Parametric design method for building envelope based on forward mapping and reverse generation

Abstract

The application discloses a building envelope parameterized design method based on forward mapping and reverse generation, and relates to the technical field of building envelope design, which comprises the following steps: constructing a rule set containing a material set allowed in each layer and a material compatibility matrix of adjacent layers; training a neural network to be a forward mapping model with the thickness and material code of each layer as input and the thermal parameter as output; determining the target thermal parameter according to the building performance requirement; inputting the target thermal parameter into the condition generation model of the non-light-transmitting envelope structure, and decoding from the outside to the inside layer by layer, generating an allowed mask according to the allowed material set, generating a dynamic mask according to the material type determined in the previous layer and searching the compatibility matrix, applying the two masks to the current layer material category to make the illegal material probability zero to obtain the alternative scheme; under the constraint of the rule set, the light-transmitting envelope structure traverses the legal material and thickness combination to filter the alternative scheme with the minimum deviation from the target thermal parameter; and finally, the alternative scheme is input into the forward mapping model to check the output final recommended scheme.

Description

Technical Field

[0001] This invention relates to the field of building energy conservation technology, specifically to a parametric design method for building envelope based on forward mapping and reverse generation. Background Technology

[0002] Building envelope is a crucial element in building energy efficiency design. In tropical climates with high temperature and humidity, building envelopes need to simultaneously meet performance requirements in multiple aspects, including thermal insulation, shading, lighting, and ventilation. The design of the building envelope involves several stages: material selection, structural assembly, determination of thermal parameters, and building performance assessment. Material selection includes choosing the main wall materials, insulation materials, finishing materials, and window glass types. Structural assembly includes the layering order and thickness configuration of materials. Thermal parameters include heat transfer coefficient, solar heat gain coefficient, visible light transmittance, shading coefficient, and thermal inertia. Building performance encompasses multiple dimensions such as operational energy consumption, thermal comfort, humidity comfort, and light comfort. These design stages are interconnected by complex nonlinear relationships; adjusting any single parameter can trigger a chain reaction of changes in multiple performance indicators.

[0003] Currently, the design of building envelopes typically involves designers first determining structural parameters such as the materials and thicknesses of each wall layer, window types, and window-to-wall ratios based on personal experience or standard drawings. Then, steady-state heat transfer formulas or energy simulation software are used to calculate the corresponding thermal parameters and building performance indicators. If the calculation results do not meet the design objectives, some parameters are manually modified and recalculated, iterating repeatedly until the results fall within the target range. In engineering practice, multi-parameter combined design models established through secondary development of Revit software can automate the calculation of thermal parameters, carbon emissions, and costs of building envelopes. However, the parametric design of the four types of building envelopes—exterior walls, roofs, floors, and windows—involves nearly 50 main parameters. Adding other characteristic parameters brings the total number of parameters in the design model to nearly 70, resulting in exceptionally complex model construction, long simulation times, and difficulties in obtaining quantitative data.

[0004] However, existing building envelope design technologies still face three pressing problems. First, the mapping relationship between material physical parameters, structural layer combinations, and the thermal parameters of the building envelope is discrete. Designers cannot obtain the precise thermal performance corresponding to any structural scheme in real time. Each change in the scheme requires rerunning time-consuming simulation software or manually substituting simplified formulas, resulting in very low efficiency in forward solving. Second, when the design task specifies a target thermal performance range, the lack of a systematic method for deriving feasible structural schemes from performance indicators forces designers to rely on experience to guess several candidate schemes and then verify them one by one. The backward derivation process almost entirely depends on manual trial and error, compromising the scientific validity and coverage of the generated schemes. Third, there is a trade-off between building operating energy consumption and multiple performance objectives such as thermal comfort, humidity comfort, and light comfort. The trial-and-error process used in existing design technologies struggles to simultaneously balance these mutually constraining objectives within a vast space of structural schemes, often resulting in design outcomes that only meet one aspect of the requirements at the expense of other performance aspects. In addition, since the wall has a certain heat storage function, there is a quantitative deviation between the simulation results modeled directly with the heat transfer coefficient K value and the results of the integrated simulation from materials to structure. The measured data shows that the difference between the two in terms of annual energy consumption can reach more than 250 GJ. How to find and fit this deviation law is also a key problem that needs to be solved urgently. Summary of the Invention

[0005] This invention proposes a parametric design method for building envelope based on forward mapping and backward generation, which solves the problem that existing building envelope design methods are difficult to efficiently derive feasible construction schemes from target thermal parameters under multi-objective performance constraints.

