A high-density residential form optimization method based on high-difference ventilation quantification guidelines

Abstract

The application discloses a high-density residential form optimization method based on height difference ventilation quantity quantitative guidance, and belongs to the technical field of building ventilation design. A building design parameter set is acquired, and a category identification vector and a numerical vector are respectively extracted through a parameter dimension encoder and a parameter numerical encoder; the ventilation coupling calculation network is input, the coupling strength between each two of the building width, the front row building height and the building spacing is calculated, and a three-dimensional coupling feature tensor is generated; then low-rank compression is carried out along three dimension directions respectively, and the compressed feature tensor is fused through ring product operation; the influence weight coefficients of each parameter are dynamically calculated through the adaptive gating unit, and the comprehensive feature vector is generated by weighted summation; the height difference prediction layer is input, and the minimum north-south building height difference layer meeting the ventilation requirement is output. Through the construction of the end-to-end framework of parameter decoupling, coupling extraction, low-rank compression, adaptive weighting and height difference prediction, the application provides quantitative guidance for high-density residential form optimization.

Description

Technical Field

[0001] This invention relates to the field of building ventilation design technology, and more specifically, to a method for optimizing the form of high-density residential buildings based on quantitative guidance for ventilation based on elevation differences. Background Technology

[0002] In high-density urban living environments, natural ventilation performance has a significant impact on resident health, thermal comfort, and building energy efficiency. Especially in regions like Shanghai with high temperatures and humidity in summer, improving the wind environment through rational building layout has become an important topic in architectural design and urban planning. Research shows that, under row-and-column layout conditions with the prevailing wind direction being due south, by creating a height difference between the front and rear rows of buildings (i.e., "height difference ventilation"), the descending airflow generated when air encounters the taller rear buildings can significantly improve ventilation velocity between buildings. A definite quantitative relationship exists between building width, front row building height, building spacing, and the minimum number of north-south building height differences required to meet summer ventilation requirements.

[0003] Currently, the assessment of wind environment performance in residential areas mainly relies on computational fluid dynamics simulations or wind tunnel experiments, but this approach suffers from low efficiency, high cost, and strong dependence on design experience. While some studies have attempted to establish table-based quantitative guidelines, these tables only cover discrete parameter combinations, failing to generalize to unsimulated combinations and struggling to model nonlinear interactions between parameters. Furthermore, building width, height, and spacing are design parameters of different dimensions; traditional methods co-encode them, ignoring the semantic differences between parameter dimensional types and specific values, leading to representational confusion and decreased prediction accuracy. Therefore, constructing a quantitative guidance method for elevation difference ventilation that can decouple the heterogeneous semantics of design parameters, capture high-order interactions between parameters, and generalize to unsimulated parameter combinations has become a pressing technical problem in this field. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a method for optimizing the form of high-density residential buildings based on quantitative guidance for ventilation differences. This method enables intelligent recommendation of the number of floors with elevation differences for any combination of input parameters, providing a new technical approach for optimizing the ventilation performance of high-density residential buildings.

[0005] This invention provides a method for optimizing the form of high-density residential buildings based on elevation difference ventilation quantification guidelines, comprising the following steps: Step S1: Obtain the building design parameter set, extract the category identifier vector of each parameter through the parameter dimension encoder, and extract the numerical vector of each parameter through the parameter value encoder. The building design parameter set includes building width, front building height and building spacing. Step S2: Input the category identifier vector and the numerical vector into the ventilation coupling calculation network to calculate the pairwise coupling strength between building width, front building height and building spacing, and generate a three-dimensional coupling feature tensor; Step S3: Perform low-rank compression on the three-dimensional coupled feature tensor along the three dimensions respectively, and then fuse the compressed features in the three directions into a compressed feature tensor through a ring product operation; Step S4: Input the compressed feature tensor into the adaptive gating unit, dynamically calculate the influence weight coefficients of the three parameters of building width, front building height and building spacing, and perform weighted summation of the feature components in the compressed feature tensor corresponding to each parameter according to the influence weight coefficients to generate a comprehensive feature vector; Step S5: Input the comprehensive feature vector into the elevation difference prediction layer, and output the minimum number of north-south building elevation difference layers that meet the ventilation requirements under the current parameter combination, as the elevation difference ventilation quantification guidance result.

[0006] In this solution, step S1 specifically includes: A parameter dimension vocabulary is constructed, which includes building width dimension, front building height dimension, and building spacing dimension. The building design parameters to be processed are input into the parameter dimension encoder, and the index of the parameter in the dimension vocabulary is used as the query key to generate the category identifier vector of the parameter through independent hot encoding. The architectural design parameter values ​​are physically discretized and modularized to generate physical realization values ​​for each parameter. The architectural design parameter values ​​include building width, front building height, and building spacing. Physically perceived random mask perturbation operation is performed on the physical implementation value respectively. The perturbation is randomly determined by the set mask probability. If it is added, the offset is sampled and superimposed within the preset perturbation radius according to the probability distribution to generate perturbed parameter value. The perturbated parameter values ​​are input into the corresponding competitive discrete codebook for encoding. The matching degree of each perturbated parameter value is calculated with each codeword in the corresponding codebook. A preset number of codeword vectors are selected through a competitive selection mechanism as the encoded output of the parameters, which serves as the numerical vector of the parameters. The category identifier vector is combined with the numerical vector to obtain the complete encoded representation of each parameter.

[0007] In this solution, step S2 specifically includes: The category identifier vector and numerical vector of each parameter are respectively input into three physically isolated coupled calculation paths, including the wide-height coupling path, the wide-spacing coupling path, and the high-spacing coupling path; Within each coupling path, the two sets of input category identifier vectors are multiplied by an outer product to generate a category interaction matrix, and the two sets of input numerical vectors are multiplied by an outer product to generate a numerical interaction matrix. The category interaction matrix and the numerical interaction matrix are added element by element to obtain a combined feature matrix. Within each coupling path, linear interaction paths and non-linear interaction paths are executed in parallel. The linear interaction path transforms the two sets of numerical vectors linearly, multiplies them element by element, and then projects them to output the first interaction feature. The non-linear interaction path concatenates the two sets of category identifier vectors with the two sets of numerical vectors and outputs the second interaction feature through a multi-layer fully connected network. Within each coupling path, the first interaction feature and the second interaction feature are weighted and fused according to the physical prior interaction kernel corresponding to the coupling type to generate the coupling strength feature of the path.

