A multimodal fusion method for evaluating natural ventilation in tall atriums of exhibition buildings

Abstract

The present invention discloses a multimodal fusion method for assessing the natural ventilation of a large atrium in an exhibition building. The method comprises: acquiring multimodal data, the multimodal data comprising meteorological time series parameters, a three-dimensional building point cloud and opening geometry parameters, and air conditioning and cooling parameters; inputting the multimodal data into a multimodal fusion neural network model to obtain a natural ventilation assessment result, the natural ventilation assessment result comprising the vertical temperature distribution of the atrium and the ventilation volume of each opening in the atrium; wherein the multimodal fusion neural network model comprises 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 present invention establishes a spatiotemporal collaborative perception mechanism for meteorological time series dynamics, building space topology, and air conditioning parameter features, breaking through the technical bottlenecks of traditional methods in terms of cross-modal feature space alignment and interaction efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of building ventilation prediction, and in particular relates to a multi-modal fusion natural ventilation assessment method for tall atriums of exhibition buildings. Background Art

[0002] Current 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 it still has significant limitations in multi-physics field coupling modeling and engineering practicality.

[0003] Traditional hybrid models often employ static parameterized mechanism equations and global average turbulence correction schemes, making them ineffective in addressing the spatial heterogeneity of flow coefficients caused by the geometric diversity of building openings. This is particularly true for modeling the unsteady characteristics of vortex motion within complex atriums. Existing methods generally employ the Reynolds-averaged Navier-Stokes equations combined with the k-ε turbulence model, which takes hours to compute and struggles to meet real-time control requirements. While data-driven approaches have improved time series feature extraction capabilities through networks such as LSTM and CNN, most studies have neglected to explicitly link meteorological time series dynamics with three-dimensional spatial topology, resulting in inaccurate modeling of the spatiotemporal coupling of key physical phenomena (such as the heat pressure-driven chimney effect and wind pressure-dominated through-drafts). Existing fusion models often employ simple concatenation or weighted averaging for cross-modal feature alignment, failing to effectively address the spatial scale mismatch between building point cloud data and air conditioning parameter features. In particular, significant errors are observed in capturing sudden thermal gradient changes near air outlets. These limitations severely restrict the ability of intelligent building control systems to accurately control ventilation volume in complex environments, and there is an urgent need to develop a new fusion prediction framework that takes into account both physical explainability and data-driven advantages. Summary of the Invention

[0004] In order to solve at least one of the problems existing in the prior art, the present invention provides a multi-modal fusion natural ventilation evaluation method for tall atriums in exhibition buildings, which provides a new direction and technical solution for the optimal design of atrium spaces.

[0005] The present invention provides a multi-modal fusion natural ventilation assessment method for a tall atrium in an exhibition building, comprising the following steps:

[0006] Acquiring multimodal data, the multimodal data including meteorological time series parameters, building three-dimensional point cloud and opening geometric parameters, and air conditioning and cooling parameters;

[0007] Inputting the multimodal data into a multimodal fusion neural network model to obtain a natural ventilation assessment result, wherein the natural ventilation assessment result includes the vertical temperature distribution of the atrium and the ventilation volume of each opening in the atrium;

[0008] 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 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 opening position and area parameters based on voxel preprocessing, feature extraction and dynamic graph convolution; the air-conditioning system parameter encoding module is used to map the air-conditioning operating parameters into a three-dimensional parameter field to obtain air-conditioning parameter features; the cross-modal feature fusion module is used to spatially align the meteorological feature vector, building geometry features and air-conditioning parameter features and perform multimodal interaction modeling, perform feature compression and semantic enhancement, obtain fusion features, and make the fusion features consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to collaboratively realize vertical temperature field modeling through basis function superposition residual correction to obtain the vertical temperature distribution of the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening.

[0009] The present invention also provides a multi-modal fusion natural ventilation evaluation system for a large atrium in an exhibition building, comprising:

[0010] A data acquisition unit, configured to acquire multimodal data, wherein the multimodal data includes meteorological time series parameters, building three-dimensional point clouds and opening geometric parameters, and air conditioning and cooling parameters;

[0011] The evaluation unit is used to input the multimodal data into the multimodal fusion neural network model to obtain a natural ventilation evaluation result, wherein the natural ventilation evaluation result includes the vertical temperature distribution of the atrium and the ventilation volume of each opening in the atrium.

[0012] The present invention also provides a computer device, wherein the processor thereof implements the multimodal fusion natural ventilation evaluation method for tall atriums of exhibition buildings when executing a computer program.

[0013] The present invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the multimodal fusion method for evaluating natural ventilation in a tall atrium of an exhibition building is implemented.

[0014] Compared with the existing technology, this invention has achieved multi-dimensional technological innovation in the field of building ventilation prediction. Its core advantages are reflected in the following aspects:

[0015] (1) By constructing a multimodal fusion architecture of bidirectional LSTM-attention collaborative encoding and 3D point cloud geometry learning, an innovative spatiotemporal collaborative perception mechanism of meteorological time series dynamics, building space topology, and air conditioning parameter characteristics was established, breaking through the technical bottlenecks of traditional methods in cross-modal feature space alignment and interaction efficiency;

[0016] (2) A geometric topology-sensitive graph neural network is proposed, which uses a dynamic neighborhood to construct a multi-head attention aggregation strategy to achieve refined modeling of the spatial layout of building openings and turbulence interference effects. Compared with conventional CNN encoding methods, the ability to analyze the connectivity characteristics of complex cavities is improved;

[0017] (3) A dynamic fusion framework for Bayesian optimization is constructed, which generates adaptive weights for the mechanism model and data-driven branches through variational inference. While retaining the physical core of the Darcy-Weisbach equation, the prediction robustness of the model under critical pressure difference conditions is significantly enhanced.

[0018] (4) Through the collaborative calculation of spatiotemporal dilation convolution and sparse attention, and through adjustable receptive field and k-nearest neighbor optimization, the computational complexity is reduced while maintaining the accuracy of capturing turbulent pulsation characteristics, providing technical support for intelligent control. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flowchart of the steps of a multi-modal fusion natural ventilation assessment method for a tall atrium of an exhibition building provided by an embodiment of the present invention.

[0020] Figure 2 Schematic diagram of the composition of the multimodal fusion neural network model in an embodiment of the present invention.

