Renovated building surface radiation prediction and optimization method fused with meteorological radiation potential layer

Abstract

The application discloses a kind of regenerative building surface radiation prediction and optimization method of fusion meteorological radiation potential layer, comprising the following steps: S1, target building body is dispersed into regular three-dimensional voxel array;S2, target building surface solar radiation cumulative value is respectively mapped to corresponding surface voxel according to outer surface direction, and effective outer surface radiation label is identified by six-direction surface mask;S3, construct geometric enhancement feature field;S4, input tensor of target building surface radiation prediction model is constructed;S5, meteorological radiation potential accumulation layer is constructed;S6: target building surface radiation prediction model is established;S8, parameterized generation building body adjustment scheme;S9, quickly predict building surface radiation distribution;S10, building body adjustment scheme regenerative optimization.Compared with prior art, the application can quickly obtain the building surface radiation distribution result of different building body scheme, and can be used for building body regenerative optimization analysis in urban renewal.

Description

Technical Field

[0001] This invention relates to the field of building surface radiation prediction and optimization technology, and in particular to a method for predicting and optimizing regenerative building surface radiation by integrating a meteorological radiation potential layer. Background Technology

[0002] In urban renewal and the regeneration of existing buildings, the solar radiation distribution on the building form and its surface has a significant impact on the thermal environment, heat load, photothermal performance, and subsequent energy-saving renovation strategies. Traditional building surface radiation analysis typically relies on physical simulation software such as Radiance, Ladybug, and EnergyPlus to obtain radiation results through solar path calculation, ray tracing, and shading analysis. However, this method has the following five shortcomings in practical applications:

[0003] (1) Traditional methods rely heavily on physical ray tracing and hourly weather calculations. The calculation time is long in complex shapes, densely occluded scenes and large-scale scheme iterations, which makes it difficult to meet the needs of rapid feedback and high-frequency optimization in urban renewal and building regeneration design.

[0004] (2) Existing analysis processes are mostly based on single simulations and lack the ability to quickly predict changes in form combinations. When building form parameters, shading components or spatial relationships change, a complete physical simulation usually needs to be performed again, which seriously restricts the efficiency of scheme optimization.

[0005] (3) Research on building performance prediction based on deep learning often uses two-dimensional images, elevation drawings or overall performance indicators as input and output, which makes it difficult to effectively retain the three-dimensional spatial topological relationship of building volume, multi-directional occlusion depth characteristics and radiation distribution information facing different external surface directions.

[0006] (4) Existing prediction methods tend to focus on the direct fitting of geometric inputs to radiation results, and rarely introduce the geometric enhancement features of building surfaces and meteorological radiation potential information of the target period into the model at the same time. It is difficult to uniformly express the comprehensive influence of factors such as self-shading, spatial openness, solar incident direction, direct radiation intensity, sky diffuse radiation and surface accessibility on radiation distribution.

[0007] (5) Existing radiation prediction methods have not yet been combined with building form regeneration optimization, making it difficult to support the integrated application requirements of "form generation - radiation assessment - optimization iteration" in urban renewal scenarios. Summary of the Invention

[0008] This invention provides a method for predicting and optimizing the radiation of regenerative building surfaces by integrating meteorological radiation potential layers, in order to solve one or more problems in the background art.

[0009] This invention provides a method for predicting the surface radiation of regenerative buildings by incorporating meteorological radiation potential layers, comprising the following steps:

[0010] S1: Establish a three-dimensional voxel representation model of the target building, and discretize the shape of the target building into a regular three-dimensional voxel array;

[0011] S2: Map the cumulative solar radiation value of the target building surface to the corresponding surface voxels according to the outer surface direction, and identify the valid outer surface radiation label through a six-directional surface mask;

[0012] S3: Construct a geometrically enhanced feature field to characterize the local openness, self-occlusion relationship, and geometric state of edges and corners of the target building surface;

[0013] S4: The target building regular three-dimensional voxel array, six-directional surface mask and geometric enhancement feature field are channel-stitched to form the input tensor of the target building surface radiation prediction model in step S6;

[0014] S5: Calculate the solar direction, direct radiation potential, and diffuse radiation potential based on meteorological data consistent with the predicted target period, and construct a meteorological radiation potential accumulation layer;

[0015] S6: Establish a target building surface radiation prediction model that integrates the geometric enhancement feature field and the meteorological radiation potential accumulation layer. The target building surface radiation prediction model includes a data-driven radiation prediction branch and a physical approximation radiation branch. The physical approximation radiation branch is obtained by fusing the direct accessibility prediction branch, the sky accessibility prediction branch, and the meteorological radiation potential accumulation layer. The radiation prediction results output by the data-driven radiation prediction branch and the radiation prediction results output by the physical approximation radiation branch are fused to obtain the six-directional building surface radiation distribution prediction results.

[0016] Preferably, in step S2, the surface voxel is obtained through the following steps:

[0017] The voxels in the three-dimensional voxel array in step S1 are represented in a binary manner:

[0018] ,

[0019] When an entity voxel has at least one air neighbor in the six-neighbor direction, it is determined to be a surface voxel, and the voxel coordinates, outer surface direction and corresponding cumulative radiation value of the surface voxel are recorded.