[0008] In this solution, step S2 further includes: The coupling strength features of the three output paths are stacked along the coupling type dimension to generate a three-dimensional coupling feature tensor, where the output of the wide and high coupling paths is used as the first layer, the output of the wide spacing coupling path is used as the second layer, and the output of the high spacing coupling path is used as the third layer. The three-dimensional coupled feature tensor undergoes a second interaction in the coupling type dimension to calculate the mutual attention weights among width-height coupling, width-distance coupling, and height-distance coupling. Based on these attention weights, the features of each layer are reweighted to output the final three-dimensional coupled feature tensor.

[0009] In this solution, step S3 specifically includes: The three-dimensional coupled feature tensor is expanded along the three parameter dimensions of width, height, and spacing to generate a width direction view, a height direction view, and a spacing direction view. The three generated directional views are decomposed into low-rank matrices respectively. Each view matrix is ​​approximately decomposed into the product of two low-rank matrices. During the decomposition, the decomposition rank of each view is independently set according to the physical importance of each parameter dimension to the ventilation effect, and physical constraint regularization terms are introduced in the decomposition process. The low-rank decomposition results of the three directional views are fused by a circular product. The fusion order is executed step by step according to the cyclic path of width, height and spacing to generate the base mode tensor. The base mode tensor and the original 3D coupled feature tensor are condensed together. The dimensions of the base mode tensor are weighted and summed according to the weights provided by the original tensor to output the compressed feature tensor.

[0010] In this solution, step S4 specifically includes: A dual-input structure including a scene path and a feature path is constructed. The scene path receives the numerical vectors of each parameter and compresses them into a scene compressed vector through physical pooling. The feature path receives the compressed feature tensor. Obtain the magnitude of the partial derivatives of the compressed feature tensor with respect to the numerical vectors of each parameter as the first-order sensitivity. Obtain the sum of the magnitudes of the mixed partial derivatives of the compressed feature tensor with respect to each parameter pair as the second-order sensitivity. The weighted sum of the first-order sensitivity and the second-order sensitivity is used as the comprehensive sensitivity of each parameter. The scene compression vector, the comprehensive sensitivity of each parameter, and the part of the compressed feature tensor related to each parameter are fused together to calculate the original scores of building width, front building height, and building spacing respectively. The three raw scores obtained from the calculation are input into a competitive normalization unit, and temperature scaling and Softmax normalization operations are performed to generate the influence weight coefficients of the three parameters.

[0011] In this solution, step S4 further includes: The width, height, and spacing features are extracted from the compressed feature tensor and mapped to feature vectors of fixed dimensions, respectively. The extracted width, height, and spacing features are weighted and summed using the aforementioned influence weight coefficients to generate a fused feature vector. This fused feature vector is then mapped through an output projection layer to generate a comprehensive feature vector.

[0012] In this scheme, the elevation difference prediction layer in step S5 is implemented using a multi-level category-aware graph neural network, specifically including: An ordered category-aware embedding space with varying height difference layers is constructed, wherein the height difference layer number is a non-negative integer. Sine-cosine position coding is used to generate position coding vectors, such that the similarity between coding vectors of any two layers depends only on the difference in layer number. A three-dimensional mesh graph structure is constructed using discrete level combinations of width, height, and spacing as nodes and node pairs that differ by one level in parameter dimension as edges. Message passing is performed in the graph through a graph neural network, so that each node in the graph obtains information from its neighboring nodes and updates its own representation. The elevation difference layers are divided into four levels: no elevation difference layer, low elevation difference layer, medium elevation difference layer, and high elevation difference layer. A level predictor is constructed, which takes the comprehensive feature vector as input and outputs the probability distribution of each level to determine the semantic level to which the current parameter combination belongs.

