them is the multi-scale feature fusion result after trilinear interpolation upsampling, indicating trilinear interpolation upsampling to a unified resolution, and Conv3D is a 3×3×3 convolution kernel with 256 channels.

[0051] The air conditioning system parameter encoding module is used to construct a non-linear mapping relationship from air conditioning operation parameters to a three-dimensional space field, mapping the cooling capacity and supply air temperature to a three-dimensional parameter field. The input parameter, i.e., the air conditioning operation parameter, is the cooling capacity (kW) and the supply air temperature (℃), the output is a control parameter field that matches the dimension of the building geometric feature field.

[0052] First, the physical dimensions of the input parameters are unified: ; in kW, kW is the design range of cooling capacity, ℃、 ℃ is the feasible region of the supply air temperature. This process eliminates the dimension difference of the parameters and constrains the input to the interval [0,1]; is the normalized value of cooling capacity, is the normalized value of the supply air temperature.

[0053] The nonlinear relationship between air conditioning operating parameters is learned through a three-layer fully connected network (16-32-64 neurons), and the output is a 64-dimensional feature vector , encoding the coupling effect of cooling power and thermodynamic parameters.

[0054]

[0055] 、 、 are the first layer hidden state, the second layer hidden state and the third layer output features respectively, 、 、 are the corresponding weight matrices, 、 、 are the corresponding bias terms respectively.

[0056] Using the spatial broadcast mechanism, the feature vector Expanded into a three-dimensional parameter field, i.e., air conditioning parameter characteristics The spatial broadcasting mechanism is to use bilinear interpolation to generate a continuous distribution in the XY plane and apply an exponential decay function in the vertical direction: ; ; ; in, It is the XY plane interpolation feature, a continuous distribution generated by bilinear interpolation; is the interpolation weight coefficient, is the grid spacing; Indicates spatial dimension, referring to the x or y direction; 、 are grid indices, indicating the row and column numbers of the nearest neighbor grid points; is the installation height of the air supply opening (m). is the total height of the atrium space (m). is the tensor dimension, a 64×64×64 spatial grid, with each point containing 64-dimensional features.

[0057] The cross-modal feature fusion module is used to perform spatial alignment and multi-modal interaction modeling on meteorological feature vectors, building geometric features, and air-conditioning parameter features. Through multi-scale feature matching and spatio-temporal correlation modeling, a three-dimensional spatial feature field that combines environmental dynamics (including meteorological fluctuations and air-conditioning parameter features) and building physical characteristics is obtained, providing a unified feature representation for jointly predicting ventilation volume and vertical temperature distribution.

[0058] The cross-modal feature fusion module includes five sub-modules: spatial alignment preprocessing, spatio-temporal attention mechanism, feature modulation and dynamic fusion, three-dimensional convolutional feature abstraction, and geometric consistency constraint. The spatial alignment preprocessing sub-module is used to uniformly align the three features to a 16×16×16 spatial grid through multi-resolution downsampling and feature compression, eliminating the dimensional differences of multi-source heterogeneous data and ensuring the point-by-point computability of features at the spatial voxel level. The spatio-temporal attention mechanism sub-module establishes a non-linear correlation between meteorological time-series parameters and three-dimensional spatial local features through time decay constraint, gradient stabilization calculation, and k-nearest neighbor sparse calculation, generates a spatio-temporal collaborative attention weight distribution, and enhances the weight interpretability through physical priors. The feature modulation and dynamic fusion sub-module maps building geometric features and air-conditioning parameter features to a unified latent space based on a dual projection network and uses a gating mechanism to dynamically adjust the contribution degrees of different modalities to achieve spatially adaptive feature fusion. The three-dimensional convolutional feature abstraction sub-module realizes feature compression and semantic enhancement through cascaded 3D convolutions (including residual connections and transposed convolutions), and finally outputs 16-channel decoupled features, retaining the implicit coding relationship of thermodynamics and fluid mechanics parameters. The geometric consistency constraint sub-module introduces a gradient alignment loss function, calculates the feature field gradient using a three-dimensional Sobel operator, prevents spatial topological contradictions caused by feature distortion, and ensures the credibility of physical laws.

[0059] To achieve the spatial computability of multi-source heterogeneous data, the cross-modal feature fusion module first performs multi-scale feature matching operations through the spatial alignment preprocessing sub-module to perform spatial alignment preprocessing on the input features (including meteorological feature vectors, building geometric features, and air-conditioning parameter features).

[0060] For the meteorological feature vector (including time-series encodings such as temperature, humidity, and wind speed), building geometric features and air-conditioning parameter features , a hierarchical downsampling strategy is adopted to achieve spatial dimension alignment.