[0020] The outer surface directions include six directions: +X, -X, +Y, -Y, +Z, and -Z, where x, y, and z are the three-dimensional spatial coordinates of voxel V.

[0021] Preferably, in step S3, the construction steps of the geometric enhancement feature field are as follows:

[0022] For each surface voxel, perform a neighborhood scan in j discrete spatial directions;

[0023] For each surface voxel, search outwards along each discrete direction and count the results at the preset maximum occlusion distance D. max Length of the contiguous airspace or distance to the nearest entity occlusion;

[0024] Each surface voxel forms a occlusion distance feature in the j-direction:

[0025] Occlusion(v) = [d1, d2, ..., d j ]

[0026] Where, d i Let d represent the cutoff occlusion distance or open distance in the i-th discrete direction, and 0 ≤ d. i ≤D max , 1≤i≤j, D max The maximum occlusion distance is the preset value, and v is the surface voxel.

[0027] Preferably, in step S4, the input tensor can be represented as:

[0028] X input =Concat(Vocc,M face 0,...,M face 5. Occlusion(v))

[0029] Where Vocc represents the target building regular three-dimensional voxel array, M face 0 to M face 5 represents the effective surface mask in the six outer surface directions, and Occlusion(v) represents the geometric enhancement feature field in the j-direction.

[0030] Preferably, in step S5, the meteorological radiation potential accumulation layer includes:

[0031] The solar direction at each moment is calculated based on meteorological data, and the solar direction is divided into several discrete solar direction sectors according to azimuth and elevation angles. For each outer surface direction d, a corresponding normal vector nd is established; for each moment t, a solar direction vector st is established.

[0032] Construct the directional radiation potential matrix P direct (d,k) can be represented as:

[0033]

[0034] Where DNI(t) represents the normal direct radiation at time t, and T kΔt represents the set of all times when the sun's direction falls into the k-th solar direction sector, and Δt represents the time step.

[0035] Constructing the potential for scattered radiation accumulated along the outer surface direction

[0036]

[0037] in, This represents all times within the target time period. Indicates time Sky-scattered radiation; Indicates the direction of the outer surface The corresponding unobstructed sky scattering reduction factor.

[0038] Preferably, in step S5, the meteorological radiation potential accumulation layer uses meteorological data for the time period in which the cumulative solar radiation value of the target building surface is located.

[0039] Preferably, in step S6, the data-driven radiation prediction branch outputs the data-driven radiation prediction value Rad. data (d,v), the direct reachability prediction branch output DA pred (d,v,k), whose values ​​range from 0 to 1; the sky reachability prediction branch output SA pred (d,v), whose value ranges from 0 to 1; based on the meteorological radiation potential accumulation layer, the physical approximate radiation branch The result is obtained by fusing the direct reachability prediction branch and the sky reachability prediction branch, and can be expressed as:

[0040]

[0041] Preferably, in step S6, the predicted results of the radiation distribution on the six-directional building surface are... It can be represented as:

[0042]

[0043] Where α is the fusion weight, which can be set as a fixed parameter or a learnable parameter.

[0044] The present invention also provides an optimized method for predicting the surface radiation of regenerated buildings, including the above-described method for integrating meteorological radiation potential layers, and further includes the following steps:

[0045] S8: By setting different building shape control parameters, multiple building shape adjustment schemes can be automatically generated;

[0046] S9: Adjust the building shape generated in S8 and input it into the target building surface radiation prediction model to obtain prediction data; evaluate the building radiation performance indicators based on the prediction data and generate a building surface radiation distribution cloud map, high radiation area identification results, building surface radiation statistical results, and radiation performance comparison results between different schemes;

[0047] S10: When there are high radiation areas, local heat load concentration areas, or uneven radiation distribution areas on the building surface, adjust the building shape control parameters.

[0048] Preferably, in step S8, the building shape control parameters include:

[0049] (1) Building height parameters;

[0050] (2) Building setback parameters;

[0051] (3) Parameters of building opening location;

[0052] (4) Parameters of building shading components;

[0053] (5) Building volume combination parameters;

[0054] (6) Parameters of local gaps in the building;

[0055] (7) Architectural regeneration morphological parameters;

[0056] In step S9, the building radiation performance indicators include:

[0057] (1) Average radiation value of building surface;

[0058] (2) Percentage of high-radiation areas;

[0059] (3) Local radiation peak;

[0060] (4) Surface radiation uniformity of the building;

[0061] (5) Surface radiation distribution characteristics under different orientations;

[0062] (6) Local heat load sensitive areas of the building;

[0063] In step S10, the adjustment of the building shape control parameters includes:

[0064] (1) Adjust the building setback relationship;

[0065] (2) Adjust the occlusion relationship of building volumes;

[0066] (3) Adjust the shape of the gaps in the local buildings;

[0067] (4) Adjust the building volume combination method.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] (1) The problem of predicting building surface radiation is transformed into a joint mapping problem between the three-dimensional spatial representation of buildings, the geometric enhancement feature field, the meteorological radiation potential accumulation layer and the surface radiation distribution tensor. This realizes the unified expression of complex building shape shading relationship, surface directional radiation label and meteorological solar direction potential, and improves the modeling ability of complex building spatial relationship and radiation influencing factors.