[0013] In this solution, step S5 further includes: For the low elevation difference demand level, an ordered regression method is used to predict the specific elevation difference level. For the medium elevation difference demand level, a recommendation method based on graph neural networks is used. The recommendation level is determined by retrieving the labels of neighboring nodes in the parameter combination graph and performing weighted voting. For the high elevation difference demand level, a special simulation suggestion is output. Calculate the confidence level of the prediction results and perform a neighborhood consistency check. Calculate the degree of difference in the labels of the target node's neighboring nodes in the graph. When the confidence level is lower than a preset threshold or the neighborhood consistency is lower than a preset standard, output a specific simulation suggestion instead of a specific number of elevation difference layers. Output the minimum number of floors required for north-south building elevation difference or a specific simulation suggestion, as a quantitative guideline for elevation difference ventilation.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively solves the semantic confusion problem caused by the mixed encoding of parameter types and values ​​in traditional methods by constructing a dual-encoding architecture that decouples parameter dimensions and numerical values. This allows the network to independently learn the dimensional semantics and numerical features of parameters, improving the model's ability to represent architectural design parameters. Through a ventilation coupling computation network, pairwise coupling strength calculations are performed on building width, front-row building height, and building spacing, enabling explicit modeling of the synergistic effects of multiple parameters on ventilation, enhancing the model's ability to capture nonlinear interactions between parameters. Ring-shaped low-rank compression fusion significantly reduces feature dimensions while maintaining the cyclical transmission relationship between the three parameter dimensions, reducing the number of model parameters and computational complexity, and improving the model's generalization performance. Adaptive gating units dynamically calculate the influence weight coefficients of each parameter, achieving scene-adaptive feature weighting and summarization. This allows the model to flexibly adjust its focus based on the characteristics of different design parameter combinations, enhancing its adaptability to different design scenarios. The elevation difference prediction layer outputs the minimum number of north-south building elevation differences, providing quantifiable ventilation performance guidance for designers and planning management departments, avoiding the inefficient process of repeated trial and error in traditional CFD simulations, and reducing design cycles and costs. Overall, this invention forms an end-to-end quantitative guidance generation framework, which improves the interpretability and engineering applicability of the model while ensuring prediction accuracy. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0016] Figure 1A flowchart of a high-density residential form optimization method based on elevation difference ventilation quantification guidelines is shown. Figure 2 A flowchart is shown to obtain the coupling strength of building width, front row building height, and building spacing. Figure 3 The flowchart for obtaining the compressed feature tensor is shown; Figure 4 A flowchart is shown showing the semantic level to which the current parameter combination belongs, output by the elevation difference prediction layer. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0019] like Figure 1 As shown, this embodiment provides a method for optimizing the form of high-density residential buildings based on elevation difference ventilation quantification guidelines, including: Step S1: Obtain the building design parameter set, extract the category identifier vector of each parameter through the parameter dimension encoder, and extract the numerical vector of each parameter through the parameter value encoder. The building design parameter set includes building width, front building height and building spacing. Step S2: Input the category identifier vector and the numerical vector into the ventilation coupling calculation network to calculate the pairwise coupling strength between building width, front building height and building spacing, and generate a three-dimensional coupling feature tensor; Step S3: Perform low-rank compression on the three-dimensional coupled feature tensor along the three dimensions respectively, and then fuse the compressed features in the three directions into a compressed feature tensor through a ring product operation; Step S4: Input the compressed feature tensor into the adaptive gating unit, dynamically calculate the influence weight coefficients of the three parameters of building width, front building height and building spacing, and perform weighted summation of the feature components in the compressed feature tensor corresponding to each parameter according to the influence weight coefficients to generate a comprehensive feature vector; Step S5: Input the comprehensive feature vector into the elevation difference prediction layer, and output the minimum number of north-south building elevation difference layers that meet the ventilation requirements under the current parameter combination, as the elevation difference ventilation quantification guidance result.

[0020] The minimum north-south building height difference refers to the minimum number of stories required for a height difference between two adjacent rows of buildings in the same row-column layout. In the example, the experimental model consists of two buildings, one in front and one behind, with the incident wind direction being due south, and the southern residence being the front row. For a residential area composed of multiple rows of buildings, the above method is applied to each pair of adjacent rows of buildings to determine the minimum number of stories required for a height difference between each rear row of buildings and the front row of buildings.

[0021] It should be noted that a parameter dimension vocabulary is constructed, which includes dimensions such as building width, front building height, and building spacing. The vocabulary can be further expanded to include dimensions such as building depth, building orientation angle, and prevailing wind direction angle to adapt to more complex application scenarios. The architectural design parameters to be processed are input into the parameter dimension encoder. Using the index of the parameter in the dimension vocabulary as the query key, a category identifier vector for the parameter is generated through independent hot encoding. The architectural design parameter values ​​are then subjected to physical discrete modular normalization to generate physical realization values ​​for each parameter. These architectural design parameter values ​​include building width, front building height, and building spacing. The building width is normalized using the structural column grid module as the basic unit, the front building height is normalized using the building floor height as the basic unit, and the building spacing is normalized using the spacing module as the basic unit. A physical-aware random mask perturbation operation is performed on each of the physical realization values. Whether to add perturbation is randomly determined by a set mask probability. If perturbation is added, the offset is sampled and superimposed within a preset perturbation radius according to a probability distribution to generate perturbed parameter values.

[0022] The perturbed parameter values ​​are input into the corresponding competitive discrete codebook for encoding. The discrete codebook is a learnable codeword matrix corresponding to each parameter type. Each codeword represents the prototype representation of a type of parameter value. The initial values ​​of the codeword vectors are physically initialized based on the commonly used value set or modular system of that parameter type in architectural design. The matching degree of each perturbed parameter value is calculated with each codeword in the corresponding codebook. A predetermined number of codeword vectors are selected as the encoded output of the parameters through a competitive selection mechanism, serving as the numerical vectors of the parameters. The matching degree is measured using negative Euclidean distance or cosine similarity. The competitive selection mechanism can be a hard competition mechanism or a soft competition mechanism. The hard competition mechanism outputs only the single codeword vector with the highest matching degree, while the soft competition mechanism outputs a weighted sum of all codeword vectors with normalized weights according to the matching degree. The category identifier vector is combined with the numerical vector to obtain the complete encoded representation of each parameter.

[0023] It should be noted that, as Figure 2As shown, the category identifier vector and numerical vector of each parameter are input into three physically isolated coupled calculation paths, including the width-height coupling path, the width-spacing coupling path, and the height-spacing coupling path. The width-height coupling path receives the encoded vectors of the building width and the height of the front row buildings, the width-spacing coupling path receives the encoded vectors of the building width and the building spacing, and the height-spacing coupling path receives the encoded vectors of the height of the front row buildings and the building spacing. The three paths have the same structure and the parameters are independent and do not share.