[0021] Figure 3 Schematic diagram of the composition of a multimodal fusion natural ventilation evaluation system for a tall atrium in an exhibition building according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] This embodiment addresses the existing shortcomings of multi-physics coupling modeling and cross-modal feature fusion in the field of building ventilation volume prediction, focusing on overcoming key technical challenges in heterogeneous data collaborative computing and physical constraint embedding. It primarily addresses the shortcomings of existing hybrid models in terms of dynamic environmental adaptability and spatial topology perception. First, the static parameterized equations in traditional mechanism-data fusion architectures struggle to represent the spatial heterogeneity of flow coefficients caused by the diversity of opening geometry, and lack graph-structured modeling of turbulent interference between adjacent openings. Second, the spatial-temporal-physical mismatch of multi-source heterogeneous data (meteorological time series, building point clouds, and air conditioning parameters) leads to inefficient cross-modal feature interaction, making existing methods unable to effectively establish nonlinear correlations between hourly meteorological fluctuations and local physical fields in three-dimensional space. Third, conventional fusion strategies rely on empirical weights to adjust the contribution of mechanism models and neural networks, failing to achieve adaptive allocation based on geometric features such as opening aspect ratios and azimuth angles, resulting in insufficient stability in flow direction prediction in critical pressure difference regions. Fourth, the weak constraints on thermodynamic conservation laws and dimensional consistency in existing deep learning models can easily lead to nonphysical oscillation solutions for small openings or sudden boundary conditions. The embodiments of the present invention achieve a synergistic improvement in the physical rationality of ventilation volume prediction and the generalization ability of working conditions in complex building environments by constructing a geometric topology-sensitive graph attention network, a spatiotemporal collaborative cross-modal interaction mechanism, a Bayesian optimized dynamic fusion framework, and a feature space constrained by differential equations.

[0024] See also Figure 1 This embodiment provides a multi-modal fusion method for evaluating the natural ventilation of a large atrium in an exhibition building, comprising the following steps:

[0025] Step 1: Obtain input data, which includes meteorological time series parameters, building three-dimensional point cloud and opening geometric parameters, and air conditioning and cooling parameters.

[0026] In one embodiment of the present invention, the three types of core input data required include:

[0027] (1) Meteorological time series parameters: total solar radiation intensity, outdoor temperature, outdoor wind speed and direction;

[0028] (2) Architectural 3D point cloud and opening geometric parameters: exhibition building form (used for CFD simulation to obtain wind pressure on the building's exterior surface), atrium opening position, atrium opening area, atrium skylight position, atrium skylight area, atrium internal form (3D form), collect architectural 3D point cloud data as architectural form input;

[0029] (3) Air conditioning cooling parameters: stratified air conditioning cooling capacity.

[0030] Step 2: Input the multimodal data into the multimodal fusion neural network model to obtain a natural ventilation evaluation result, which includes the vertical temperature distribution of the atrium and the ventilation volume of each opening in the atrium.

[0031] The multimodal fusion neural network model constructed in this embodiment is a deep neural network architecture that integrates multimodal spatiotemporal features. Figure 2 Through modules such as meteorological time series feature extraction, building geometry encoding, cross-modal feature fusion, vertical temperature field prediction, and ventilation volume prediction, the system achieves joint prediction of atrium opening ventilation volume and vertical temperature distribution. Based on meteorological time series feature extraction, building geometry encoding, and air conditioning parameter fusion, a dual-task decoder generates high-precision prediction results, providing key environmental parameter prediction capabilities for intelligent building control systems.

[0032] The meteorological time series feature extraction module is used to extract the temporal dynamic features of meteorological time series parameters through the bidirectional long short-term memory network (BiLSTM) and the multi-head self-attention mechanism, obtain the meteorological feature vector, and realize the temporal feature modeling of meteorological time series parameters.

[0033] The input data of the meteorological time series feature extraction module, namely the meteorological time series parameters, include outdoor temperature (℃), wind speed (m / s), wind direction (°) and solar radiation intensity (W / m²)4D time series matrix , represents a matrix with 24 rows and 4 columns, Contains 4 types of meteorological parameter data for 24 hours.

[0034] In one embodiment of the present invention, data preprocessing is also required for the meteorological time series parameters. In the data preprocessing stage, dynamic standardization is used to eliminate dimensional differences and retain the diurnal cycle characteristics of the meteorological time series parameters:

[0035] ;

[0036] in, After normalization, Meteorological time series parameters, Indicates The moment The original values ​​of meteorological time series parameters, Indicates the historical data Meteorological time series parameters are The mean of the time, is the corresponding standard deviation, Indicates 24 hours time, , is a constant used to ensure numerical stability. For the meteorological time series parameters, in one embodiment of the present invention, =1,2,3,4, corresponding to the four meteorological parameters of outdoor temperature, wind speed, wind direction and solar radiation intensity.

[0037] Data preprocessing makes parameters of different dimensions comparable and avoids the standardization process from eliminating real physical fluctuations.

[0038] Bidirectional long short-term memory network (BiLSTM) consists of two independent LSTM units, namely forward LSTM unit and backward LSTM unit, which are respectively from the forward ( ) and backward ( ) to process meteorological time series parameters. is the hidden state of the previous time step, The dimension of the hidden state vector is 128, and the unit state update process is implemented through a gating mechanism:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] in, Is the output of the forget gate, controlling the state of the previous unit Current unit status the extent of the impact; Is the output of the input gate, which determines the current input Impact on the current unit state; is the current candidate cell state, which is a combination of the current input and the previous hidden state processed by the tanh activation function to generate the current cell state. It 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, Indicates that the hidden state of the previous time step and the current input Perform splicing operations. represents the Hadamard product, is the current input. In one embodiment of the present invention, ,Right now The 4 meteorological parameter data after data preprocessing at this moment, Represents a 4-dimensional vector.

[0044] Forward propagation and backpropagation The splicing forms a 256-dimensional time series feature, which simultaneously captures the forward cumulative effect and backward causal relationship of meteorological time series parameters. is the final temporal feature output of the bidirectional long short-term memory network (BiLSTM), which is formed by concatenating the outputs of the LSTM units in two directions, namely the outputs of the forward LSTM unit and the backward LSTM unit, and its dimension is , which means that at 24 time steps, the feature dimension of each time step is 256.

[0045] In order to enhance the feature extraction capability of key time nodes, the final output of the Bidirectional Long Short-Term Memory Network (BiLSTM) is used to output the time series features. A multi-head self-attention mechanism is introduced (in one embodiment of the present invention, a 4-head self-attention mechanism). The query vector of each attention head in the multi-head self-attention mechanism is , key vector Sum vector All are generated through independent fully connected layers; the attention weight calculation uses the scaled dot product method to generate the feature importance distribution in the time dimension; the output of the multi-head attention head is normalized after splicing, and the time-aware gating mechanism is introduced to finally generate a 128-dimensional meteorological feature vector This design enables the multimodal fusion neural network model to autonomously identify the impact of key meteorological events such as strong midday radiation and nighttime temperature inversion. The multi-head self-attention mechanism enhances feature expression capabilities through multiple groups of parallel computations.

[0046] Project the temporal feature output of the bidirectional long short-term memory network (BiLSTM) into a 64-dimensional subspace:

[0047] ;

[0048] in, 、 、 are query matrix, key matrix and value matrix respectively, is the learnable parameter matrix, Indicates its dimension. This operation compresses the 256-dimensional LSTM feature into a 64-dimensional subspace to achieve multi-head parallel computing. Single-head attention calculation uses the scaled dot product form:

[0049] ;

[0050] in, For the The output matrix of the attention head has the dimension , represents the temporal feature pattern extracted by the attention head through the dot product operation; , , Respectively The query matrix, key matrix, and value matrix of each attention head, Features used to match key time nodes; in is the matrix transpose operator, and the overall expression is The key matrix of the attention head is transposed to calculate the similarity with the query vector; Carries the time series feature information that needs to be aggregated. Scaling factor Used to control the magnitude of the dot product result and prevent the gradient from disappearing.