[0006] To address the aforementioned technical problems, this invention provides a parametric design method for building envelopes based on forward mapping and backward generation, comprising the following steps: Step S1: Construct a set of construction rules for the enclosure structure, which includes the set of permissible materials for each level of the enclosure structure and the material compatibility matrix of adjacent levels; Step S2: Using the thickness and material codes of each layer of the building envelope as input and thermal parameters as output, train a neural network as a forward mapping model; determine the target thermal parameters according to the building performance requirements. Step S3: For non-transparent enclosure structures, the target thermal parameters are used as conditions and input into the trained conditional generation model. The decoding end of the conditional generation model performs autoregressive decoding layer by layer from the outside to the inside of the enclosure structure. The decoding output of each layer generates an allowable mask based on the allowable material set, and generates a dynamic mask by querying the material compatibility matrix based on the material type determined in the previous layer. The allowable mask and the dynamic mask are applied to the original logical value of the current layer material category prediction branch output, so that the output probability of materials that do not belong to the allowable material set or are incompatible with the materials in the previous layer is zero, thus obtaining alternative construction schemes. For a light-transmitting enclosure structure, under the constraints of the construction rule set, legal material and thickness combinations are traversed, and the combination with the smallest deviation from the target thermal parameters is selected as the alternative construction scheme. Step S4: Input the alternative construction schemes into the forward mapping model for forward verification and then output the final recommended scheme.

[0007] Preferably, the conditional generation model in step S3 is a conditional variational autoencoder. During the training phase, the encoder concatenates the thermal parameter conditional vector with the real construction scheme vector and outputs the mean and logarithmic variance of the latent space variables through a fully connected layer. The training loss function includes a reconstruction error term, a KL divergence regularization term, and a physical consistency check term. The physical consistency check term inputs the construction scheme vector generated by the decoder into the forward mapping model to obtain the predicted thermal parameter values, extracts the first three terms of the predicted thermal parameter values, calculates the sum of squared deviations between the first three terms and the thermal parameter conditional vector, and includes the sum of squared deviations in the loss calculation.

[0008] Preferably, the expression for the training loss function is: ; ; In the formula, For the first i Reconstruction error of each sample; For the first i The true construction scheme vector of each sample; The first generation generated for decoder reconstruction i A vector of construction schemes for each sample; It is a conditional vector; It is a forward mapping model; Let KL divergence be a metric. The posterior distribution for encoder parameterization; It follows a standard normal prior distribution; The weighting coefficients for the KL divergence regularization term; The weighting coefficient for the physical consistency check item; This represents the number of samples in the training batch. This represents the first three thermal parameters output by the forward mapping model; J The number of layers in the building envelope; The cross-entropy loss function; For the first i The first sample j The actual material category label of the layer; The first predicted by the decoder j Probability distribution of layer material categories; For the first i The first sample j The actual thickness value of the layer; The first predicted by the decoder j Layer thickness value.

[0009] Preferably, in step S3, the length of the allowed mask is equal to the number of selectable materials in the corresponding layer, the position corresponding to the material belonging to the allowed material set is set to 1, and the position corresponding to the material not belonging to the allowed material set is set to 0; in the dynamic mask, the position corresponding to the material that is incompatible with the material type determined in the previous layer is set to 0, and the other positions are set to 1.

[0010] Preferably, the training of the forward mapping model in step S2 is divided into a pre-training stage and a fine-tuning stage; in the pre-training stage, under the constraints of the construction rule set, a construction scheme is generated by Latin hypercube sampling, and the corresponding thermal parameters are calculated using the heat transfer analytical formula as training samples, with the loss function being the mean square error; in the fine-tuning stage, the forward mapping model is transferred to the data generated by the building energy consumption simulation software, and the loss function in the fine-tuning stage adds a physical consistency gradient penalty term to the mean square error.

[0011] Preferably, the forward mapping model automatically identifies the type of enclosure structure based on the material codes of each layer of the input enclosure structure, and applies the corresponding physical consistency gradient penalty term accordingly; The expression for the physical consistency gradient penalty term of the non-transparent enclosure structure is: ; The expression for the physical consistency gradient penalty term of the light-transmitting enclosure structure is: ; In the formula, For the physical consistency gradient penalty term of the non-transparent enclosure structure; The physical consistency gradient penalty term for the light-transmitting enclosure structure; λ is the penalty coefficient; j For building envelope hierarchy index; n The number of layers in a non-transparent building envelope; The number of glass substrate layers in a light-transmitting enclosure structure; The number of gas interlayers in a light-transmitting enclosure structure; This refers to the thickness of the insulation layer; The heat transfer coefficient K value is output by the forward mapping model; The solar heat gain coefficient (SHGC) value is the output of the forward mapping model. This represents the visible light transmittance output by the forward mapping model. For the first j Glass substrate thickness; For the first j Interlayer thickness of gas layers; The thermal inertia index output by the forward mapping model; For the first j Layer material density; For the first j Layer material thickness.