[0024] Within each coupling path, the two sets of input category identifier vectors are multiplied by an outer product to generate a category interaction matrix, and the two sets of input numerical vectors are multiplied by an outer product to generate a numerical interaction matrix. The category interaction matrix and the numerical interaction matrix are then added element-wise to obtain a combined feature matrix. Within each coupling path, linear and nonlinear interaction paths are executed in parallel. The linear interaction path linearly transforms the two sets of numerical vectors, multiplies them element-wise, and then projects them to output a first interaction feature. The nonlinear interaction path concatenates the two sets of category identifier vectors and the two sets of numerical vectors, and then outputs a second interaction feature through a multi-layer fully connected network. Within each coupling path, the first and second interaction features are weighted and fused according to the physical prior interaction kernel corresponding to the coupling type to generate the coupling strength feature of the path. The physical prior interaction kernel of the width-to-height coupling path is initialized as a normalized width-to-height ratio matrix, where the width-to-height ratio is defined as the building width divided by the height of the preceding buildings. The physical prior interaction kernel of the wide-spacing coupling path is initialized as a normalized channel width-to-height ratio matrix, where the channel width-to-height ratio is defined as the building width divided by the building spacing. The physical prior interaction kernel of the height-spacing coupling path is initialized as a normalized spacing-to-height ratio matrix, where the spacing-to-height ratio is defined as the building spacing divided by the height of the preceding buildings. The weights of the weighted fusion are dynamically calculated based on the confidence level of the physical prior interaction kernel. When the ratio corresponding to the physical prior interaction kernel is within a preset reliability interval, the weight of the linear interaction path is higher than that of the nonlinear interaction path; when the ratio exceeds the preset reliability interval, the weight of the nonlinear interaction path is higher than that of the linear interaction path.

[0025] The coupling strength features of the three output paths are stacked along the coupling type dimension to generate a three-dimensional coupling feature tensor. The outputs of the wide-height coupling path serve as the first layer, the outputs of the wide-spacing coupling path as the second layer, and the outputs of the high-spacing coupling path as the third layer. A secondary interaction is performed on this three-dimensional coupling feature tensor along the coupling type dimension using a self-attention mechanism. Each layer of the three-dimensional coupling feature tensor along the coupling type dimension is used as a query unit. The attention score between this layer and other layers is calculated, and the features of other layers are weighted and summed using the attention score as a weight. This achieves information interaction between coupling types, outputting the final three-dimensional coupling feature tensor. Each layer of the three-dimensional coupling feature tensor explicitly corresponds to a physical coupling type, clearly tracing the dominant role of the coupling type in the final ventilation effect. During the design optimization process, the parameters corresponding to a specific coupling type can be consciously adjusted.

[0026] It should be noted that, as Figure 3 As shown, the three-dimensional coupled feature tensor is expanded along the three parameter dimensions of width, height, and spacing to generate a width-oriented view, a height-oriented view, and a spacing-oriented view. The width-oriented view has rows with width as the row and columns with the joint dimension of height and spacing as the column; the height-oriented view has rows with height as the row and columns with the joint dimension of width and spacing as the column; and the spacing-oriented view has rows with spacing as the row and columns with the joint dimension of width and height as the column. Low-rank matrix decomposition is performed on the three generated views, approximating each view matrix as the product of two low-rank matrices. During decomposition, the decomposition rank of each view is independently set according to the physical importance of each parameter dimension's impact on ventilation, with the width-oriented view having the highest decomposition rank, followed by the height-oriented view, and the spacing-oriented view having the lowest. Physical constraint regularization terms are introduced during the decomposition process, including at least one of monotonicity constraints, smoothness constraints, and non-negativity constraints.

[0027] The low-rank decomposition results of the three directional views are fused using a circular product, with the fusion order proceeding sequentially along a cyclic path of width, height, and spacing. First, the low-rank decomposition results of the width and height views are fused by an outer product to generate an intermediate tensor. Then, this intermediate tensor is fused by an outer product with the low-rank decomposition results of the spacing view to generate a base pattern tensor. A cyclic correspondence exists between the three views; information from any one view can be approximated by a combination of the other two views. The circular product fusion is performed sequentially in the cyclic order of width, height, spacing, and width. The base pattern tensor is then compressed with the original 3D coupled feature tensor to reduce the high-dimensional base pattern tensor to the target dimension. In ventilation problems, the influence of the three parameters—width, height, and spacing—is cyclically propagated. Width affects the lateral distribution of airflow, lateral distribution affects the airflow pattern in the height direction, airflow pattern affects attenuation in the spacing direction, and the attenuation of spacing, in turn, affects the effective influence range of width. The dimensions of the base pattern tensor are weighted and summed according to the weights provided by the original tensor to output a compressed feature tensor. Compressed feature tensors remove redundant components in the interaction information between parameters through low-rank decomposition, and the number of parameters is significantly reduced.

[0028] It should be noted that a dual-input structure is constructed, comprising an adaptive gating unit with scene paths and feature paths. The scene path receives the numerical vectors of each parameter, which are compressed into a scene compressed vector through physical pooling, serving as basic information describing the scene in which the current parameter combination exists. The feature path receives compressed feature tensors, serving as deeper information describing the interaction patterns between parameters. The physical pooling operation groups the parameters according to their physical classifications: width parameters are grouped into narrow, medium, and wide; height parameters into low, medium, and high; and spacing parameters into near, medium, and far groups. This is defined based on the physical classification of parameter values ​​in architectural design, giving the compressed scene vectors clear physical semantics.

[0029] Based on the compressed feature tensor, a preset intermediate layer is used to obtain the predicted height difference layer number. The predicted height difference layer number represents the minimum north-south building height difference layer required to meet summer ventilation requirements (inter-building wind speed exceeding 3 m / s) under the current input combination of building width, front building height, and building spacing parameters. It should be noted that the predicted height difference layer number here is based on the forward calculation result of the current network parameters, and is not the final output of step S5. The intermediate calculation results are only used for sensitivity calculation and do not participate in feature transfer in subsequent steps.

[0030] The multi-scale sensitivity of each parameter, including first-order and second-order sensitivity, is calculated within the adaptive gating unit. The predicted elevation difference level is then determined. The magnitude of the partial derivative of each parameter's numerical vector is used as the first-order sensitivity. The first-order sensitivity measures the direct impact of a change in a single parameter on the ventilation effect. The first-order sensitivity corresponding to the preferred width parameter is expressed as: Obtain the predicted elevation difference level. The sum of the magnitudes of the mixed partial derivatives of each parameter pair is taken as the second-order sensitivity. The second-order sensitivity measures the indirect influence of a parameter through its interaction with other parameters. The preferred second-order sensitivity corresponds to the width parameter. Represented as: The width parameter vector is The height parameter numerical vector is The spacing parameter numerical vector is The weighted sum of the first-order and second-order sensitivities is used as the overall sensitivity of each parameter. The sensitivity of each parameter is explicitly calculated; parameters that are more sensitive to ventilation effects should be assigned higher weights during design optimization, as even small adjustments can lead to significant improvements.