[0051] After multiple outputs are concatenated, they are linearly transformed and connected with the residuals:

[0052] ;

[0053] ;

[0054] in, represents the result of linear transformation after concatenating the multi-head attention output. In one embodiment of the present invention, the dimension after concatenation is (Each head is , four-head splicing is ); is the weight matrix of the linear transformation, which is used to map the spliced ​​multi-head output back to the 256-dimensional space. express The dimensions of input and output are both 256-dimensional, ensuring the consistency of dimensions; is the output after residual connection and layer normalization, With the original Add, then perform layer normalization to keep the dimension unchanged. The layer normalization operation accelerates training convergence by standardizing the hidden layer activation values. Its calculation includes a learnable scaling parameter and translation parameters (The trainable parameters automatically learned during the model training process are automatically optimized through backpropagation).

[0055] The meteorological time series feature extraction module dynamically adjusts the feature weights of key meteorological events by introducing a time-aware gating mechanism. Key meteorological events are meteorological state mutation points that have a significant impact on building environmental parameters and are automatically identified through attention weights.

[0056] Define time difference gating:

[0057] ;

[0058] ;

[0059] in, is the time difference gating coefficient, which generates the channel attention weight through the Sigmoid function to dynamically adjust the importance of the hidden state features; The hidden state of the current time step output by the Bidirectional Long Short-Term Memory Network (BiLSTM) represents the temporal dynamic characteristics of the meteorological time series parameters at the current moment; It is the time difference between the current time and the peak time of solar radiation intensity, which is used to enhance period sensitivity; is the gating weight matrix, Indicates input dimension 257 ( and The concatenation operation forms a 257-dimensional joint input vector and an output dimension of 128, which corresponds to the number of channels of the gating coefficients and is aligned with the hidden layer dimensions of the BiLSTM network to achieve feature-level gating adjustment. The moment of peak solar radiation intensity.

[0060] In one embodiment of the present invention, the time-aware gating mechanism enables the multimodal fusion neural network model to automatically enhance the midday period ( The attention weight of key meteorological events is calculated as:

[0061] ;

[0062] in is the projection matrix, is the learnable parameter vector, is the time position (corresponding to the 24-hour system), is the timing position The hidden state of is the hidden state at the current time step.

[0063] This time-aware gating mechanism can enhance the attention weight of key meteorological events such as morning and evening inversions.

[0064] The temporal convolution layer has a one-dimensional convolution dimension , using causal convolution with a time span of kernel_size=24, outputs a 24×128 time series feature matrix, i.e., the meteorological feature vector:

[0065] ;

[0066] The building geometry encoding module is used to generate building geometric features by fusing opening position and area parameters based on voxel preprocessing, feature extraction and dynamic graph convolution.

[0067] The building geometry encoding module, based on an improved PointNet++ architecture, implements multi-level geometric feature learning for 3D building point cloud data. It includes a voxelization preprocessing layer, a feature extraction network (including 3D convolution), a dynamic graph convolution layer, an opening feature fusion layer, and a multi-scale feature fusion layer.

[0068] The input data of the building geometry encoding module contains the coordinates of the original point cloud of the building atrium and opening parameters , the opening parameters include position coordinates and area parameters, where The first The coordinates of the location, is the opening area, represents the 3D point cloud data matrix, Indicates the total number of points in the building atrium point cloud, For the The eigenvectors of the openings, , is the number of all openings in the building's atrium, is the opening parameter matrix.

[0069] The voxelization preprocessing layer is used to perform voxelization preprocessing on the original point cloud of the building atrium to convert the disordered point cloud into a structured feature field. The processing steps include:

[0070] The input original point cloud is pre-processed by voxelization, and the atrium space is divided into a cubic grid with a preset voxel size (in one embodiment of the present invention, the preset voxel size is 1m³). The coordinates of the center point of each voxel are used as the initial feature, and a multi-dimensional feature vector is output. :

[0071]

[0072] in are the voxel center coordinates, is the principal component of the normal vector, is the point density standard deviation, is the statistic of the number of openings.

[0073] This voxelization preprocessing transforms the disordered point cloud into a structured feature field , is the dimension of structured feature field, 、 、 Represents the number of divisions of the voxel grid in the length, width, and height directions, is the feature dimension of each voxel, corresponding to The 7 characteristic components of (3+3+1+1).

[0074] The first-level feature extraction network is used to extract features from the structured feature field and generate primary geometric features.

[0075] The feature extraction network is implemented by a 64-channel 1×1 convolution kernel, which generates primary geometric features through batch normalization and ReLU activation function. :

[0076] ;

[0077] in, Represents a three-dimensional convolution operation, BN is a batch normalization layer, is the convolution kernel parameter of 64 channels. The feature extraction network outputs 64-dimensional primary geometric features . middle, is the spatial size of the 3D convolution kernel (single point convolution), is the number of input channels (corresponding to voxel feature dimensions), is the number of output channels (feature enhancement dimension), which is the primary geometric feature dimension extracted by convolution. is the spatial resolution.

[0078] The dynamic graph convolution layer is used to extract local geometric features and aggregate features. Through dynamic graph convolution, the original point cloud of the building atrium is converted into 128-dimensional intermediate geometric features. .

[0079] Introducing dynamic graph convolution operation in the second level, each position of the atrium A local sphere with a radius of 2m is constructed as the center, and the neighborhood features are aggregated using maximum pooling.

[0080]

[0081] in express The neighborhood point set of It is a fully connected network with 64 and 128 layers. For the The aggregate feature vector of the position (128 dimensions) is obtained by max pooling of neighborhood point features; For Neighborhood point set No. Neighborhood points; Neighborhood points The first layer of 64-dimensional features, from Extracted.

[0082] right Perform feature aggregation to generate 128-dimensional intermediate geometric features .

[0083] The opening feature fusion layer is used to transform the opening position parameters O Three-dimensional coordinates It is fused with area information to enhance the opening feature and is spliced ​​with intermediate geometric features to obtain channel splicing features.

[0084] The influence range of each opening is modeled by a Gaussian kernel function:

[0085] ;

[0086] in is the coordinate of the atrium position, For controlling the influence radius, in one embodiment of the present invention, .

[0087] The weight distribution of all openings is normalized by Softmax and compared with the main network features (i.e., the intermediate geometric features ) are weighted fused so that the multimodal fusion neural network model can accurately perceive the impact of the opening layout on the ventilation path.

[0088] After normalization, weighted fusion with opening area is performed:

[0089] ;

[0090] in, is the opening feature, is the opening index variable, traverse all An opening, For the Gaussian weights of the openings at the atrium coordinates, The opening coordinates Encoded as 64-dimensional features, i.e., opening features The 64-dimensional feature is spliced ​​with the main network feature to obtain the channel splicing feature :

[0091] ;

[0092] in, Intermediate geometric features With opening features The channel splicing characteristics of To maintain spatial resolution It remains unchanged, indicating that each voxel point in the three-dimensional space grid has 192-dimensional features (128-dimensional mid-level geometric features plus 64-dimensional features of the opening).