[0070] (2) A method for constructing the radiation distribution tensor of a six-directional building surface was established, which enables the cumulative radiation values ​​of the same surface voxel in different external surface directions to be expressed separately, avoiding the mutual overlap of multi-faceted radiation labels of edge or corner voxels;

[0071] (3) A geometric enhancement feature field construction method based on multi-directional occlusion distance statistics is proposed, which can enhance the prediction model’s ability to identify the openness of local space on building surface, self-shaped occlusion relationship, edge region and corner region, and improve the accuracy of radiation prediction of complex building surface.

[0072] (4) A modeling mechanism for meteorological radiation potential accumulation layer and surface radiation accessibility was proposed. Without repeating the complete ray tracing for each scheme to be evaluated, the direct and scattered radiation potentials were constructed using meteorological data of the target period. Radiation accumulation prediction was made by combining the direct accessibility of building surfaces and sky accessibility, thereby enhancing the model's comprehensive ability to express meteorological conditions, solar incident direction, sky scattering influence and building self-shading relationship.

[0073] (5) The present invention can quickly obtain the radiation distribution results of building surface of different building form schemes, which can be used for building form regeneration optimization analysis, building thermal environment optimization and rapid comparison of building schemes in urban renewal, thereby improving the efficiency of scheme iteration and auxiliary decision-making ability in the process of complex building renewal design. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0075] Figure 1 This is a flowchart of the process of the present invention;

[0076] Figure 2 This is a schematic diagram of the target building surface radiation prediction model of the present invention;

[0077] Figure 3 This is the actual radiation cloud map of the present invention;

[0078] Figure 4 This is a predicted radiation cloud map for the present invention;

[0079] Figure 5 This is a radiation difference cloud map of the present invention. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0081] See attached document Figure 1 This embodiment provides a method for predicting the surface radiation of a regenerated target building by fusing geometrically enhanced feature fields, including the following steps:

[0082] Step S1: 3D voxelization of the target building shape

[0083] A three-dimensional voxel representation model of the target building is established, and the three-dimensional shape of the target building is discretized into a regular three-dimensional voxel array.

[0084] Establish a three-dimensional voxel space of fixed size:

[0085] in:

[0086] X, Y, and Z represent the discrete quantities of the voxel space in the three directions, respectively;

[0087] Each voxel is represented in a binary format:

[0088]

[0089] in, denoted as the three-dimensional spatial coordinates of voxel V.

[0090] Specifically, in this embodiment, the following is adopted: A three-dimensional voxel array.

[0091] Step S2: Construct radiation distribution data for the target building surface

[0092] Based on target building radiation simulation software, cumulative solar radiation analysis was performed on the target building surface to obtain the cumulative radiation value of the target building surface during the summer extreme heat cycle. The cumulative solar radiation value of the target building surface was mapped to the corresponding surface voxels according to the outer surface direction, and the effective outer surface radiation label was identified by a six-directional surface mask. The specific steps are as follows:

[0093] First, identify the target building surface voxels: a voxel is identified as a surface voxel when it has at least one air neighbor in its six neighborhood directions; simultaneously, record the actual exposed outer surface direction of the voxel, which includes six directions: +X, -X, +Y, -Y, +Z, and -Z. For each surface voxel, record its voxel coordinates, outer surface direction, and corresponding cumulative radiation value.

[0094] The cumulative radiation value of surface voxels is obtained by mapping the radiation simulation results of the target building surface. Specifically, firstly, the outer surface (surface exposed to the outside air) of each entity voxel is extracted from the 3D model of the target building, generating corresponding surface simulation test points. Subsequently, the Ladybug plugin of the Rhino-Grasshopper platform is used to perform radiation simulation on the 3D model of the target building during the summer extreme heat week, obtaining the cumulative radiation value of each surface simulation test point during the target period, including contributions from direct radiation and sky diffuse radiation. Based on the voxel and outer surface orientation corresponding to each surface simulation test point, the cumulative radiation value is mapped to the corresponding surface voxel. When a surface voxel is located at an edge or corner and has multiple outer surface orientations, the cumulative radiation value of the voxel in different outer surface orientations is recorded separately, without merging the radiation values ​​from multiple orientations into a single value.

[0095] Construct the six-directional surface true radiation tensor Y corresponding to the voxel space true (d,v), where d represents the outer surface direction and v represents the voxel. A corresponding cumulative radiation value is assigned only if the voxel v has an effective outer surface in direction d, and this is verified by a six-directional surface mask M. face (d,v) indicates its validity; for non-surface regions or invalid directions, it is not included in the training error calculation.

[0096] In this way, the multifaceted radiation values ​​of the same edge or corner voxel can be expressed separately, avoiding the compression of the radiation values ​​of multiple outer surface directions of the voxel into a single scalar, which would cause training label errors.

[0097] Step S3: Construct the geometrically enhanced feature field

[0098] To enhance the ability of the target building surface radiation prediction model in step S6 to identify local shading relationships, spatial openness, and radiation differences in edge and corner areas of the target building surface, a multi-directional shading depth statistical field is constructed as a geometric enhancement feature field. In this embodiment, the number of discrete spaces j is 18.

[0099] For each surface voxel, a neighborhood scan is performed in 18 discrete spatial directions. The 18 discrete spatial directions include 6 axial directions and 12 face diagonal directions in the neighborhood of the 3D cube, excluding the volume diagonal directions.