[0061] Among them, the meteorological feature vector Generate a pseudo - three - dimensional tensor through time - dimension expansion and spatial broadcasting, that is, the aligned meteorological feature vector , the meteorological feature vector at each voxel position is updated synchronously with the time step.

[0062] Architectural geometric features Achieve spatial dimensionality reduction through a cascaded operation (cascading represents the sequential connection of two heterogeneous operations: spatial dimensionality reduction and channel compression).

[0063] Among them, the first layer performs average pooling downsampling to resolution, and the stride of the pooling kernel is set to 4 to maintain key spatial topological information: ; In the formula, represents the architectural geometric feature tensor after the first - layer average pooling, is a three - dimensional average pooling operation, calculating the average value of the eigenvalues within the local cube; kernel represents the pooling window size of 4x4x4 voxels, and stride represents the pooling sliding step, ensuring non - overlapping downsampling; represents a 16x16x16 spatial grid, and each position contains a 256 - dimensional feature vector.

[0064] The second layer uses a 1×1×1 convolutional kernel for channel compression to generate the aligned architectural geometric features . The 1x1x1 convolution operation compresses the number of feature channels from 256 to 64 while maintaining key geometric attributes such as the opening position and cavity connectivity.

[0065] Air - conditioning parameter features Achieve resolution matching through adaptive max - pooling: ;[[ID=4B]] Among them, is the aligned air - conditioning parameter feature, represents three - dimensional max - pooling, extracting the maximum value in the local area, represents the spatial grid, with 64 - dimensional features at each position.

[0066] The dynamic kernel size of the MaxPool3D layer is , the stride , and the output is the aligned air - conditioning parameter feature This processing ensures the computability of multi-source data in the spatial dimension. This processing ensures that the spatial distribution patterns of the supply air temperature and the cooling power are not blurred during the downsampling process, and realizes the dimensional alignment with other modality data. This stage enables each spatial voxel position of features from different sources to have direct comparability. This preprocessing ensures that different modality features have voxel-by-voxel alignment characteristics within a 16×16×16 grid space, laying a foundation for subsequent cross-modal interactions.

[0067] In this embodiment, spatio-temporal attention mechanism is used for spatio-temporal correlation modeling. The spatio-temporal attention mechanism sub-module is the core component for this multi-modal fusion neural network model to realize the coupling of meteorological time series dynamics and building space static features. Its core function is to establish the non-linear correlation between hourly meteorological fluctuations and the three-dimensional local physical field. This spatio-temporal attention mechanism can autonomously identify the influence intensity of meteorological time series parameters on different regions of the building through a learnable correlation matrix. Compared with traditional feature splicing methods, it can explicitly model the complex interaction among "time - space - physical quantity", enabling the fused features to have spatio-temporal collaborative perception ability.

[0068] The spatio-temporal attention mechanism adopts a multi-head parallel attention structure (in one embodiment of the present invention, a four-head parallel attention structure, where 4 heads achieve an optimal balance between the number of parameters and feature diversity, corresponding to four coupling modes of temperature conduction, momentum exchange, radiative heat flux, and pressure gradient respectively), and each attention head independently learns different physical coupling modes.

[0069] In one embodiment of the present invention, the spatio-temporal attention mechanism adopts a four-head parallel attention structure, and defines a group of learnable parameter matrices , projects the original 128-dimensional aligned meteorological feature vector into a 32-dimensional subspace. Each head calculates an independent correlation score: ; In the formula is the spatial feature projection of the th attention head, is a learnable projection matrix, is the meteorological feature vector at time , is the learnable correlation matrix of the th attention head, denotes the operation, which linearly projects the concatenated 128-dimensional spatial features into a 32-dimensional subspace.

[0070] The outputs of the four attention heads are aggregated through gated weighting: ; is the spatio-temporal position The final attention weight, gating coefficient is dynamically generated by a learnable parameter vector , which enables the multi-modal fusion neural network model to automatically allocate the importance of each attention head.

[0071] ; The time decay constraint: An exponential decay factor is introduced to strengthen the physical time series characteristics, and a time influence function is defined: ; where is the time decay factor, is the current prediction time, hours is the learnable decay coefficient. The corrected attention score is calculated as: ; where is the original attention score, the correlation coefficient without considering time decay, which is composed of the independently calculated correlation scores of each attention head spliced together. This time decay constraint enables the multi-modal fusion neural network model to conform to the law of thermodynamic inertia, and the decay coefficient is automatically optimized through gradient backpropagation.

[0072] The gradient stabilization calculation: To prevent gradient explosion in high-dimensional tensor operations, a hierarchical normalization strategy is designed. First, the scores of each attention head are zero-meaned: ; where is the normalized value of the score of the th attention head, is the mean of the scores of the th attention head within the current batch, is the mean calculation, is the original score matrix of the th attention head, is the standard deviation of the scores of the th attention head within the current batch, is the standard deviation calculation.