[0031] Since each combination of geometric parameters (width, front row height, spacing) matches the minimum number of elevation differences required to meet preset ventilation requirements, the predicted number of elevation differences is essentially a proxy indicator of ventilation effectiveness. A larger number of elevation differences indicates that a greater elevation difference is needed to meet ventilation requirements under the current combination of geometric parameters, proving a poorer ventilation effect; a smaller number of elevation differences indicates that good ventilation can be achieved even with a small elevation difference under the current combination of geometric parameters, proving a better ventilation effect. Predicting the number of elevation differences is equivalent to evaluating the ventilation effect under the current combination of geometric parameters.

[0032] The scene compression vector, the comprehensive sensitivity of each parameter, and the parameter-related parts of the compressed feature tensor are fused to calculate the original scores for building width, front building height, and building spacing. The width-related features include a width-to-height coupling layer and a width-to-spacing coupling layer; the height-related features include a width-to-height coupling layer and a height-to-spacing coupling layer; and the spacing-related features include a width-to-spacing coupling layer and a height-to-spacing coupling layer. A preferred original width score is... Represented as: , For learnable mapping functions, For width-related features extracted from the compressed feature tensor, This is a physical prior bias. This is a compressed representation of the width after physical pooling. This refers to the overall sensitivity of the width.

[0033] During the integration process, physical prior biases are injected. These biases are pre-calculated based on ventilation physics formulas and include width, height, and spacing physical importance factors. The width physical importance factor is set based on the building's width-to-height ratio. In high-density residential areas, the presence of front-row buildings forces rear-row buildings to rely primarily on descending airflow brought by the overhead airflow for ventilation. Therefore, the larger the width-to-height ratio of the front-row buildings, the stronger the overhead airflow, and the more the rear-row buildings rely on height differences to capture descending airflow. The greater the width-to-height ratio deviates from 1, the larger the width physical importance factor. The formula for calculating the width physical importance factor is: The width of the front row buildings is The height of the front row buildings is , The scale parameter controls the rate at which the aspect ratio deviates from 1. Based on typical building wind tunnel test data, the scale parameter is set to 0.5 to 0.8.

[0034] The height-based physical importance factor is set based on the relationship between the height of the front buildings and the thickness of the atmospheric boundary layer. For ventilation issues in high-density residential buildings, the height of the front buildings determines the speed of the airflow at the top; the taller the building, the greater the wind speed at the top, and the stronger the downdraft generated, making it easier for the buildings behind to achieve good ventilation. The boundary layer thickness is defined as the height at which wind speeds reach gradient wind speeds (unaffected by ground friction), which is approximately 300 to 500 meters in densely populated urban areas. In the lower boundary layer (approximately 0 to 100 meters), wind speed increases significantly with height; in the upper boundary layer (above 100 meters), the rate of increase in wind speed slows down. The formula for calculating the height-based physical importance factor is: The height of the front row buildings is The atmospheric boundary layer thickness in the region is , This is a power-law exponent, controlling the rate at which the control factor increases with height. Based on boundary layer wind speed profile theory, it is typically taken as 0.25 to 0.35.

[0035] The physical importance factor of spacing is set based on the ratio of spacing to height; the larger the ratio, the smaller the physical importance factor. The shading effect of front-row buildings on rear-row buildings mainly depends on the ratio of spacing to height. In ventilation design optimization, when the ratio of spacing to height is already large, further increasing the spacing has negligible improvement on ventilation. Rear-row buildings are basically out of the influence of the wake zone of the front-row buildings, and ventilation conditions mainly depend on the environmental wind field; the importance of the spacing parameter approaches zero. However, when the ratio of spacing to height is small, rear-row buildings are located near the wake zone of the front-row buildings, and airflow is severely blocked, resulting in extremely poor ventilation. Even small adjustments to the spacing parameter will significantly affect the ventilation effect, making it highly important. The formula for calculating the physical importance factor of spacing is: The height of the front row buildings is The building spacing is , The attenuation coefficient is the rate at which the control factor decreases as the distance-to-height ratio increases. Based on wake region studies, it is typically taken as 1.0 to 1.5.

[0036] The three raw scores obtained from the calculation are input into a competitive normalization unit, and temperature scaling and Softmax normalization operations are performed to generate influence weight coefficients for the three parameters. These influence weight coefficients satisfy the constraint that they sum to 1 and are all non-negative, reflecting the competitive contribution relationship of the three parameters to the ventilation effect. Width, height, and spacing features are extracted from the compressed feature tensor and mapped to fixed-dimensional feature vectors. The width feature is extracted from the width-height coupling layer and the width-spacing coupling layer in the compressed feature tensor; the height feature is extracted from the width-height coupling layer and the height-spacing coupling layer; and the spacing feature is extracted from the width-spacing coupling layer and the height-spacing coupling layer. The extracted width, height, and spacing features are weighted and summed using the influence weight coefficients to generate a fused feature vector. This fused feature vector is then mapped through an output projection layer to generate a comprehensive feature vector.

[0037] It should be noted that, as Figure 4 As shown, the elevation difference prediction layer is implemented using a multi-level category-aware graph neural network. The overall architecture of the network consists of a cascaded ordered category embedding module, a graph structure parameter spatial encoding module, a multi-level recommendation decision module, and an uncertainty estimation and output module. The ordered category embedding module transforms the ordered nature of the elevation difference layers into positional encoding, which serves as the semantic representation of the prediction target. The graph structure parameter spatial encoding module constructs a three-dimensional mesh graph from the discrete values ​​of building width, front building height, and building spacing, and performs structured encoding of parameter combinations through a graph neural network. The multi-level recommendation decision module divides the elevation difference layers into four semantic levels, first predicting the level affiliation, and then predicting the specific layer number within each level. The uncertainty estimation and output module calculates the prediction confidence, determines whether a special simulation suggestion is needed, and outputs the final quantitative guidance result for elevation difference ventilation.