[0093] The multi-scale feature fusion layer is used to integrate primary geometric features, intermediate geometric features and channel splicing features to obtain building geometric features.

[0094] Among them, the multi-scale feature fusion layer integrates features of different scales through jump connections to obtain spatial topological features, namely building geometric features :

[0095] ;

[0096] in It is the multi-scale feature fusion result after trilinear interpolation upsampling. Indicates trilinear interpolation upsampling to a uniform resolution, and Conv3D is a 3×3×3 convolution kernel with 256 channels.

[0097] The air conditioning system parameter encoding module is used to construct a nonlinear mapping relationship between air conditioning operating parameters and three-dimensional space fields, mapping cooling capacity and supply air temperature into a three-dimensional parameter field. The input parameter is the air conditioning operating parameter, cooling capacity. (kW) and supply air temperature (℃), the output is a control parameter field that matches the dimension of the building geometric feature field.

[0098] First, the physical dimensions of the input parameters are unified:

[0099] ;

[0100] 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.

[0101] 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.

[0102]

[0103] 、 、 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.

[0104] 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:

[0105] ;

[0106] ;

[0107] ;

[0108] 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 outlet (m), is the total height of the atrium space (m), It is the tensor dimension, 64×64×64 spatial grid, and each point contains 64-dimensional features.

[0109] The cross-modal feature fusion module is used to spatially align meteorological feature vectors, building geometric features, and air-conditioning parameter features and perform multimodal interaction modeling. Through multi-scale feature matching and spatiotemporal correlation modeling, a three-dimensional spatial feature field that integrates environmental dynamics (including meteorological fluctuations and air-conditioning parameter features) with building physical properties is obtained, providing a unified feature representation for the joint prediction of ventilation volume and vertical temperature distribution.

[0110] The cross-modal feature fusion module consists of five submodules: spatial alignment preprocessing, spatiotemporal attention mechanism, feature modulation and dynamic fusion, 3D convolutional feature abstraction, and geometric consistency constraints. The spatial alignment preprocessing submodule aligns the three features onto a 16×16×16 spatial grid through multi-resolution downsampling and feature compression, eliminating dimensionality differences among heterogeneous data sources and ensuring point-by-point computability of features at the voxel level. The spatiotemporal attention mechanism submodule establishes a nonlinear relationship between meteorological time series parameters and local 3D spatial features through time decay constraints, gradient stabilization, and k-nearest neighbor sparsification. It generates a spatiotemporally coordinated attention weight distribution and enhances the interpretability of the weights through physical priors. The feature modulation and dynamic fusion submodule maps building geometry and air conditioning parameter features into a unified latent space based on a dual-projection network. It then utilizes a gating mechanism to dynamically adjust the contributions of different modalities, achieving spatially adaptive feature fusion. The 3D convolutional feature abstraction submodule achieves feature compression and semantic enhancement through cascaded 3D convolutions (including residual connections and deconvolutions), ultimately outputting 16-channel decoupled features that preserve the implicit relationship between thermodynamic and fluid dynamics parameters. The geometric consistency constraint submodule introduces a gradient alignment loss function and uses the 3D Sobel operator to calculate feature field gradients. This prevents spatial topological inconsistencies caused by feature distortion and ensures the credibility of physical laws.

[0111] To achieve 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).

[0112] For meteorological characteristic vectors (including time series coding of temperature, humidity, wind speed, etc.), building geometric characteristics And air conditioning parameter characteristics ,A hierarchical downsampling strategy is adopted to achieve spatial dimension alignment.

[0113] Among them, the meteorological characteristic vector After time dimension expansion and spatial broadcasting, a pseudo three-dimensional tensor is generated, that is, the aligned meteorological feature vector , each voxel position The meteorological characteristic vector Updated synchronously with the time step.

[0114] Building geometry Spatial dimensionality reduction is achieved through cascade operations (cascade means the sequential connection of two heterogeneous operations, spatial dimensionality reduction and channel compression).

[0115] Among them, the first layer execution Average pooling downsampling to The resolution and stride of the pooling kernel are set to 4 to preserve key spatial topology information:

[0116] ;

[0117] Where, Represents the building geometric feature tensor after the first layer average pooling, This is a 3D average pooling operation that calculates the average of the eigenvalues ​​within a local cube. kernel indicates that the pooling window size is 4x4x4 voxels, and stride indicates the pooling sliding step size to ensure no overlapping downsampling. Represents a 16x16x16 spatial grid, with each position containing a 256-dimensional feature vector.

[0118] The second layer uses a 1×1×1 convolution kernel to perform channel compression and generate aligned building geometric features. , the 1x1x1 convolution operation compresses the number of feature channels from 256 to 64 while maintaining key geometric properties such as opening position and cavity connectivity.

[0119] Air conditioning parameter characteristics Resolution matching through adaptive max pooling:

[0120] ;

[0121] in, is the aligned air conditioning parameter feature, Represents three-dimensional maximum pooling, extracting the maximum value of the local area, express Spatial grid, 64-dimensional features at each location.

[0122] The dynamic kernel size of the MaxPool3D layer is , step length , output the aligned air conditioning parameter features This processing ensures the computability of multi-source data in the spatial dimension. It ensures that the spatial distribution patterns of supply air temperature and cooling power are not blurred during the downsampling process, and achieves dimensional alignment with other modal data. This stage makes each spatial voxel position of features from different sources directly comparable. This preprocessing ensures that different modal features have voxel-by-voxel alignment within the 16×16×16 grid space, laying the foundation for subsequent cross-modal interaction.

[0123] This embodiment uses a spatiotemporal attention mechanism to model spatiotemporal correlations. The spatiotemporal attention mechanism submodule is the core component of this multimodal fusion neural network model that couples the dynamics of meteorological time series with the static features of architectural space. Its core function is to establish a nonlinear correlation between hourly meteorological fluctuations and the local physical field in three-dimensional space. This spatiotemporal attention mechanism uses a learnable correlation matrix to autonomously identify the impact of meteorological time series parameters on different areas of a building at a specific moment. Compared to traditional feature splicing methods, it can explicitly model the complex interactions between "time, space, and physical quantities," giving the fused features the ability to perceive spatiotemporal synergy.

[0124] The spatiotemporal attention mechanism adopts a multi-head parallel attention structure (in one embodiment of the present invention, a four-head parallel attention structure, where the four heads achieve the optimal balance between parameter quantity and feature diversity, corresponding to the four coupling modes of temperature conduction, momentum exchange, radiation heat flow, and pressure gradient respectively), and each attention head independently learns a different physical coupling mode.

[0125] In one embodiment of the present invention, the spatiotemporal attention mechanism adopts a 4-head parallel attention structure to define a learnable parameter matrix group , projecting the original 128-dimensional aligned meteorological feature vectors into a 32-dimensional subspace. Each head calculates an independent correlation score:

[0126] ;

[0127] In the formula For the Spatial feature projection of the attention head, is the learnable projection matrix, For the moment The meteorological characteristic vector of For the The learnable correlation matrix of the attention heads, Reference Operation, linearly projecting the spliced ​​128-dimensional spatial features into a 32-dimensional subspace.