[0073] Then, an L2 norm constraint is imposed on the output of multiple attention heads to reduce the gradient variance during training. The normalization parameter is dynamically calculated in each training batch and does not participate in parameter updates.

[0074] ; where represents the final attention weight after standardization and gating weighting.

[0075] In addition, to improve the computational efficiency, sparse attention calculation based on k-nearest neighbors is performed.

[0076] First is the spatial domain sparsity. For each voxel position , only calculate the attention of the 8 meteorological time points with the closest Euclidean distance to it (the 8 time points belong to the hyperparameter setting, and the 8-nearest neighbors balance the computational cost and accuracy), and the threshold is adaptively adjusted: ;

[0077] Among them, is the set of valid time points of the voxel position , is the feature vector of the aligned geometric feature field at the voxel position , is the meteorological feature encoding at the moment, is the calculation of the expected value in the time dimension, is the cross-modal distance standard deviation between the meteorological feature and the geometric feature.

[0078] Then is the time domain segmentation. Divide 24 hours into 6 time periods (00-04, 04-08, ···, 20-24), and share the basic attention pattern within each time period. This optimization can reduce the computational complexity.

[0079] The physical interpretability is enhanced by introducing prior knowledge to constrain the attention weight distribution.

[0080] First is the radiation constraint: Force the lower limit of the attention weight on the sunny side during the noon period (11:00-13:00): ; Among them is the constraint strength coefficient, is the sunshine function calculated according to the building orientation.

[0081] Secondly is the ventilation constraint. When the wind speed at the opening , force the attention weight of the relevant time step . This prior injection can improve the prediction stability of the multi-modal fusion neural network model under extreme meteorological conditions and increase the spatial correlation coefficient value between the attention weight distribution and the CFD simulation results.

[0082] In the feature modulation and dynamic fusion sub-module, first project the multi-modal features (including the aligned air-conditioning parameter features and the aligned building geometric features ) to a unified latent space through a double projection network: ; ; Among them, and are respectively the projection weight matrices of the building geometric features and the air conditioning parameter features, and are respectively the corresponding bias terms, is the projection of the building geometric features in the latent space, is the projection of the air conditioning parameter features in the latent space.

[0083] Perform spatially adaptive feature fusion to obtain the fused feature vector, and the gating mechanism is introduced in the fusion process to dynamically adjust the modal contribution degree: ; ; Among them is the spatial gating coefficient, which controls the weights of the building geometric features and the air conditioning parameter features, is the gating weight matrix, is the corresponding bias term, is the Sigmoid function, which realizes the normalization of the gating coefficient, is the fused feature vector. This design can make the multi-modal fusion neural network model focus on the influence of the air conditioning parameter features in the area near the air supply outlet ( ), while strengthening the role of the building geometric features in the enclosed space ( ).

[0084] The three-dimensional convolutional feature abstraction sub-module realizes feature compression and semantic enhancement through a cascaded three-dimensional convolutional network: ; ; ; ; And the cascaded three-dimensional convolutional network uses residual connections to retain the shallow details: ; The cascaded three-dimensional convolutional network finally outputs a decoupled feature with 16 channels Each channel of which corresponds to the latent encoding of physical quantities such as temperature and flow velocity, and feature decoupling is achieved through 1×1 convolution between channels.

[0085] In the geometric consistency constraint sub-module, in order to maintain the spatial topological consistency between the fused features (decoupled features) and the building structure, the gradient alignment loss function is defined : ; In the formula represents the three-dimensional Sobel gradient operator, represents the decoupled feature at the voxel position ; represents the solid structure region, represents being masked the number of valid voxels marked as the structure region.

[0086] The gradient calculation adopts the central difference format: ; Among them, , , are three-dimensional grid coordinates, representing the voxel positions in the building space; and represent the eigenvalue of adjacent voxels in the x direction, used to calculate the horizontal gradient; is the grid spacing, determining the scale of gradient calculation. This constraint makes the characteristic gradient of boundaries such as walls and floors consistent with the original geometric field in direction, preventing feature distortion.

[0087] The vertical temperature field prediction module adopts a prediction architecture of the cooperation of basis function superposition and residual correction to realize the modeling of the vertical temperature field conforming to the thermodynamic law and obtain the vertical temperature distribution of the atrium.

[0088] The vertical temperature field prediction module includes a basis function parameter generation network, a residual correction network, and an interpolation optimization module.

[0089] The basis function parameter generation network adopts a dual-channel feature fusion architecture to generate basis function parameters , and the input decoupled feature is compressed into a 256-dimensional global feature vector through global average pooling, and at the same time, height-related features are extracted (20 height layers, 64 dimensions per layer).