[0100] Searching outwards from the current surface voxel in each direction, and counting the results at the preset maximum occlusion distance D. max The length of the contiguous airspace or the distance to the nearest occlusion. When in D max When a target building entity voxel is encountered within the range, the open distance before encountering the entity is recorded; when it exceeds D... max If the target building entity voxel is not encountered, the directional feature is uniformly denoted as D. max This indicates that the direction is not obscured by the object itself.

[0101] During feature computation, the finite boundaries of the voxel mesh are not considered occlusion boundaries. That is, when the search direction leaves the current voxel mesh and is within D... max If no target building entity is encountered within the range, occlusion is not considered to have occurred, and the maximum occlusion distance D is used instead. max This method avoids misjudging the computational domain boundary as self-occlusion by the target building.

[0102] Therefore, 18-directional occlusion distance features are formed for each surface voxel:

[0103] Occlusion(v) = [d1, d2, ..., d 18 ]

[0104] Where, d i Let d represent the cutoff / occlusion distance or open distance in the i-th direction, where 1 ≤ i ≤ 18 and 0 ≤ d. i ≤D max , where v is the surface voxel.

[0105] This geometrically enhanced feature field can characterize the degree of openness of the local space around the target building surface, the occlusion relationship of the self-form, the geometric state of the edges and corners, and the changes in the local occlusion gradient, which can enhance the learning ability of the target building surface radiation prediction model to the differences in radiation distribution on the target building surface.

[0106] Step S4: Construct the 3D input tensor

[0107] The target building's regular three-dimensional voxel array, six-directional surface mask, and 18-directional geometric enhancement feature field are channel-stitched together to form the input tensor of the target building's surface radiation prediction model.

[0108] In this embodiment, the input tensor includes 25 channels, which are 1 target building entity placeholder channel, 6 outer surface orientation mask channels, and 18 geometry enhancement feature channels.

[0109] The input tensor can be represented as:

[0110] X input =Concat(Vocc,M face 0,...,M face 5. Occlusion(v))

[0111] Where Vocc represents the target building regular three-dimensional voxel array, M face 0 to M face 5 represents the effective surface mask in the six outer surface directions, and Occlusion(v) represents the 18-direction geometric enhancement feature field.

[0112] Step S5: Constructing a meteorological radiation potential accumulation layer

[0113] To enable the prediction process to represent the influence of solar direction and meteorological radiation conditions on the radiation distribution of the target building surface, a meteorological radiation potential accumulation layer is constructed. This layer accumulates the theoretical radiation potential of each outer surface direction under different solar direction sectors based on meteorological data consistent with the training label time range.

[0114] In this embodiment, the radiation tag corresponds to the cumulative radiation during the summer extreme heat week, and the meteorological radiation potential accumulation layer uses meteorological data within the same time range, such as hourly meteorological records from 0:00 on July 13 to 23:00 on July 19, instead of using annual meteorological data. The meteorological data can be derived from EPW format meteorological files or other data sources containing parameters required for calculating direct radiation, diffuse radiation, and solar position.

[0115] First, the solar direction at each moment is calculated based on meteorological data, and the solar direction is divided into several discrete solar direction sectors according to azimuth and elevation angles, with the number of discrete solar direction sectors being K; for each outer surface direction d, a corresponding normal vector nd is established; for each moment t, a solar direction vector st is established.

[0116] For the direct radiation component, construct the directional radiation potential matrix P. direct (d,k) can be represented as:

[0117]

[0118] Where DNI(t) represents the normal direct radiation at time t, T k Indicates the direction of the sun falling into the first The set of all times for a solar sector, 1≤ ≤K, where Δt represents the time step.

[0119] For the scattered radiation component, the potential for scattered radiation accumulated along the outer surface direction can be further constructed:

[0120]

[0121] in This indicates all times within the target period, such as July 13 to July 19; Indicates time Sky-scattered radiation; Indicates the direction of the outer surface The corresponding unobstructed sky scattering reduction factor.

[0122] The meteorological radiation potential accumulation layer does not perform scheme-by-scheme ray tracing on the shape of the target building to be predicted, nor does it directly calculate the real sky visibility factor. Instead, it pre-accumulates the solar directional radiation potential within the same meteorological period into a directional potential matrix that can be called by the neural network, thereby reducing the amount of repetitive calculations during large-scale scheme iterations.

[0123] This results in the cumulative meteorological radiation potential associated with the outer surface direction and the solar direction sector: P(d,k)={P direct (d,k),P diffuse (d)}. This result can be used as input for fixed meteorological conditions during the training and forecasting phases, or it can be recalculated based on different cities, seasons, or target time periods.

[0124] Step S6: Construct a radiation prediction model for the target building surface

[0125] To achieve a joint mapping between the three-dimensional spatial representation of the target building, the geometrically enhanced feature field, and the accumulated results of meteorological radiation potential, a surface radiation prediction model for the target building is constructed. This embodiment selects a three-dimensional convolutional encoder-decoder network as the feature extraction backbone; specifically, a 3D U-Net structure can be used, but this network structure is only an optional implementation and does not constitute a limitation of the present invention.

[0126] The target building surface radiation prediction model includes a data-driven radiation prediction branch and a physical approximation radiation branch:

[0127] (1) Data-driven radiation prediction branch outputs data-driven radiation prediction value Rad. data (d,v) is used to learn the spatial distribution of radiation on the surface of the target building obtained by direct supervision from the training samples.