[0128] The outputs of the four attention heads are aggregated through gated weighting:

[0129] ;

[0130] is the space-time position The final attention weight, gate coefficient Through the learnable parameter vector Dynamic generation, this design enables the multimodal fusion neural network model to automatically assign the importance of each attention head.

[0131] ;

[0132] The time decay constraint:

[0133] An exponential decay factor is introduced to enhance the physical timing characteristics and define the time influence function:

[0134] ;

[0135] in is the time decay factor, is the current prediction time, Hour is the learnable decay coefficient. The corrected attention score is calculated as:

[0136] ;

[0137] in, This is the raw attention score, without considering the time-decay correlation coefficient. It is composed of the concatenation of independent correlation scores calculated by each attention head. This time-decay constraint ensures that the multimodal fusion neural network model conforms to the thermodynamic inertia law, and the decay coefficient is automatically optimized through gradient backpropagation.

[0138] The gradient stabilization calculation:

[0139] To prevent gradient explosion in high-dimensional tensor operations, a layered normalization strategy is designed. First, the score of each attention head is zero-meaned:

[0140] ;

[0141] in, For the The normalized value of the attention head score, The first The mean of the scores of the attention heads, is the mean calculation, For the The original score matrix of the attention heads, The first The standard deviation of the attention head scores, Calculates the standard deviation.

[0142] Then, L2 norm constraints are applied to the output of multiple attention heads to reduce the gradient variance of the training process and normalize the parameters. It is dynamically calculated in each training batch and does not participate in parameter updates.

[0143] ;

[0144] in, Represents the final attention weight after normalization and gated weighting.

[0145] In addition, in order to improve computational efficiency, sparse attention calculation based on k-nearest neighbors is performed.

[0146] First, the spatial domain is sparse, and for each voxel position , only calculates the attention of the 8 meteorological time points with the closest Euclidean distance (8 time points belong to the hyperparameter setting, 8 nearest neighbors are more balanced between computing consumption and accuracy), and the threshold Adaptive Adjustment:

[0147] ;

[0148]

[0149] in, is the voxel position The set of valid time points, The geometric feature field after alignment is at the voxel position The eigenvector at for Meteorological characteristic code at the time, Calculate the expected value of the time dimension. is the standard deviation of the cross-modal distance between meteorological features and geometric features.

[0150] Next is time domain segmentation, which divides 24 hours into 6 periods (00-04, 04-08,..., 20-24). The basic attention pattern is shared within each period. This optimization can reduce computational complexity.

[0151] Physical interpretability is enhanced by introducing prior knowledge to constrain the attention weight distribution.

[0152] The first is the radiation constraint: it imposes a lower limit on the attention weight on the sun-facing side during the noon period (11:00-13:00):

[0153] ;

[0154] in is the constraint strength coefficient, is the sunshine function calculated according to the building orientation.

[0155] The second is ventilation constraint. When the wind speed at the opening is When , the attention weights of the relevant time steps are enforced This prior injection can improve the prediction stability of the multimodal 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.

[0156] In the feature modulation and dynamic fusion submodule, the multimodal features (including the aligned air conditioning parameter features) are first transformed into and aligned building geometry ) is mapped to a unified latent space:

[0157] ;

[0158] ;

[0159] in, and are the projection weight matrices of building geometry features and air conditioning parameter features, and are the corresponding bias terms, is the projection of the building’s geometric features in the latent space, is the projection of the air conditioning parameter features in the latent space.

[0160] Perform spatially adaptive feature fusion to obtain the fused feature vector, and introduce a gating mechanism into the fusion process to dynamically adjust the modal contribution:

[0161] ;

[0162] ;

[0163] in is the spatial gating coefficient, which controls the weight of building geometric features and air conditioning parameter features. is the gating weight matrix, is the corresponding bias term, is the Sigmoid function, which realizes the normalization of the gate coefficient. is the fused feature vector. This design enables the multimodal fusion neural network model to be used in the area near the air outlet ( ) focuses on the influence of air conditioning parameter characteristics, while in confined spaces ( ) Strengthen the geometric characteristics of the building.

[0164] The 3D convolutional feature abstraction submodule achieves feature compression and semantic enhancement through a cascaded 3D convolutional network:

[0165] ;

[0166] ;

[0167] ;

[0168] ;

[0169] And the cascaded 3D convolutional network uses residual connections to retain shallow details:

[0170] ;

[0171] The cascaded 3D convolutional network finally outputs 16-channel decoupled features Each channel corresponds to the implicit encoding of physical quantities such as temperature and flow rate, and feature decoupling is achieved between channels through 1×1 convolution.

[0172] In the geometric consistency constraint submodule, in order to maintain the spatial topological consistency of the fusion features (decoupled features) and the building structure, the gradient alignment loss function is defined :

[0173] ;

[0174] In the formula represents the three-dimensional Sobel gradient operator, Indicates the voxel position The decoupling characteristics of Represents the physical structure area, Indicates masked The number of valid voxels marked as structural regions.

[0175] The gradient is calculated using the central difference format:

[0176] ;

[0177] in, 、 、 is a three-dimensional grid coordinate, representing the voxel position in the building space; and Represents the eigenvalue of the adjacent voxel in the x direction, which is used to calculate the horizontal gradient; is the grid spacing, which determines the scale of gradient calculation. This constraint ensures that the feature gradients of boundaries such as walls and floors maintain directional consistency with the original geometric field, preventing feature distortion.

[0178] The vertical temperature field prediction module adopts a prediction architecture that combines the synergistic effects of basis functions and residual corrections to achieve vertical temperature field modeling that conforms to the laws of thermodynamics and obtain the vertical temperature distribution of the atrium.

[0179] The vertical temperature field prediction module includes a basis function parameter generation network, a residual correction network and an interpolation optimization module.

[0180] The basis function parameter generation network uses a dual-channel feature fusion architecture to generate basis function parameters , decoupling characteristics of the input The global feature vector is compressed into 256 dimensions by global average pooling , while extracting highly relevant features (20 height layers, 64 dimensions each).

[0181] Perform feature cross projection:

[0182] ;

[0183] Where, Represents the fused feature vector formed by cross-projecting the global feature vector and the highly correlated features, is the projection matrix, LayerNorm is a bias term in the feature cross-projection process, which normalizes the channel dimension and eliminates dimensional differences.

[0184] Secondly, two-way attention enhancement is performed to calculate the response weight of each altitude layer to the global feature vector.

[0185] ;

[0186] in, For the The 64-dimensional feature vector of each altitude layer encodes the spatial thermodynamic characteristics of the altitude layer; For the The characteristic vector of the altitude layer, To traverse the index variable of all height layers, is the total number of floors after discretization in the vertical direction of the building.