[0090] Perform feature cross-projection: ; In the formula, represents the fusion feature vector formed by the cross-projection of the global feature vector and the height-related features, is the projection matrix, is the bias term in the feature cross-projection process, and LayerNorm realizes the normalization of the channel dimension to eliminate the dimension difference.

[0091] Secondly, perform dual-channel attention enhancement to calculate the response weights of each height layer to the global feature vector.

[0092] ; Among them, is the 64-dimensional feature vector of the th height layer, encoding the spatial thermodynamic characteristics of this height layer; is the feature vector of the th height layer, is the index variable traversing all height layers, is the total number of layers after discretization in the vertical direction of the building.

[0093] Next, calculate the local attention and construct a height neighborhood relationship graph (neighborhood radius = 2m): ; Among them, is the attention weight of the th height layer to the features of the th height layer in the neighborhood, reflecting the influence degree of the th height layer on the thermodynamic state of the th height layer; and are the absolute height values of the th height layer and the th height layer respectively, used to calculate the vertical spacing; is a multi-layer perceptron with the absolute value of the height difference as the input, and outputs intermediate features for generating attention weights; is to normalize the weights of all layers in the neighborhood to ensure , making the weights have a probabilistic interpretation.

[0094] Then perform the synthetic feature operation:

[0095] Among them, is the updated feature of the th height layer enhanced by global-local attention, fusing the global thermodynamic state and the local influence of neighboring height layers; is the set of neighboring height layers of the th height layer, reflecting the vertical heat conduction continuity.

[0096] Finally, perform parameter regression: ; Among them is the weight matrix parameter, is the cubic polynomial coefficient of the basis function, used to construct the vertical temperature distribution model, is the bias term of the fully connected operation of the regression layer. The output parameters satisfy physical constraints: 。

[0097] The residual correction network adopts a spatio-temporal joint modeling architecture. The residual correction network constructs a highly discretized feature space, divides the building space into 20 height layers at 1m intervals, and the temperature correction amount of each height layer is calculated by gated convolution.

[0098] First, perform three-dimensional gated convolution: ; where is the temperature residual correction amount of the th height layer, dynamically calculated from convolutional features; is the 64-dimensional feature vector corresponding to the th height layer, derived from the height-related features output by cross-modal fusion; is the concatenation of the feature tensors of the , , three adjacent height layers, forming a 3-layer feature slice for 3×3×3 convolution operation; is the weight matrix parameter, is the convolution kernel parameter, represents three-dimensional convolution operation.

[0099] Secondly, there is LSTM height propagation, constructing a bidirectional LSTM network to transfer the thermodynamic state along the height dimension: ; ; ; where, is the LSTM hidden state when propagating from the bottom layer to the top layer ( =1→20), containing the thermodynamic influence of the lower layer; is the LSTM hidden state when propagating from the top layer to the bottom layer ( =20→1), capturing the high-rise thermal pressure effect; and together constitute the input vector of the LSTM, transferring the change amount of the thermal state; is the weight matrix parameter; the hidden state dimension is set to 32, and the final gating coefficient controls the residual correction intensity.

[0100] Finally, add a physical constraint loss function, introducing a second-order differential operator to strengthen thermodynamic consistency: ; where is the physical consistency loss function, which constructs the energy conservation constraint through the second-order spatial derivative and the time derivative of the temperature field; is the height at the predicted temperature value; is at the moment the height at the predicted temperature value; is the air density, is the specific heat capacity at constant pressure of air, is the height interval, is the time step. This constraint makes the prediction result satisfy the unsteady heat conduction equation, so as to reduce the energy conservation error.

[0101] The interpolation optimization module performs cubic spline interpolation optimization.

[0102] To achieve continuous temperature field prediction with a resolution of 0.1 m, a piecewise cubic polynomial is constructed for 20 discrete nodes through cubic spline with constraint conditions: ; By applying the continuity condition , and the natural boundary condition , a tridiagonal system of equations is constructed to solve the interpolation coefficients.

[0103] Among them, is the cubic polynomial function defined on the interval; , , , is the th cubic polynomial coefficient of the interval segment; is the th height coordinate of the interpolation node; is the th segment of the polynomial at at the function value, which needs to be equal to the th segment starting value; is the th segment of the polynomial at at the first derivative continuity condition; and are natural boundary conditions, which force the second derivatives at the top and bottom to be zero to ensure the curve is smooth.

[0104] The ventilation volume prediction module adopts a prediction framework that combines fluid mechanics mechanism and deep learning to achieve accurate estimation of building opening flow.