[0128] (2) The physical approximation radiation branch is composed of the direct reachability prediction branch and the sky reachability prediction branch. The direct reachability prediction branch outputs DA. pred (d,v,k), whose value ranges from 0 to 1, is used to represent the degree of direct or obstructed sunlight accessibility of voxel v relative to the k-th solar sector in the outer surface direction d; the sky accessibility prediction branch output SA pred (d,v), whose value ranges from 0 to 1, is used to represent the degree of sky openness or the degree of scattered radiation reachable by voxel v in the direction d on the outer surface.

[0129] Based on the accumulation layer of meteorological radiation potential, a differentiable physical approximate radiation branch can be formed:

[0130]

[0131] The predicted radiation distribution of building surfaces in six directions is obtained by fusing the data-driven radiation prediction branch and the physical approximation radiation branch:

[0132]

[0133] Where α is the fusion weight, which can be set as a fixed parameter or a learnable parameter.

[0134] In this embodiment, the input tensor size is 25×30×30×16, and the six-directional radiation output tensor size is 6×30×30×16; the direct reachability branch further includes K solar direction sector dimensions, while the sky reachability branch does not include solar direction sector dimensions. The above dimensions, number of channels, and number of solar sectors can all be adjusted according to the application scenario.

[0135] For details, please refer to the appendix. Figure 2 The operation process of the target building surface radiation prediction model is as follows:

[0136] First, the target building's regular 3D voxel array, 18-directional geometry-enhanced feature field, and six-directional surface mask are combined into a geometry-enhanced input tensor with dimensions of 25×30×30×16, and then input into the 3D encoder-decoder backbone. This embodiment uses a 3D U-Net structure.

[0137]

[0138] In the encoding stage, the building geometry, spatial occlusion relationship and surface orientation information are extracted step by step through enc1 double convolution, pool1 three-dimensional max pooling, enc2 double convolution, pool2 three-dimensional max pooling and enc3 double convolution to obtain the high-rise spatial features.

[0139] Subsequently, in the decoding stage, the feature resolution is restored by upsampling up2, and then skipped and concatenated with the feature e2 from the encoding stage. After double convolution processing by dec2, it is upsampled again by up1 and skipped and concatenated with the feature e1 from the encoding stage. Finally, a shared feature with a size of 16×30×30×16 is obtained by double convolution by dec1.

[0140] Then, the shared features are input into three output branches (data-driven radiation prediction branch, direct reachability prediction branch, and sky reachability prediction branch). Specifically, the data-driven radiation prediction branch outputs the data-driven radiation prediction value Radiation_data(d,v) for each surface voxel in different outer surface directions through a 1×1×1 convolution, i.e., Rad. data (d,v), with dimensions of 6×30×30×16; the direct accessibility prediction branch outputs the direct accessibility prediction value DirectAccess_pred(d,v,k) for each surface voxel under different external surface directions and solar direction sectors through convolution, Sigmoid activation, and dimension restructuring, i.e., DA. pred (d,v,k), with dimensions of 6×K×30×30×16; the sky accessibility prediction branch, activated by convolution and sigmoid, outputs the sky accessibility prediction value SkyAccess_pred(d,v) for each surface voxel in different outer surface directions, i.e., SA. pred (d,v), with dimensions of 6×30×30×16. In this embodiment, the solar direction is divided into 24 sectors.

[0141] Based on meteorological data, the direct radiation potential Potential_direct(d,k) corresponding to different outer surface orientations and solar orientation sectors is calculated, i.e., the directional radiation potential matrix P. direct (d,k), and calculate the potential diffuse radiation (Potential_diffuse(d)) corresponding to different outer surface directions, i.e. Potential_diffuse(d) is broadcast to the voxel space according to the outer surface direction.

[0142] Subsequently, based on the meteorological radiation potential accumulation layer, DirectAccess_pred(d,v,k) is combined with the corresponding Potential_direct(d,k) item by item, and accumulation is performed for all solar direction sectors; simultaneously, SkyAccess_pred(d,v) is combined with the Potential_diffuse(d) for the corresponding outer surface direction. Then, the accumulated direct radiation portion and the scattered radiation portion are added together to obtain the meteorological physical approximate radiation Radiation_physics(d,v), i.e. .

[0143] Finally, according to the fusion weight α, the data-driven radiation prediction value Radiation_data(d,v) and the meteorological physical approximate radiation Radiation_physics(d,v) are weighted and fused to obtain the final fused radiation prediction result Radiation_final(d,v) for each surface voxel in different external surface directions, which is the prediction result of the radiation distribution of the building surface in six directions. Its dimensions are 6×30×30×16.

[0144] In this embodiment, the feature extraction backbone includes an encoding path, a bottleneck feature extraction layer, a decoding path, and a skip connection structure, which are used to extract the overall spatial relationship, local occlusion features, and surface directional features of the target building.

[0145] The encoding path is used to extract the spatial occlusion relationships, volumetric composition relationships, and local surface geometric features of the target building's 3D shape step by step. Each level of encoding unit may include a 3D convolutional layer, a nonlinear activation layer, and a downsampling layer.