[0038] In the ordered category embedding module, an ordered category-aware embedding space with a height difference layer number is constructed. The height difference layer number is a non-negative integer, and its value range is determined based on the height difference range covered by the experimental data. Preferably, the set of values ​​for the height difference layer number is as follows: In this system, 0 floors indicate that the rear buildings only need to be no taller than the front buildings to meet ventilation requirements. Floors 1 to 11 indicate the specific number of floors the rear buildings need to be higher than the front buildings. "Extra-high" indicates that the rear buildings need to exceed 100 meters in height, exceeding the scope of the quantitative guidelines, and a specific simulation is recommended. Positional coding is used to inject ordered information into the category representation. Sine-cosine positional coding is used to generate positional coding vectors, ensuring that the similarity between the coding vectors of any two floor numbers depends only on the difference in floor number. Therefore, the coding similarity between floors 3 and 4 is higher than that between floors 3 and 10, reflecting the ordered nature. For the "extra-high" category, all elements of the positional coding vector of the largest numerical category (11) are multiplied by a decay factor, and a marker bit is appended to the end of the vector to separate it from the numerical category coding space.

[0039] In the graph structure parameter space encoding module, discrete level combinations of width, height, and spacing are used as nodes. If two nodes differ by one level in a certain parameter dimension, the node pair that differs by one level in the parameter dimension is used as an edge to construct a three-dimensional mesh graph structure.

[0040] Message passing is performed in the graph using a graph neural network, with the initial representation of each node set as the comprehensive feature vector output in step S4. In the... In each layer, each node collects the representation of the previous layer from all its neighbors, generates a neighborhood aggregation vector through an aggregation function, and employs an attention aggregation mechanism to assign different attention weights to different neighbor nodes. These attention weights are determined by the node... With nodes The feature similarity determines the representation. The representation of the current node is fused with the neighborhood aggregation vector, and mapped to a new representation through a fully connected network. The fusion method employs a gating mechanism, calculating reset and update gates to control the proportion of self-information and neighborhood information retained. After layer message passing, the final representation of each node not only includes its own ventilation characteristics, but also incorporates information from neighboring nodes.

[0041] In the multi-level recommendation decision module, the number of elevation difference levels is divided into four levels: no elevation difference required, low elevation difference required, medium elevation difference required, and high elevation difference required. The no elevation difference required level means that under the current parameter combination, the ventilation requirement of 3m / s can be met by the buildings in front and behind being of equal height or the front buildings being taller, without the need for the rear buildings to be deliberately taller. The low elevation difference required level means that a small difference in the number of floors (about 3-9 meters in height) can effectively drive the airflow between buildings to meet the ventilation requirements. The medium elevation difference required level requires a moderate difference in the number of floors (about 12-21 meters in height) to overcome the ventilation obstruction caused by the large building width or small spacing. The high elevation difference required level requires the rear buildings to be significantly taller than the front buildings (more than 24 meters) to meet the ventilation requirements, or the height of the rear buildings needs to exceed 100 meters. In this case, the quantitative guidance no longer provides specific values, and it is recommended to conduct special CFD simulations or wind tunnel physical model experiments.

[0042] A hierarchical predictor consisting of three fully connected layers is constructed. Using the comprehensive feature vector as input, it outputs the probability distribution of each level to determine the semantic level to which the current parameter combination belongs. For levels requiring no elevation difference, the elevation difference level is directly output as 0. For levels with low elevation difference requirements, an ordered regression method is used, employing a three-layer fully connected network to output the number of consecutive, specific elevation difference levels. For levels with medium elevation difference requirements, a graph neural network-based recommendation method is used. The recommendation level is determined by retrieving the labels of neighboring nodes in the parameter combination graph and performing weighted voting, with the voting weight determined by the graph distance and feature similarity between nodes. For levels with high elevation difference requirements, if the number of elevation difference levels under the output parameter combination is greater than a preset threshold, a specific simulation suggestion is output.

[0043] In the uncertainty estimation and output module, the confidence level of the prediction results is calculated, and a neighborhood consistency check is performed. This involves calculating the degree of difference in the labels of the target node's neighboring nodes in the graph. When the confidence level is lower than a preset threshold or the neighborhood consistency is lower than a preset standard, a specific simulation suggestion is output instead of a specific number of elevation differences. In the specific simulation suggestion, if the predicted height of the rear buildings exceeds 100 meters, wind tunnel experiments are preferred; otherwise, CFD simulations are recommended. The confidence level of the prediction results is defined as the maximum probability value output by the hierarchical predictor. Neighborhood consistency is defined as the dispersion of the labels of all labeled nodes within the k-hop neighborhood of the target node in the graph. Label variance or information entropy is used as the metric. If the label variance within the neighborhood is greater than a preset threshold, it indicates that similar combinations in the parameter space have significantly different elevation requirements. The current parameter combination is in a sensitive area for ventilation performance; minor adjustments may lead to drastic performance changes, proving that the reliability of the prediction results is low.

[0044] The output of the elevation difference prediction layer adopts a structured format, outputting confidence level, minimum number of north-south building elevation difference layers, or special simulation suggestions, as quantitative guidance results for elevation difference ventilation.

[0045] The network training for the elevation difference prediction layer employs a multi-task joint learning strategy. Each sample in the training data includes building width level, front building height level, and building spacing level as input features, and the corresponding elevation difference level as the output label. For combinations labeled as "super high," the level label is set to the "high elevation difference demand layer," and it does not participate in the loss calculation for specific level prediction. Given the limited number of training samples, a parameter space interpolation method is used for data augmentation. Virtual samples are generated at the midpoint of the line connecting two adjacent labeled nodes in the graph, and the label is the mean of the labels at both endpoints, rounded to the nearest integer.