[0105] The ventilation volume prediction module includes a graph neural network, a turbulence-corrected spatio-temporal convolutional network, a geometric-physical joint encoding network, and a physical constraint output layer. The graph neural network dynamically predicts the flow coefficient of each atrium opening by modeling the geometric topological relationship between building openings; the turbulence-corrected spatio-temporal convolutional network captures the spatio-temporal characteristics of vortex motion to compensate for the turbulence effect error of the traditional mechanism model; the geometric-physical joint encoding network uses an adaptive weight allocation mechanism to generate the dynamic fusion weights of the mechanism model and the data-driven model according to the opening shape and physical parameters; the physical constraint output layer ensures that the prediction results conform to the laws of fluid mechanics through dimensional consistency and area threshold suppression.

[0106] The input features include the opening geometric parameter matrix and the decoupled features output by the cross-modal feature fusion module and the environmental features obtained after processing . The opening geometric parameter matrix integrates the key geometric attributes such as the position coordinates and aspect ratio of the opening.

[0107] is from the decoupled features after cross-modal fusion through spatial alignment and feature sampling, according to the opening position information is extracted. The decoupled features output by cross-modal feature fusion implicitly encode the three-dimensional spatial environmental features. For each opening , first perform voxel coordinate mapping to convert the physical coordinates to the grid coordinate system , and then perform trilinear interpolation sampling to extract 8 voxel features in the neighborhood in the decoupled features to generate the environmental features of the opening position through distance-weighted interpolation: ; The weight is calculated as: ; where is the three-dimensional physical size of the building space, extracted from the building three-dimensional point cloud data, , and respectively represent the physical space sizes of each voxel grid in the x-axis, y-axis, and z-axis directions.

[0108] The prediction network of the flow coefficient adopts a three-level graph attention architecture to construct an opening geometric topology graph , where the node represents the building opening, Represents a set of graph nodes, where each node corresponds to a building opening; Represents a set of graph edges, indicating the spatial adjacency relationship between openings; the edge Satisfies , Is the first three columns of the geometric parameter matrix of the opening , that is, the three-dimensional spatial coordinates, used to calculate the Euclidean distance between openings, Is the first three columns of the geometric parameter matrix of the opening . First, initialize the node features: ; In the formula Is the feature encoding matrix, Represents the dimension of the weight matrix, with the input being 3D geometric features + 128D environmental features and the output being 32D, Is the environmental feature for cross-modal fusion, Is the dimension of the environmental feature vector output by cross-modal fusion, Is the initial feature vector of the node , generated by fusing geometric parameters and environmental features; The geometric parameter matrix of the opening , including attributes such as position coordinates and aspect ratio.

[0109] Then calculate the graph attention weights: ; Among them Is the normalized attention weight of node in the -th layer of the graph attention network for the neighborhood node ; Is the layer index of the graph attention network ( ); Is the natural exponential function, Is the operation of the leaky rectified linear unit activation function, and Are the feature vectors of the neighborhood nodes and in the -th layer respectively; Is the query, key-value projection matrix, Is the learnable parameter vector; Is the set of neighborhood nodes of the node , Is the neighborhood node index variable, Is the feature vector of the node in the -th layer.

[0110] Next, feature aggregation and update are performed. Each layer of the graph attention network uses the PReLU activation function: ; Among them, is the feature vector of the node at the th layer; ; is the feature vector of the node at the th layer; ; is the layer normalization operation, is the set of neighboring nodes of node ; is the neighboring node index variable; is the value projection matrix that maps the neighboring node feature to the attention space. Final output, constraining to be within the range of engineering experience: ; Among them, compresses the node feature to the interval through the Sigmoid function to generate the scaling coefficient of , is the feature vector of node after passing through 3 layers of the graph attention network; is the output layer weight matrix.

[0111] The mainstream prediction channel models the pressure difference-driven flow based on the improved Darcy-Weisbach equation, and its core formula is: ; Among them, is the predicted value of the theoretical volume flow rate of the opening , which is the ideal state flow rate calculated based on the Darcy-Weisbach equation. is the pressure difference between both sides of the opening , which is composed of the thermal pressure effect and the wind pressure, is the geometric area of the opening , which directly affects the flow cross-sectional area for flow rate calculation.

[0112] The turbulence-corrected spatio-temporal convolutional network includes a dilated convolutional layer, a feature extraction layer, a spatio-temporal compression layer, and an output layer.

[0113] The turbulence-corrected spatio-temporal convolutional network uses spatio-temporal three-dimensional convolution to capture the vortex motion characteristics, and the dilated convolutional layer uses the dilated convolutional structure to enhance the receptive field: ; Among them, represents the The predicted value of the turbulent correction flow rate at an opening; is the time offset parameter, is the time window radius; is a three-dimensional convolutional kernel parameter tensor with a dilation rate and, is the coordinate of a 7×7×7 convolutional kernel, and the dilation rate increases with the network depth; is the slice tensor of the spatio-temporal fusion feature field, is the time dimension offset, is the offset coordinate of the three-dimensional convolutional kernel in the spatial dimension.