[0187] Then calculate the local attention and build a height neighborhood relationship graph (neighborhood radius = 2m):

[0188] ;

[0189] in, For the The altitude level is the first The attention weight of the features of the first layer reflects the Altitude level The degree of influence of the thermodynamic state of each altitude layer; and Respectively altitude layer and The absolute height value of each altitude layer is used to calculate the vertical spacing; It is a multi-layer perceptron that takes the absolute value of the height difference as input and outputs the intermediate features used to generate the attention weights; For all the neighbors The weights of the layers are normalized to ensure , making the weights probabilistically interpretable.

[0190] Then perform the synthetic feature operation:

[0191]

[0192] in, For the The updated features of each altitude layer after global-local attention enhancement integrate the local influence of the global thermodynamic state and the neighboring altitude layers; For the The set of neighboring height layers of a height layer reflects the continuity of heat conduction in the vertical direction.

[0193] Finally, perform parameter regression:

[0194] ;

[0195] in is the weight matrix parameter, is the cubic polynomial coefficient of the basis function, which is 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 the physical constraints: .

[0196] The residual correction network adopts a spatiotemporal joint modeling architecture. The residual correction network constructs a highly discretized feature space and divides the building space into 20 height layers with an interval of 1m. Each height layer The temperature correction is calculated via gated convolution.

[0197] First, perform a three-dimensional gated convolution:

[0198] ;

[0199] in For the The temperature residual correction of the altitude layer is obtained by dynamic calculation of the convolution feature; For the The 64-dimensional feature vector corresponding to the height layer comes from the highly correlated features output by cross-modal fusion; For the 、 、 The feature tensors of three adjacent height layers are concatenated to form three layers of feature slices for 3×3×3 convolution operations; is the weight matrix parameter, is the convolution kernel parameter, Represents a 3D convolution operation.

[0200] The second is LSTM height propagation, which constructs a bidirectional LSTM network to transmit thermodynamic states along the height dimension:

[0201] ;

[0202] ;

[0203] ;

[0204] in, From bottom to top ( =1→20) LSTM hidden state during propagation, including low-level thermodynamic effects; From top to bottom ( =20→1) LSTM hidden state during propagation, capturing high-level thermal pressure effects; and Together they form the input vector of LSTM and transmit the thermal state change; is the weight matrix parameter; the hidden state dimension is set to 32, and the final gating coefficient Controls the strength of residual correction.

[0205] Finally, a physical constraint loss function is added and a second-order differential operator is introduced to enhance thermodynamic consistency:

[0206] ;

[0207] in It is a physical consistency loss function, and the energy conservation constraint is constructed through the second-order spatial derivative and time derivative of the temperature field; For height The predicted temperature value at ; for Time altitude The predicted temperature value at ; is the air density, is the specific heat capacity of air at constant pressure, is the height interval, This constraint makes the prediction results satisfy the unsteady heat conduction equation and reduces the energy conservation error.

[0208] The interpolation optimization module performs cubic spline interpolation optimization.

[0209] In order to achieve continuous temperature field prediction with a resolution of 0.1 m, a piecewise cubic polynomial is constructed for 20 discrete nodes using cubic splines with constraints:

[0210] ;

[0211] By imposing the continuity condition 、 and natural boundary conditions , construct a tridiagonal system of equations to solve the interpolation coefficients.

[0212] in, For the definition cubic polynomial functions on intervals; 、 、 、 For the The coefficients of the cubic polynomial for the interval segments; For the The height coordinates of the interpolation nodes; For the Segment polynomials in The function value at must be equal to Segment start value; For the Segment polynomials in Continuity condition of the first-order derivative at ; and For natural boundary conditions, the top and bottom second-order derivatives are forced to be zero to ensure a smooth curve.

[0213] The ventilation volume prediction module adopts a prediction framework that integrates fluid mechanics mechanisms and deep learning to achieve accurate estimation of building opening flow.

[0214] The ventilation volume prediction module includes a graph neural network, a turbulence-corrected spatiotemporal convolutional network, a geometry-physics 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 spatiotemporal convolutional network captures the spatiotemporal characteristics of vortex motion to compensate for the turbulence effect error of the traditional mechanism model; the geometry-physics joint encoding network adopts an adaptive weight distribution mechanism to generate 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.

[0215] Input features include the opening geometry matrix Decoupled features output from the cross-modal feature fusion module Environmental characteristics obtained after processing ,The opening geometric parameter matrix integrates key geometric properties such as the ,position coordinates and aspect ratio of the opening.

[0216] It is the decoupled feature after cross-modal fusion In the process, through spatial alignment and feature sampling, according to the opening The position information is extracted. The decoupled features output by cross-modal feature fusion are Implicitly encode the three-dimensional spatial environment features, for each opening , first perform voxel coordinate mapping, and convert the physical coordinates Convert to Grid coordinate system , followed by trilinear interpolation sampling, decoupling features Extract The 8 voxel features of the neighborhood are interpolated using distance weights to generate the environmental features of the opening position:

[0217] ;

[0218] Weight Calculated as:

[0219] ;

[0220] in, The three-dimensional physical size of the building space is extracted from the three-dimensional point cloud data of the building. 、 and Represents the size of each voxel grid in the physical space in the x-axis, y-axis, and z-axis directions respectively.

[0221] Flow coefficient The prediction network adopts a three-level graph attention architecture to construct an opening geometric topology graph , where the nodes Indicates a building opening, Represents a set of graph nodes, each node corresponds to a building opening; Represents the graph edge set, indicating the spatial adjacency relationship between openings; edge satisfy , For opening The first three columns of the geometric parameter matrix, i.e. the three-dimensional space coordinates, are used to calculate the Euclidean distance between openings. For opening The first three columns of the geometric parameter matrix. Initialize the node features first:

[0222] ;

[0223] In the formula is the feature encoding matrix, Represents the dimension of the weight matrix. The input is 3D geometric features + 128D environmental features, and the output is 32D. is the environmental feature of cross-modal fusion, is the dimension of the environment feature vector output by cross-modal fusion, For nodes The initial feature vector of is generated by fusing geometric parameters with environmental features; Opening The geometric parameter matrix includes attributes such as position coordinates and aspect ratio.

[0224] Then calculate the graph attention weight:

[0225] ;

[0226] in For the In the layer graph attention network, nodes For neighboring nodes Normalized attention weights of ; is the level index of the graph attention network ( ); is the natural exponential function, is the activation function operation of the rectified linear unit with leakage, and Respectively Neighbor nodes in the layer and The eigenvector of is the query,key-value projection matrix, is the learnable parameter vector; For nodes The set of neighboring nodes, is the neighborhood node index variable, For the Layer Node The eigenvector of .

[0227] Next, feature aggregation is updated, and each layer of the graph attention network uses the PReLU activation function:

[0228] ;

[0229] in, For the Layer Node The eigenvector of For the Layer Node The eigenvector of is the layer normalization operation, For nodes The set of neighboring nodes; Index variable for neighborhood nodes; is the value projection matrix, which transforms the neighborhood node features Mapping to the attention space.