[0046] A multi-task joint loss function is adopted, consisting of a weighted average of four terms: hierarchical classification loss for the low-to-high-difference demand layer, ordered regression loss for the low-to-high-difference demand layer, and L2 regularization term. The weight coefficients of each loss are determined through grid search, with the optimal values ​​being: hierarchical classification loss weight 1.0, ordered regression loss weight 1.5, and regularization term weight 0.0005. Let the semantic hierarchy set be... These correspond to the following levels: no elevation difference required, low elevation difference required, medium elevation difference required, and high elevation difference required, respectively. The hierarchical classification loss uses cross-entropy loss to train the hierarchical predictor, and is expressed as: ,in For all training sample sets, For the sample Real-world hierarchical labels Semantic hierarchy output by the hierarchical predictor The probability. Samples from the high-difference demand level only participate in the calculation of hierarchical classification loss, not in the calculation of ordered regression loss. Ordered regression loss is only calculated for samples in the low-difference demand level with specific height difference level labels. The set of levels for the low-difference demand level is... The sample set is The ordered regression loss for the low-to-high demand hierarchy is expressed as: , The actual number of floors with elevation differences. These are the continuous predicted values ​​output by the model. number of layers Learnable threshold parameters The activation function is used. For the high-to-medium demand levels (layers 4-7), a weighted voting method based on graph neural networks is adopted, without setting a separate loss function.

[0047] The Adam optimizer was used during training, with an initial learning rate of 0.001, which decayed to 0.9 times the original rate every 50 training epochs. Since graph neural networks require complete graph structure information, a full-graph training approach was adopted, using the entire parameter combination graph as input for each training epoch. The preferred number of graph neural network layers was 3, the preferred position encoding dimension was 64, and the number of attention heads for neighbor aggregation was set to 4. An early stopping mechanism was employed: training was stopped and the optimal model parameters were saved when the hierarchical classification accuracy on the validation set no longer improved for 10 consecutive epochs. The parameter combinations were randomly divided into training and validation sets at a ratio of 80% and 20%, respectively, maintaining a relatively consistent sample proportion across all levels. Evaluation metrics included hierarchical classification accuracy, height difference layer prediction accuracy, neighbor error tolerance accuracy, and specialized simulated recommendation accuracy. Neighbor error tolerance accuracy was defined as the proportion where the predicted layer number differs from the actual layer number by no more than one layer.

[0048] For a new parameter combination, the parameter combination is mapped to the corresponding node in the graph structure (if the parameter value is continuous, it is first discretized to the nearest modulus level). Message passing in the graph neural network is executed to obtain information from neighboring nodes. The comprehensive feature vector is input into the hierarchical predictor to obtain the hierarchical distribution. If the maximum hierarchical probability is lower than the threshold, the output indicates that a specific simulation is needed; otherwise, the corresponding specific hierarchical predictor is selected based on the predicted hierarchical level, and the recommended height difference hierarchical level is output. Specific simulation is used when the model prediction confidence is low or the parameter combination is in a boundary condition. In this case, the specific height difference hierarchical level is not output, but rather a suggestion is made that computational fluid dynamics (CFD) numerical simulation or wind tunnel physical model experiment is needed to more accurately evaluate wind environment performance.

[0049] The second embodiment of the present invention provides a computer-readable storage medium, which includes a program for optimizing the form of high-density residential buildings based on elevation difference ventilation quantification guidelines. When the program for optimizing the form of high-density residential buildings based on elevation difference ventilation quantification guidelines is executed by a processor, it implements the steps of the method for optimizing the form of high-density residential buildings based on elevation difference ventilation quantification guidelines.

[0050] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation, characterized in that, Includes the following steps: Step S1: Obtain the building design parameter set, extract the category identifier vector of each parameter through the parameter dimension encoder, and extract the numerical vector of each parameter through the parameter value encoder. The building design parameter set includes building width, front building height and building spacing. Step S2: Input the category identifier vector and the numerical vector into the ventilation coupling calculation network to calculate the pairwise coupling strength between building width, front building height and building spacing, and generate a three-dimensional coupling feature tensor; Step S3: Perform low-rank compression on the three-dimensional coupled feature tensor along the three dimensions respectively, and then fuse the compressed features in the three directions into a compressed feature tensor through a ring product operation; Step S4: Input the compressed feature tensor into the adaptive gating unit, dynamically calculate the influence weight coefficients of the three parameters of building width, front building height and building spacing, and perform weighted summation of the feature components in the compressed feature tensor corresponding to each parameter according to the influence weight coefficients to generate a comprehensive feature vector; Step S5: Input the comprehensive feature vector into the elevation difference prediction layer, and output the minimum number of north-south building elevation difference layers that meet the ventilation requirements under the current parameter combination, as the elevation difference ventilation quantification guidance result.

2. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 1, characterized in that, Step S1 specifically includes: A parameter dimension vocabulary is constructed, which includes building width dimension, front row building height dimension, and building spacing dimension. The building design parameters to be processed are input into the parameter dimension encoder, and the index of the parameter in the dimension vocabulary is used as the query key to generate the category identifier vector of the parameter through independent hot encoding. The architectural design parameter values ​​are physically discretized and modularized to generate physical realization values ​​for each parameter. The architectural design parameter values ​​include building width, front building height, and building spacing. Physically perceived random mask perturbation operation is performed on the physical implementation value respectively. The perturbation is randomly determined by the set mask probability. If it is added, the offset is sampled and superimposed within the preset perturbation radius according to the probability distribution to generate perturbed parameter value. The perturbated parameter values ​​are input into the corresponding competitive discrete codebook for encoding. The matching degree of each perturbated parameter value is calculated with each codeword in the corresponding codebook. A preset number of codeword vectors are selected through a competitive selection mechanism as the encoded output of the parameters, which serves as the numerical vector of the parameters. The category identifier vector is combined with the numerical vector to obtain the complete encoded representation of each parameter.

3. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 1, characterized in that, Step S2 specifically includes: The category identifier vector and numerical vector of each parameter are respectively input into three physically isolated coupled calculation paths, including the wide-height coupling path, the wide-spacing coupling path, and the high-spacing coupling path; Within each coupling path, the two sets of input category identifier vectors are multiplied by an outer product to generate a category interaction matrix, and the two sets of input numerical vectors are multiplied by an outer product to generate a numerical interaction matrix. The category interaction matrix and the numerical interaction matrix are added element by element to obtain a combined feature matrix. Within each coupling path, linear interaction paths and non-linear interaction paths are executed in parallel. The linear interaction path transforms the two sets of numerical vectors linearly, multiplies them element by element, and then projects them to output the first interaction feature. The non-linear interaction path concatenates the two sets of category identifier vectors with the two sets of numerical vectors and outputs the second interaction feature through a multi-layer fully connected network. Within each coupling path, the first interaction feature and the second interaction feature are weighted and fused according to the physical prior interaction kernel corresponding to the coupling type to generate the coupling strength feature of the path.

4. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 3, characterized in that, Step S2 further includes: The coupling strength features of the three output paths are stacked along the coupling type dimension to generate a three-dimensional coupling feature tensor, where the output of the wide and high coupling paths is used as the first layer, the output of the wide spacing coupling path is used as the second layer, and the output of the high spacing coupling path is used as the third layer. The three-dimensional coupled feature tensor undergoes a second interaction in the coupling type dimension to calculate the mutual attention weights among width-height coupling, width-distance coupling, and height-distance coupling. Based on these attention weights, the features of each layer are reweighted to output the final three-dimensional coupled feature tensor.

5. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 1, characterized in that, Step S3 specifically includes: The three-dimensional coupled feature tensor is expanded along the three parameter dimensions of width, height, and spacing to generate a width direction view, a height direction view, and a spacing direction view. The three generated directional views are decomposed into low-rank matrices respectively. Each view matrix is ​​approximately decomposed into the product of two low-rank matrices. During the decomposition, the decomposition rank of each view is independently set according to the physical importance of each parameter dimension to the ventilation effect, and physical constraint regularization terms are introduced in the decomposition process. The low-rank decomposition results of the three directional views are fused by a circular product. The fusion order is executed step by step according to the cyclic path of width, height and spacing to generate the base mode tensor. The base mode tensor and the original 3D coupled feature tensor are condensed together. The dimensions of the base mode tensor are weighted and summed according to the weights provided by the original tensor to output the compressed feature tensor.

6. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 1, characterized in that, Step S4 specifically includes: A dual-input structure including a scene path and a feature path is constructed. The scene path receives the numerical vectors of each parameter and compresses them into a scene compressed vector through physical pooling. The feature path receives the compressed feature tensor. Based on the compressed feature tensor, the number of predicted elevation difference layers is obtained using a preset intermediate layer; Obtain the magnitude of the partial derivative of the predicted elevation difference layer number with respect to the numerical vector of each parameter as the first-order sensitivity. Obtain the sum of the magnitudes of the mixed partial derivatives of the predicted elevation difference layer number with respect to each parameter pair as the second-order sensitivity. The weighted sum of the first-order sensitivity and the second-order sensitivity is used as the comprehensive sensitivity of each parameter. The scene compression vector, the comprehensive sensitivity of each parameter, and the part of the compressed feature tensor related to each parameter are fused together to calculate the original scores of building width, front building height, and building spacing respectively. The three raw scores obtained from the calculation are input into a competitive normalization unit, and temperature scaling and Softmax normalization operations are performed to generate the influence weight coefficients of the three parameters.

7. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 6, characterized in that, Step S4 further includes: The width, height, and spacing features are extracted from the compressed feature tensor and mapped to feature vectors of fixed dimensions, respectively. The extracted width, height, and spacing features are weighted and summed using the aforementioned influence weight coefficients to generate a fused feature vector. This fused feature vector is then mapped through an output projection layer to generate a comprehensive feature vector.

8. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 1, characterized in that, The elevation difference prediction layer in step S5 is implemented using a multi-level category-aware graph neural network, specifically including: An ordered category-aware embedding space with varying height difference layers is constructed, wherein the height difference layer number is a non-negative integer. Sine-cosine position coding is used to generate position coding vectors, such that the similarity between coding vectors of any two layers depends only on the difference in layer number. A three-dimensional mesh graph structure is constructed using discrete level combinations of width, height, and spacing as nodes and node pairs that differ by one level in parameter dimension as edges. Message passing is performed in the graph through a graph neural network, so that each node in the graph obtains information from its neighboring nodes and updates its own representation. The elevation difference layers are divided into four levels: no elevation difference layer, low elevation difference layer, medium elevation difference layer, and high elevation difference layer. A level predictor is constructed, which takes the comprehensive feature vector as input and outputs the probability distribution of each level to determine the semantic level to which the current parameter combination belongs.

9. The method for optimizing the form of high-density residential buildings based on quantitative guidance for elevation difference ventilation according to claim 8, characterized in that, Step S5 further includes: For the low elevation difference demand level, an ordered regression method is used to predict the specific elevation difference level. For the medium elevation difference demand level, a recommendation method based on graph neural networks is used. The recommendation level is determined by retrieving the labels of neighboring nodes in the parameter combination graph and performing weighted voting. For the high elevation difference demand level, a special simulation suggestion is output. Calculate the confidence level of the prediction results and perform a neighborhood consistency check. Calculate the degree of difference in the labels of the target node's neighboring nodes in the graph. When the confidence level is lower than a preset threshold or the neighborhood consistency is lower than a preset standard, output a specific simulation suggestion instead of a specific number of elevation difference layers. Output the minimum number of floors required for north-south building elevation difference or a specific simulation suggestion, as a quantitative guideline for elevation difference ventilation.

Images