[0114] First is the feature extraction layer. For the 16-channel → 32-channel convolution, a gating mechanism is adopted: ; where, is the gating coefficient matrix, is the gating convolutional kernel parameter, is the feature convolutional kernel parameter, is the output of the feature extraction layer.

[0115] Secondly is the spatio-temporal compression layer. For the 32-channel → 16-channel convolution, channel attention is introduced: , ; where, is the channel attention weight coefficient, which after normalization has a range and represents the importance of the th channel; is the feature channel index variable, with a value range , is the number of input channels, is the three-dimensional tensor of the th feature channel, with dimensions × × (height × width × time); is the volume parameter of the three-dimensional feature field, used for global average pooling calculation; is the output of the spatio-temporal compression layer.

[0116] Finally is the output layer. For the 16-channel → 4-channel convolution, residual connection ensures stable gradients: ; where, is the output of the turbulent correction flow rate after convolution processing; It is the trainable parameter matrix of the convolutional layer. By correcting the spatio-temporal convolutional network through turbulence, the multi-modal fusion neural network model can effectively capture the turbulence pulsation characteristics of 0.5 - 2.0 Hz, so as to improve the prediction accuracy under the condition of sudden wind speed change.

[0117] In the flow fusion stage, the geometric-physical joint encoding network adopts a geometric feature-sensitive adaptive weight allocation mechanism to generate dynamic fusion weights , and the adaptive weight allocation mechanism generates dynamic fusion weights through geometric feature encoding, multi-head attention aggregation and variational inference.

[0118] First, perform geometric feature encoding: ; Among them is a 16-dimensional opening geometric feature vector generated by a 3-layer MLP (16→8→4), represents the 16-dimensional space of the geometric feature vector; is the sum of the areas of all openings in the building atrium ( ), which is used to calculate the opening area ratio parameter; is the aspect ratio, is the angle between the opening normal and the main wind direction, and the MLP includes 16→8→4 dimensional compression. Secondly, perform physical feature interaction, adopt a 4-head attention mechanism, and calculate the turbulent flow interaction effect of neighboring openings : ; Among them, is the physical interaction feature vector aggregated by multi-head attention, encoding the turbulent transfer effect between openings; is the set of geometric parameters of the neighboring openings of the opening , defined as: , including geometric parameters such as the position and area of neighboring openings.

[0119] Finally, perform Bayesian optimization: ; Among them, is the expectation calculation of the variational distribution , is the mean in the normal distribution, is the corresponding standard deviation, represents the identity matrix; is the Sigmoid activation function, which constrains the linear combination to the interval; and are the geometric feature projection weight and the physical feature projection weight respectively, and the feature importance is automatically learned through variational inference.

[0120] Next, it is optimized through the evidence lower bound ELBO, and this mechanism enables dynamic fusion weights In a narrow opening (aspect ratio ), it focuses on turbulent correction ( ), and in a conventional opening (aspect ratio ), it focuses on the mechanism model ( ).

[0121] The multi-modal fusion neural network model generates the final predicted ventilation volume through collaborative calculations of multiple branches (including a graph neural network, a spatio-temporal convolutional network for turbulent correction, and a geometric-physical joint encoding network), and designs the physical constraint output layer to balance the physical dimension consistency and engineering application requirements.

[0122] For each opening , the physical constraint output layer imposes a dimension consistency constraint: ; where is the final predicted ventilation volume of the th opening, is the steepness coefficient, is the minimum effective opening area, and the Sigmoid function is used to suppress the non-physical flow oscillation of small openings. In the formula is the dynamic fusion weight, which dynamically adjusts the contribution ratio of the main channel mechanism model and the turbulent correction channel through the opening geometric characteristics.

[0123] In one embodiment of the present invention, a natural ventilation assessment system for a large atrium in an exhibition building with multi-modal fusion is provided to implement the method provided in the foregoing embodiment. Please refer to Figure 3 , and the system includes the following units: A data acquisition unit for acquiring multi-modal data, where the multi-modal data includes meteorological time series parameters, building three-dimensional point clouds, opening geometric parameters, and air conditioning cooling parameters; An evaluation unit for inputting the multi-modal data into the multi-modal fusion neural network model to obtain a natural ventilation assessment result, where the natural ventilation assessment result includes the vertical temperature distribution in the atrium and the ventilation volume of each opening in the atrium; Among them, the multi-modal fusion neural network model includes a meteorological time-series feature extraction module, a building geometry encoding module, an air-conditioning system parameter encoding module, a cross-modal feature fusion module, a vertical temperature field prediction module, and a ventilation volume prediction module. The meteorological time-series feature extraction module is used to extract the time-series dynamic features of meteorological time-series parameters through a bidirectional long short-term memory network and a multi-head self-attention mechanism, and obtain a meteorological feature vector. The building geometry encoding module is used to generate building geometry features by fusing the opening position and area parameters based on voxelization preprocessing, feature extraction, and dynamic graph convolution. The air-conditioning system parameter encoding module is used to map the air-conditioning operation parameters into a three-dimensional parameter field to obtain air-conditioning parameter features. The cross-modal feature fusion module is used to perform spatial alignment and multi-modal interaction modeling on the meteorological feature vector, building geometry features, and air-conditioning parameter features, perform feature compression and semantic enhancement to obtain fused features, and make the fused features consistent with the spatial topology of the building structure. The vertical temperature field prediction module is used to jointly implement vertical temperature field modeling through basis function superposition residual correction to obtain the vertical temperature distribution in the atrium. The ventilation volume prediction module is used to predict the ventilation volume of each opening.