[0230] Final output, constraints In line with the scope of engineering experience:

[0231] ;

[0232] in, The node features are compressed into interval, used to generate The scaling factor, is the node after passing through the 3-layer graph attention network Eigenvector; is the output layer weight matrix.

[0233] The mainstream prediction channel is based on the improved Darcy-Weisbach equation to model the pressure difference driven flow. Its core formula is:

[0234] ;

[0235] in, Opening The theoretical volume flow rate prediction value is based on the ideal state flow calculated by the Darcy-Weisbach equation. For opening The pressure difference on both sides is composed of thermal pressure effect and wind pressure. For opening The geometric area directly affects the flow cross-sectional area of ​​flow calculation.

[0236] The turbulence-corrected spatiotemporal convolutional network consists of a dilated convolutional layer, a feature extraction layer, a spatiotemporal compression layer, and an output layer.

[0237] The turbulence-corrected spatiotemporal convolutional network uses spatiotemporal three-dimensional convolution to capture vortex motion characteristics, and the dilated convolution layer uses a dilated convolution structure to enhance the receptive field:

[0238] ;

[0239] in, Indicates the Turbulence-corrected flow prediction value at each opening; is the time offset parameter, is the time window radius; With expansion rate The three-dimensional convolution kernel parameter tensor, is the 7×7×7 convolution kernel coordinate, the expansion rate Increases with network depth; is the slice tensor of the spatiotemporal fusion feature field, is the time dimension offset, is the offset coordinate of the three-dimensional convolution kernel in the spatial dimension.

[0240] The first is the feature extraction layer. For 16-channel → 32-channel convolution, a gating mechanism is used:

[0241] ;

[0242] in, is the gating coefficient matrix, is the gated convolution kernel parameter, is the characteristic convolution kernel parameter, is the output of the feature extraction layer.

[0243] The second is the spatiotemporal compression layer. For the 32-channel → 16-channel convolution, channel attention is introduced:

[0244] , ;

[0245] in, is the channel attention weight coefficient, through Normalized range , representing the The importance of each channel; Is the characteristic channel index variable, the value range is , is the number of input channels, For the A three-dimensional tensor of feature channels 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 space-time compression layer.

[0246] Finally, the output layer performs 16-channel → 4-channel convolution, and the residual connection ensures gradient stability:

[0247] ;

[0248] in, It is the turbulence corrected flow output after convolution processing; is the trainable parameter matrix of the convolutional layer. By using the turbulence-corrected spatiotemporal convolutional network, the multimodal fusion neural network model can effectively capture the 0.5-2.0 Hz turbulent pulsation characteristics, thereby improving the prediction accuracy under conditions of sudden wind speed changes.

[0249] In the traffic fusion stage, the geometric-physical joint coding network uses a geometric feature-sensitive adaptive weight allocation mechanism to generate dynamic fusion weights ,The adaptive weight allocation mechanism generates dynamic fusion weights through geometric feature encoding, multi-head attention aggregation, and variational inference.

[0250] First, perform geometric feature encoding:

[0251] ;

[0252] in is the 16-dimensional opening geometric feature vector generated by 3-layer MLP (16→8→4), Represents the 16-dimensional space of geometric eigenvectors; is the sum of all opening areas in the building atrium ( ), used to calculate the opening area ratio parameters; is the aspect ratio, The angle between the opening normal and the main wind direction is calculated by MLP, which includes 16→8→4 dimensional compression. Then, physical feature interaction is performed and a 4-head attention mechanism is used to calculate the neighborhood opening. The turbulence interaction effect:

[0253] ;

[0254] in, is the physical interaction feature vector after multi-head attention aggregation, encoding the turbulence transfer effect between openings; For opening The set of neighborhood opening geometric parameters is defined as: , including geometric parameters such as the location and area of ​​the neighborhood opening.

[0255] Finally, perform Bayesian optimization:

[0256] ;

[0257] in, is the variational distribution Expected calculation of is the mean of the normal distribution, is the corresponding standard deviation, Then it represents the identity matrix; For the Sigmoid activation function, constrain the linear combination to interval; 、 They are the geometric feature projection weights and physical feature projection weights, respectively, and feature importance is automatically learned through variational inference.

[0258] Then, through the ELBO optimization of the evidence lower bound, the dynamic fusion weight In the narrow opening (aspect ratio ) focuses on turbulence correction ( ), in regular openings (aspect ratio ) focuses on mechanism models ( ).

[0259] The multimodal fusion neural network model generates the final ventilation volume prediction result through collaborative calculation of multiple branches (including graph neural network, turbulence-corrected spatiotemporal convolutional network, and geometric-physical joint coding network), and takes into account the consistency of physical dimensions and engineering application requirements through the design of the physical constraint output layer.

[0260] For each opening , the physical constraint output layer imposes dimensional consistency constraints:

[0261] ;

[0262] in For the The final predicted ventilation rate for each opening is 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 order to dynamically fuse the weights, the contribution ratio of the main channel mechanism model and the turbulence correction channel is dynamically adjusted through the opening geometric characteristics.

[0263] In one embodiment of the present invention, a multi-modal fusion natural ventilation evaluation system for a large atrium in an exhibition building is provided to implement the method provided in the above embodiment. Figure 3 , the system comprises the following units:

[0264] A data acquisition unit, configured to acquire multimodal data, wherein the multimodal data includes meteorological time series parameters, building three-dimensional point clouds and opening geometric parameters, and air conditioning and cooling parameters;

[0265] An evaluation unit, configured to input the multimodal data into a multimodal fusion neural network model to obtain a natural ventilation evaluation result, wherein the natural ventilation evaluation result includes a vertical temperature distribution of the atrium and a ventilation volume of each opening in the atrium;

[0266] 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 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; obtain the meteorological feature vector; the building geometry encoding module is used to generate the building geometry features by fusing the opening position and area parameters based on voxel 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 the air-conditioning parameter features; the cross-modal feature fusion module is used to spatially align the meteorological feature vectors, building geometry features and air-conditioning parameter features and perform multimodal interaction modeling, perform feature compression and semantic enhancement, obtain the fusion features, and make the fusion features consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to collaboratively realize the vertical temperature field modeling by superimposing the basis function and residual correction to obtain the vertical temperature distribution of the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening.

[0267] In one embodiment of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the aforementioned embodiment when executing the computer program.

[0268] In one embodiment of the present invention, a computer-readable storage medium is provided, wherein 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 above embodiment are implemented.

[0269] This embodiment of the present invention utilizes spatiotemporal attention-gating coupled feature modeling, integrating a hybrid architecture with bidirectional LSTM time series modeling and time-decay gating to achieve forward-backward causal analysis for building environment forecasting for the first time. By introducing time-difference gating and a self-identification mechanism for peak solar radiation moments, the feature weights of key meteorological events are dynamically enhanced. Combined with the physical field projection technology of a multi-head self-attention mechanism, a quantitative characterization of wind speed-radiation coupling effects is achieved.