[0146] The decoding path is used to restore the low-resolution spatial features to the original spatial dimensions step by step, and fuse them with the encoded features of the corresponding scale through skip connections to restore the surface boundary, edges, corners and local radiation variation features of the target building.

[0147] Through the above structural design, the target building surface radiation prediction model can simultaneously learn the global shading relationship of the target building shape, the directional geometric features of the target building surface, the local radiation changes in the edge and corner areas, and the influence of the meteorological solar directional radiation potential on the surface radiation distribution.

[0148] In this embodiment, 3D U-Net is only used as one of the selected models for the 3D convolutional encoder-decoder backbone. In actual implementation, other 3D convolutional networks, encoder-decoder networks, or deep learning models with 3D spatial mapping capabilities can also be used.

[0149] Step S7: Train the target building surface radiation prediction model

[0150] The constructed dataset is divided into a training set and a validation set. In this embodiment, the training set accounts for 90% and the validation set accounts for 10%. DataLoader is used for batch reading and random shuffling during training. Samples lacking effective radiation labels are removed during the training data construction stage.

[0151] The Adam optimizer is used to update parameters during model training.

[0152] Since the effective radiation value of the target building surface only exists in the direction of the actually exposed target building surface, the prediction error based on the six-directional surface mask is calculated only for the effective outer surface direction during training. In the six-directional surface mask M... face In (d,v), when voxel v has an effective outer surface radiation tag in direction d, M face (d,v)=1, otherwise M face (d,v)=0.

[0153] To enhance the model's ability to learn edge, corner, and top-face corner regions, a surface region weighting coefficient w(d,v) is further set. When a voxel has only one outer surface direction, w(d,v)=1; when a voxel has two outer surface directions, i.e., located in the edge region, w(d,v)=2; when a voxel has three or more outer surface directions, i.e., located in the corner region, w(d,v)=4; when a voxel is located in the corner region and the current prediction direction is the +Z top surface direction, w(d,v)=8.

[0154] The loss function employs a weighted surface mean square error, applied to the final fusion result, the data-driven radiation prediction branch, and the physical approximation radiation branch, respectively. Its weighted surface mean square error can be expressed as:

[0155]

[0156] in, The radiation prediction tensor participating in the calculation of the current loss term can be taken as follows: , or .

[0157] Furthermore, the total training loss can be expressed as:

[0158]

[0159] Where L final Corresponding Rad final L data Corresponding Rad data L physics Corresponding Rad physics L reg Optional regular expression terms.

[0160] During training, the input tensor containing the geometrically enhanced feature field, the accumulated meteorological radiation potential, and the true radiation distribution data in six directions are first input into the prediction model to obtain Rad. data DA pred SA pred Rad physics and Rad finalThe prediction error is then calculated based on the weighted loss function described above, and backpropagation is used to update the network parameters. After each training round, the validation loss is calculated using the validation set, and the changes in training loss and validation loss are recorded. When the validation loss is less than the current optimal validation loss, the corresponding model parameters are saved to form the optimal target building surface radiation prediction model.

[0161] After training, the system can quickly predict the radiation distribution on the surface of a target building and output the radiation distribution results, the comparison between the actual and predicted values, the error distribution results, and optional intermediate results for direct sunlight accessibility and sky accessibility.

[0162] This embodiment also provides an optimization method for predicting the radiation of regenerated target building surfaces by fusing geometrically enhanced feature fields, comprising the following steps:

[0163] Step S8: Parametrically generate building shape adjustment scheme

[0164] A module for generating and adjusting building shapes is constructed based on parametric modeling methods.

[0165] By setting different building shape control parameters, multiple building shape adjustment schemes can be automatically generated.

[0166] Building form control parameters include:

[0167] (1) Building height parameters;

[0168] (2) Building setback parameters;

[0169] (3) Parameters of building opening location;

[0170] (4) Parameters of building shading components;

[0171] (5) Building volume combination parameters;

[0172] (6) Parameters of local gaps in the building;

[0173] (7) Architectural regeneration morphological parameters.

[0174] Different combinations of parameters can create different architectural shape adjustment schemes.

[0175] The generated building shape is further converted into a corresponding three-dimensional spatial expression, and an input tensor consistent with the training phase is constructed; when meteorological conditions, city location or target time period change, the corresponding meteorological radiation potential accumulation results can be updated synchronously.

[0176] Another implementation method for adjusting the building form is through rule generation, random generation, genetic generation, or other building digitization methods.

[0177] Step S9: Rapidly predict the radiation distribution on the building surface

[0178] Input the building shape adjustment scheme generated in step S8 into the trained target building surface radiation prediction model, and call the meteorological radiation potential accumulation layer corresponding to the target time period.

[0179] Rapidly obtain prediction results of radiation distribution on building surfaces Rad final (d,v), and simultaneously output the results of the data-driven radiation prediction branch, the physical approximation radiation branch, the direct sunlight accessibility prediction branch, and the sky accessibility prediction branch for analysis. The prediction data includes the building surface radiation distribution prediction results Rad. final (d,v), data-driven radiation prediction branch, physical approximation radiation branch, direct accessibility prediction branch, and sky accessibility prediction branch.

[0180] Further evaluation of building radiation performance indicators based on predicted data.