[0124] In one embodiment of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method described in the foregoing embodiment are implemented.

[0125] In one embodiment of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in the foregoing embodiment are implemented.

[0126] Through spatio-temporal attention-gated coupled feature modeling, the embodiment of the present invention integrates a hybrid architecture of bidirectional LSTM time-series modeling and time decay gating, and for the first time realizes forward and backward causal association analysis in building environment prediction. By introducing a time difference gating and a solar radiation peak moment self-identification mechanism, the feature weights of key meteorological events are dynamically strengthened, and combined with the physical field projection technology of the multi-head self-attention mechanism, a quantitative characterization of the wind speed-radiation coupling effect is realized.

[0127] By establishing a geometric topology-aware three-dimensional point cloud dynamic encoding system, the embodiment of the present invention proposes an improved PointNet++ architecture for building cavities. Based on a three-level encoding paradigm of "voxelization-dynamic graph convolution-opening influence field fusion", by defining an opening influence intensity Gaussian kernel function and multi-scale feature fusion, the technical bottleneck that traditional point cloud processing cannot quantify ventilation paths is broken through, and geometric topology correlation modeling of irregular openings is supported.

[0128] In the embodiments of the present invention, a dynamic fusion framework driven by physical constraints is constructed, and a feature fusion framework jointly constrained by spatio-temporal attention and thermodynamic equations is established (the constructed bidirectional LSTM network transmits thermodynamic states along the height dimension). By introducing a physical consistency loss function, through sparse attention calculation based on k-nearest neighbors and enhanced physical interpretability, the deep fusion of the data-driven model and thermodynamic principles is realized.

[0129] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined in the present invention can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for evaluating the natural ventilation of a tall atrium in a convention and exhibition building with multimodal fusion, characterized in that, Including the steps: Obtain multimodal data, where the multimodal data includes meteorological time series parameters, building three-dimensional point clouds, opening geometric parameters, and air-conditioning cooling parameters; Input the multimodal data into a multimodal fusion neural network model to obtain a natural ventilation evaluation result, where the natural ventilation evaluation result includes the vertical temperature distribution in the atrium and the ventilation volume of each opening in the atrium; Among them, the multimodal fusion neural network model includes a meteorological time series feature extraction module, a building geometry encoding module, an air-conditioning system parameter encoding module, a cross-modal feature fusion module, a vertical temperature field prediction module, and a ventilation volume prediction module; the cross-modal feature fusion module is used to perform spatial alignment and multimodal interaction modeling on the meteorological feature vector, building geometry feature, and air-conditioning parameter feature, perform feature compression and semantic enhancement to obtain a fusion feature, and make the fusion feature consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to jointly implement vertical temperature field modeling through basis function superposition residual correction to obtain the vertical temperature distribution in the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening.

2. A method for evaluating the natural ventilation of a large atrium in a convention and exhibition building with multimodal fusion according to claim 1, characterized in that The meteorological time series feature extraction module is used to extract the temporal dynamic features of meteorological time series parameters through a bidirectional long short-term memory network and a multi-head self-attention mechanism to obtain a meteorological feature vector; the building geometry encoding module is used to generate a building geometry feature by fusing the opening position and area parameters based on voxelization preprocessing, feature extraction, and dynamic graph convolution; The air-conditioning system parameter encoding module is used to map the air-conditioning operation parameters into a three-dimensional parameter field to obtain an air-conditioning parameter feature.

3. A method for evaluating natural ventilation of a tall atrium in a convention and exhibition building with multimodal fusion according to claim 1, characterized in that, In the meteorological time series feature extraction module, the bidirectional long short-term memory network includes a forward LSTM unit and a backward LSTM unit, which respectively process the meteorological time series parameters from the forward and backward directions, capture the forward cumulative effect and backward causal relationship of the meteorological time series parameters, and obtain a temporal feature output; Temporal The feature output is input into the multi-head self-attention mechanism to enhance the feature extraction ability of key time nodes, and the attention weight of key meteorological events is dynamically adjusted through a time-aware gating mechanism; The air-conditioning system parameter encoding module is configured to perform the following operations: Learn the non-linear relationship between air-conditioning operation parameters to obtain a feature vector; Use the spatial broadcasting mechanism to expand the feature vector into a three-dimensional parameter field to obtain an air-conditioning parameter feature.