[0270] The embodiment of the present invention establishes a geometric topology-aware three-dimensional point cloud dynamic coding system and proposes an improved PointNet++ architecture for building cavities. Based on the three-level coding paradigm of "voxelization-dynamic graph convolution-opening influence field fusion", it defines the opening influence intensity Gaussian kernel function and multi-scale feature fusion, breaking through the technical bottleneck of traditional point cloud processing that cannot quantify ventilation paths, and supports geometric topology association modeling of irregular openings.

[0271] The embodiment of the present invention constructs a dynamic fusion framework driven by physical constraints, establishes a feature fusion framework with joint constraints of spatiotemporal attention and thermodynamic equations (the constructed bidirectional LSTM network transmits thermodynamic states along the height dimension), introduces a physical consistency loss function, and realizes the deep integration of data-driven models and thermodynamic principles through sparse attention calculation based on k-nearest neighbors and physical interpretability enhancement.

[0272] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is intended to be embodied in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A multi-modal fusion method for evaluating the natural ventilation of large atriums in exhibition buildings, characterized by: Including steps: Acquiring multimodal data, the multimodal data including meteorological time series parameters, building three-dimensional point cloud and opening geometric parameters, and air conditioning and cooling parameters; Inputting the multimodal data into a multimodal fusion neural network model to obtain a natural ventilation assessment result, wherein the natural ventilation assessment result includes the vertical temperature distribution of 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 spatially align and multimodal interactive modeling of meteorological feature vectors, building geometry features and air-conditioning parameter features, perform feature compression and semantic enhancement, obtain fusion features, and make the fusion features consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to collaboratively realize vertical temperature field modeling through basis function superposition residual correction to obtain the vertical temperature distribution of the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening; 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 encoding module is used to generate building geometry features by fusing opening position and area parameters based on voxel preprocessing, feature extraction and dynamic graph convolution; the air-conditioning system parameter encoding module is used to map air-conditioning operation parameters into a three-dimensional parameter field to obtain air-conditioning parameter features.

2. A multi-modal fusion natural ventilation assessment method for a large atrium in an exhibition building 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 process meteorological time series parameters from the forward and backward directions respectively, capture the forward cumulative effect and backward causal relationship of meteorological time series parameters, and obtain time series feature output; Timing The feature output is fed into a multi-head self-attention mechanism to enhance the feature extraction capability of key time nodes, and the attention weights of key meteorological events are dynamically adjusted through a time-aware gating mechanism. The air conditioning system parameter encoding module is configured to perform the following operations: Learn the nonlinear relationship between air conditioning operating parameters and obtain the feature vector; The spatial broadcast mechanism is used to expand the feature vector into a three-dimensional parameter field to obtain the air conditioning parameter features.

3. The multimodal fusion natural ventilation assessment method for large atriums in exhibition buildings according to claim 1 is 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 disordered point cloud into a structured feature field; the feature extraction network is used to extract features from the structured feature field and generate primary geometric features; the dynamic graph convolution layer is used to perform local geometric feature extraction and feature aggregation. Through dynamic graph convolution, the original point cloud of the building atrium is converted into intermediate geometric features. The opening feature fusion layer is used to fuse the three-dimensional coordinates of the opening position parameters with the area information and splice them with the intermediate geometric features to obtain the channel splicing features; The multi-scale feature fusion layer is used to integrate primary geometric features, intermediate geometric features and channel splicing features to obtain building geometric features.

4. The multimodal fusion natural ventilation assessment method for large atriums in exhibition buildings according to claim 3 is characterized in that: The cross-modal feature fusion module includes: The spatial alignment preprocessing submodule is used to perform spatial alignment preprocessing on meteorological feature vectors, building geometric features, and air conditioning parameter features; The spatiotemporal attention mechanism submodule is used to establish a nonlinear association between meteorological time series parameters and local features in three-dimensional space through time decay constraints, gradient stabilization calculations, and k-nearest neighbor sparse calculations, generate a spatiotemporal coordinated attention weight distribution, and enhance the interpretability of the weights through physical priors; The feature modulation and dynamic fusion submodule is used to map building geometric features and air conditioning parameter features into a unified latent space based on a dual-projection network. It also uses a gating mechanism to dynamically adjust the contribution of different modalities to achieve spatially adaptive feature fusion. The 3D convolutional feature abstraction submodule is used to perform feature compression and semantic enhancement through cascaded convolutions, output fused features, and preserve the implicit coding relationship between thermodynamic and fluid dynamics parameters; The geometric consistency constraint submodule is used to introduce the gradient alignment loss function to maintain the spatial topological consistency of the fused features and the building structure, and calculate the feature field gradient.

5. The multi-modal fusion natural ventilation assessment method for a large atrium in an exhibition building according to claim 1 is characterized in that: The vertical temperature field prediction module includes: The basis function parameter generation network is used to perform feature cross-projection, attention enhancement, and feature synthesis operations on the global feature vector and highly correlated features, and perform parameter regression to generate basis function parameters for constructing the vertical temperature distribution model; The residual correction network is used to perform residual correction on the temperature by using gated convolution, adopting a bidirectional LSTM network to transfer the thermodynamic state along the height dimension and introducing a physical constraint loss function.

6. A multimodal fusion natural ventilation assessment method for a large atrium in an exhibition building according to any one of claims 1 to 5, characterized in that: The ventilation volume prediction module includes: A graph neural network is used to dynamically predict the flow coefficient of each opening by modeling the geometric topological relationship between the openings and based on the output of the cross-modal feature fusion module; Turbulence-corrected spatiotemporal convolutional network, used to capture the spatiotemporal characteristics of vortex motion and compensate for turbulence effect errors; A joint geometry-physics encoding network for generating dynamic fusion weights based on the opening morphology and physical parameters using an adaptive weight allocation mechanism that generates dynamic fusion weights through geometric feature encoding, multi-head attention aggregation, and variational inference; The physical constraint output layer is used to ensure that the ventilation volume prediction results conform to the laws of fluid mechanics through dimensional consistency and area threshold suppression.

7. A multi-modal fusion natural ventilation assessment system for large atriums in exhibition buildings, characterized by: The following units are included: A data acquisition unit, configured to acquire multimodal data, wherein the multimodal data includes meteorological time series parameters, building three-dimensional point clouds and opening geometric parameters, and air conditioning and cooling parameters; An evaluation unit, configured to input the multimodal data into a multimodal fusion neural network model to obtain a natural ventilation evaluation result, wherein the natural ventilation evaluation result includes a vertical temperature distribution of the atrium and a 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 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 opening position and area parameters based on voxel preprocessing, feature extraction and dynamic graph convolution; the air-conditioning system parameter encoding module is used to map the air-conditioning operating parameters into a three-dimensional parameter field to obtain air-conditioning parameter features; the cross-modal feature fusion module is used to spatially align the meteorological feature vector, building geometry features and air-conditioning parameter features and perform multimodal interaction modeling, perform feature compression and semantic enhancement, obtain fusion features, and make the fusion features consistent with the spatial topology of the building structure; the vertical temperature field prediction module is used to collaboratively realize vertical temperature field modeling through basis function superposition residual correction to obtain the vertical temperature distribution of the atrium; the ventilation volume prediction module is used to predict the ventilation volume of each opening.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images