[0181] Building radiation performance indicators include:

[0182] (1) Average radiation value of building surface;

[0183] (2) Percentage of high-radiation areas;

[0184] (3) Local radiation peak;

[0185] (4) Surface radiation uniformity of the building;

[0186] (5) Surface radiation distribution characteristics under different orientations;

[0187] (6) Local heat load sensitive areas of the building.

[0188] And simultaneously generate:

[0189] (1) Radiation distribution cloud map of building surface;

[0190] (2) Results of high-radiation area identification;

[0191] (3) Statistical results of radiation from building surfaces;

[0192] (4) Comparison of radiation performance between different schemes.

[0193] Compared with traditional successive physical simulation methods, this invention can avoid repeating complete ray tracing and hourly radiation simulation on each new scheme, significantly shortening the analysis time of complex building shape adjustment schemes and improving the ability to quickly analyze multiple schemes.

[0194] Step S10: Regeneration and optimization of building form adjustment scheme

[0195] See attached document Figure 3-5The system evaluates building radiation performance indicators based on predicted data, and then iteratively adjusts building shape control parameters according to the evaluation results. Specifically, when there are high-radiation areas, localized areas of concentrated heat load, or areas with uneven radiation distribution on the building surface, the corresponding building shape control parameters are automatically adjusted.

[0196] Adjustments to building form control parameters include:

[0197] (1) Adjust the building setback relationship;

[0198] (2) Adjust the occlusion relationship of building volumes;

[0199] (3) Adjust the shape of the gaps in the building;

[0200] (4) Adjust the building volume combination method.

[0201] The adjusted building form scheme is re-expressed in three-dimensional space and its geometrically enhanced feature field is constructed. This revised scheme is then input again into the target building surface radiation prediction model for rapid prediction. Through an iterative optimization process of "form generation - feature construction - meteorological potential invocation - radiation prediction - performance evaluation - parameter adjustment," a building regeneration scheme that meets the target radiation performance requirements is gradually obtained. This invention enables rapid optimization and iteration of building form schemes without the need for repeated large-scale physical radiation simulations, improving the efficiency of scheme optimization and decision support capabilities for complex building regeneration designs in urban renewal scenarios.

[0202] The building surface radiation data and meteorological radiation potential data in this invention can be obtained through the following methods: numerical simulation, measured data, EPW meteorological files, other meteorological data sources, machine learning prediction results, or other building environment analysis methods.

[0203] Compared with the prior art, the present invention has the following beneficial effects:

[0204] (1) The problem of predicting building surface radiation is transformed into a joint mapping problem between the three-dimensional spatial representation of buildings, the geometric enhancement feature field, the meteorological radiation potential accumulation layer and the surface radiation distribution tensor. This realizes the unified expression of complex building shape shading relationship, surface directional radiation label and meteorological solar direction potential, and improves the modeling ability of complex building spatial relationship and radiation influencing factors.

[0205] (2) A method for constructing the radiation distribution tensor of a six-directional building surface was established, which enables the cumulative radiation values ​​of the same surface voxel in different external surface directions to be expressed separately, avoiding the mutual overlap of multi-faceted radiation labels of edge or corner voxels;

[0206] (3) A geometric enhancement feature field construction method based on multi-directional occlusion distance statistics is proposed, which can enhance the prediction model’s ability to identify the openness of local space on building surface, self-shaped occlusion relationship, edge region and corner region, and improve the accuracy of radiation prediction of complex building surface.

[0207] (4) A modeling mechanism for meteorological radiation potential accumulation layer and surface radiation accessibility was proposed. Without repeating the complete ray tracing for each scheme to be evaluated, the direct and scattered radiation potentials were constructed using meteorological data of the target period. Radiation accumulation prediction was made by combining the direct accessibility of building surfaces and sky accessibility, thereby enhancing the model's comprehensive ability to express meteorological conditions, solar incident direction, sky scattering influence and building self-shading relationship.

[0208] (5) The present invention can quickly obtain the radiation distribution results of building surface of different building form schemes, which can be used for building form regeneration optimization analysis, building thermal environment optimization and rapid comparison of building schemes in urban renewal, thereby improving the efficiency of scheme iteration and auxiliary decision-making ability in the process of complex building renewal design.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the surface radiation of regenerative buildings by integrating meteorological radiation potential layers, characterized in that, Includes the following steps: S1: Establish a three-dimensional voxel representation model of the target building, and discretize the shape of the target building into a regular three-dimensional voxel array; S2: Map the cumulative solar radiation value of the target building surface to the corresponding surface voxels according to the outer surface direction, and identify the valid outer surface radiation label through a six-directional surface mask; S3: Construct a geometrically enhanced feature field to characterize the local openness, self-occlusion relationship, and geometric state of edges and corners of the target building surface; S4: The target building regular three-dimensional voxel array, six-directional surface mask and geometric enhancement feature field are channel-stitched to form the input tensor of the target building surface radiation prediction model in step S6; S5: Calculate the solar direction, direct radiation potential, and diffuse radiation potential based on meteorological data consistent with the predicted target period, and construct a meteorological radiation potential accumulation layer; S6: Establish a target building surface radiation prediction model that integrates the geometric enhancement feature field and the meteorological radiation potential accumulation layer. The target building surface radiation prediction model includes a data-driven radiation prediction branch and a physical approximation radiation branch. The physical approximation radiation branch is obtained by fusing the direct accessibility prediction branch, the sky accessibility prediction branch and the meteorological radiation potential accumulation layer. The radiation prediction results output by the data-driven radiation prediction branch and the radiation prediction results output by the physical approximation radiation branch are fused to obtain the radiation distribution prediction results of the building surface in six directions.

2. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 1, characterized in that, In step S2, the surface voxel is obtained through the following steps: The voxels in the three-dimensional voxel array in step S1 are represented in a binary manner: ， When an entity voxel has at least one air neighbor in the six-neighbor direction, it is determined to be a surface voxel, and the voxel coordinates, outer surface direction and corresponding cumulative radiation value of the surface voxel are recorded. The outer surface directions include six directions: +X, -X, +Y, -Y, +Z, and -Z. denoted as the three-dimensional spatial coordinates of voxel V.

3. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 1, characterized in that, In step S3, the construction steps of the geometric enhancement feature field are as follows: For each surface voxel, perform a neighborhood scan in j discrete spatial directions; For each surface voxel, search outward along each discrete direction and count the results at the preset maximum occlusion distance D. max Length of the contiguous airspace or distance to the nearest entity occlusion; Each surface voxel forms a occlusion distance feature in the j-direction: Occlusion(v)=[d1,d2,...,d j ] Where, d i Let d represent the cutoff occlusion distance or open distance in the i-th discrete direction, and 0 ≤ d. i ≤D max , 1≤i≤j, D max The maximum occlusion distance is the preset value, and v is the surface voxel.

4. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 3, characterized in that, In step S4, the input tensor can be represented as: X input =Concat(Vocc,M face 0,...,M face 5,Occlusion(v)) Where Vocc represents the target building regular three-dimensional voxel array, M face 0 to M face 5 represents the effective surface mask in the six outer surface directions, and Occlusion(v) represents the geometric enhancement feature field in the j-direction.

5. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 1, characterized in that, In step S5, the direct radiation potential and diffuse radiation potential in the meteorological radiation potential accumulation layer include: The solar direction at each moment is calculated based on meteorological data, and the solar direction is divided into several discrete solar direction sectors according to azimuth and elevation angles. For each outer surface direction d, a corresponding normal vector nd is established; for each moment t, a solar direction vector st is established. Construct the directional radiation potential matrix P direct (d,k) can be represented as: Where DNI(t) represents the normal direct radiation at time t, T k Δt represents the set of all times when the sun's direction falls into the k-th solar direction sector, and Δt represents the time step. Constructing the potential for scattered radiation accumulated along the outer surface direction in, This represents all times within the target time period. Indicates time Sky-scattered radiation; Indicates the direction of the outer surface The corresponding unobstructed sky scattering reduction factor.

6. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 5, characterized in that, In step S5, the meteorological radiation potential accumulation layer uses meteorological data for the time period in which the cumulative solar radiation value of the target building surface is located.

7. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 1, characterized in that, In step S6, the data-driven radiation prediction branch outputs the data-driven radiation prediction value Rad. data (d,v), the direct reachability prediction branch output DA pred (d,v,k), whose values ​​range from 0 to 1; the sky reachability prediction branch output SA pred (d,v), whose value ranges from 0 to 1; based on the meteorological radiation potential accumulation layer, the physical approximate radiation branch The result is obtained by fusing the direct reachability prediction branch and the sky reachability prediction branch, and can be expressed as:

8. The method for predicting the surface radiation of regenerated buildings by integrating meteorological radiation potential layers according to claim 7, characterized in that, In step S6, the predicted results of the radiation distribution on the six-directional building surface are obtained. It can be represented as: Where α is the fusion weight, which can be set as a fixed parameter or a learnable parameter.

9. An optimized method for predicting the surface radiation of regenerative buildings, comprising the method for integrating meteorological radiation potential layers as described in any one of claims 1-8, characterized in that, Includes the following steps: S8: By setting different building shape control parameters, multiple building shape adjustment schemes can be automatically generated; S9: Adjust the building shape generated in S8 and input it into the target building surface radiation prediction model to obtain prediction data; The radiation performance indicators of buildings are evaluated based on the predicted data, and a radiation distribution cloud map of building surface, high radiation area identification results, building surface radiation statistics results, and radiation performance comparison results between different schemes are generated. S10: When there are high radiation areas, local heat load concentration areas, or uneven radiation distribution areas on the building surface, adjust the building shape control parameters.

10. The optimization method of the regenerative building surface radiation prediction method integrating meteorological radiation potential layer according to claim 9, characterized in that, In step S8, the building shape control parameters include: (1) Building height parameters; (2) Building setback parameters; (3) Parameters of building opening location; (4) Parameters of building shading components; (5) Building volume combination parameters; (6) Parameters of local gaps in the building; (7) Architectural regeneration morphological parameters; In step S9, the building radiation performance indicators include: (1) Average radiation value of building surface; (2) Percentage of high-radiation areas; (3) Local radiation peak; (4) Surface radiation uniformity of the building; (5) Surface radiation distribution characteristics under different orientations; (6) Local heat load sensitive areas of the building; In step S10, the adjustment of the building shape control parameters includes: (1) Adjust the building setback relationship; (2) Adjust the occlusion relationship of building volumes; (3) Adjust the shape of the gaps in the local buildings; (4) Adjust the building volume combination method.

Images