4. A method for evaluating the natural ventilation of a tall atrium in a convention and exhibition building with multimodal fusion according to claim 1, characterized in that, The building geometry encoding module includes a voxelization preprocessing layer, a feature extraction network, a dynamic graph convolution layer, an opening feature fusion layer, and a multi-scale feature fusion layer. The voxelization preprocessing layer is used to perform voxelization preprocessing on the original point cloud of the building atrium to convert the unordered point cloud into a structured feature field; the feature extraction network is used to extract features from the structured feature field to generate primary geometric features; the dynamic graph convolution layer is used to perform local geometric feature extraction and feature aggregation, and through dynamic graph convolution, convert the original point cloud of the building atrium into intermediate geometric features; The opening feature fusion layer is used to fuse the three-dimensional coordinates and area information of the opening position parameters and splice them with the intermediate geometric features to obtain a channel splicing feature; The multi-scale feature fusion layer is used to integrate the primary geometric features, intermediate geometric features, and channel splicing features to obtain a building geometry feature.

5. The natural ventilation evaluation method for the tall atrium of a convention and exhibition building with multimodal fusion according to claim 4, characterized in that, The cross-modal feature fusion module includes: A spatial alignment preprocessing sub-module for performing spatial alignment preprocessing on meteorological feature vectors, building geometric features, and air-conditioning parameter features; A spatio-temporal attention mechanism sub-module for establishing a non-linear association between meteorological time-series parameters and three-dimensional spatial local features through time decay constraint, gradient stabilization calculation, and k-nearest neighbor sparse calculation, generating a spatio-temporal collaborative attention weight distribution, and enhancing the weight interpretability through physical priors; A feature modulation and dynamic fusion sub-module for mapping building geometric features and air-conditioning parameter features to a unified latent space based on a dual projection network, and dynamically adjusting the contribution degrees of different modalities using a gating mechanism to achieve spatially adaptive feature fusion; A three-dimensional convolutional feature abstraction sub-module for performing feature compression and semantic enhancement through cascaded convolutions, outputting fused features, and retaining the implicit coding relationship of thermodynamics and fluid mechanics parameters; A geometric consistency constraint sub-module for introducing a gradient alignment loss function to maintain the spatial topological consistency between the fused features and the building structure, and calculating the feature field gradient.

6. The natural ventilation assessment method for the tall atrium of a convention and exhibition building with multimodal fusion according to claim 1, characterized in that The vertical temperature field prediction module includes: A basis function parameter generation network for performing feature cross-projection, attention enhancement, and synthetic feature operations on global feature vectors and height-related features, and performing parameter regression to generate basis function parameters for constructing a vertical temperature distribution model; A residual correction network for performing residual correction on the temperature by gated convolution, transmitting the thermodynamic state along the height dimension using a bidirectional LSTM network, and introducing a physical constraint loss function.

7. A method for evaluating the natural ventilation of a tall atrium in a convention and exhibition building with multimodal fusion according to any one of claims 1-6, characterized in that, The ventilation volume prediction module includes: ​ ​ ​ ​ 8. A natural ventilation evaluation system for the tall atrium of a convention and exhibition building with multimodal fusion, characterized in that, ​ ​ ​ Among them, the multi-modal fusion neural network model includes a meteorological time-series feature extraction module, a building geometry coding module, an air-conditioning system parameter coding module, a cross-modal feature fusion module, a vertical temperature field prediction module, and a ventilation volume prediction module. The meteorological time-series feature extraction module is used to extract the time-series dynamic features of meteorological time-series parameters through a bidirectional long short-term memory network and a multi-head self-attention mechanism to obtain a meteorological feature vector; the building geometry coding module is used to generate building geometry features by fusing the opening position and area parameters based on voxelization preprocessing, feature extraction, and dynamic graph convolution; the air-conditioning system parameter coding module is used to map the air-conditioning operation parameters into a three-dimensional parameter field to obtain air-conditioning parameter features; the cross-modal feature fusion module is used to perform spatial alignment and multi-modal interaction modeling on the meteorological feature vector, building geometry features, and air-conditioning parameter features, perform feature compression and semantic enhancement to obtain fused features, and make the fused features consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to jointly implement vertical temperature field modeling through basis function superposition residual correction to obtain the vertical temperature distribution in the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, the method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the method described in any one of claims 1 to 7 is implemented.

Images