[0012] Preferably, the forward mapping model includes an input layer, four hidden layers, and an output layer. The input layer selects corresponding node configurations according to the type of enclosure structure. For non-transparent enclosure structures, twelve nodes are used, of which four nodes receive integer material codes. These integer material codes are mapped into four-dimensional continuous vectors through a shared entity embedding layer and then concatenated with four minimum-maximum normalized thickness values ​​and four minimum-maximum normalized material density values ​​before being input to the first hidden layer. For translucent enclosure structures, nine nodes are used for double-layer glass and fifteen nodes are used for triple-layer glass, with the node allocation method being the same as for non-translucent enclosure structures. The material density values ​​at each level are obtained from the enclosure structure material database based on the material codes and used as continuous feature inputs. The number of neurons in the four hidden layers are 128, 64, 32, and 16, respectively. Each hidden layer is followed by a batch normalization layer and a random dropout layer, with the activation function being a modified linear unit. The output layer has 5 nodes.

[0013] Preferably, in the fine-tuning stage, the parameters of the first two hidden layers closest to the input layer in the forward mapping model are frozen, and only the parameters of the last two hidden layers closest to the output layer and the output layer are updated.

[0014] Preferably, step S2 includes establishing a mapping relationship between thermal parameters and building performance indicators, taking the building's annual operating energy consumption and indoor environmental comfort indicators as optimization objectives, using a multi-objective evolutionary algorithm to search for the Pareto front non-dominated solution set in the thermal parameter space, and outputting the target thermal parameters after multi-attribute decision sorting.

[0015] Preferably, the multi-objective evolutionary algorithm employs the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy, and the multi-attribute decision-making sorting uses the superior-inferiority distance method TOPSIS; the indoor environmental comfort index includes thermal comfort satisfaction rate, humidity comfort satisfaction rate, and light comfort satisfaction rate; the multi-objective evolutionary algorithm optimizes by minimizing the entire objective function, and the expression of the objective function is: ; ; ; ; In the formula, The decision vector is composed of the thermal parameters of the building envelope; Annual operating energy consumption; , , The first The thermal comfort indicator function, humidity comfort indicator function, and light comfort indicator function for each hour are set to 1 if the comfort conditions are met, and 0 otherwise.

[0016] The advantages of this invention include at least the following: 1. By training a deep neural network as a forward mapping model, taking the thickness and material coding of each layer of the enclosure structure as input and thermal parameters as output, after pre-training and physical consistency fine-tuning, it can achieve millisecond-level accurate solution of all thermal parameters of any construction scheme, replacing the traditional repeated manual calculation and software simulation, and significantly improving the efficiency of forward solution. 2. A reverse design model is constructed by using a conditional variational autoencoder. The target value of the optimal thermal parameters is used as the conditional input. A construction rule mask layer is set at the output of the decoder. The material allowable mask vector and the compatibility mask vector are multiplied element by element with the material prediction branch logic value to filter out illegal material combinations. Compliant alternative solutions are automatically generated and verified by the forward mapping model in closed loop, which fundamentally replaces the manual trial-and-error reverse derivation process. 3. Through the multi-objective optimization module, within the thermal parameter space covered by the forward mapping model, the annual operating energy consumption of the building is simultaneously weighed against the four objectives of thermal comfort, humidity comfort, and light comfort. After searching the Pareto front, the target value of the thermal parameter with the best comprehensive performance is output through multi-attribute decision ranking. Then, the conditional variational autoencoder is transformed into a specific construction scheme, which solves the problem of difficulty in simultaneously weighing multiple conflicting objectives. 4. By adding a physical consistency gradient penalty term to the loss function in the fine-tuning stage, constraints are imposed on the partial derivatives of the heat transfer coefficient, solar heat gain coefficient with the thickness of the insulation layer, visible light transmittance with the thickness of the glass layer, and thermal inertia index with the material density and thickness, respectively. This ensures that the prediction results of the forward mapping model fully conform to the basic physical laws of heat transfer and optics, thereby improving the physical reliability of the model. 5. By introducing a layer-by-layer autoregressive decoding mechanism into the conditional variational autoencoder decoder, the decoder is sequentially decoded from the outside to the inside according to the physical hierarchy of the building envelope. This fully utilizes the causal dependencies between the building layers. At the same time, combined with a dynamic compatibility mask, illegal materials in the current layer are filtered in real time based on the determined materials of the previous layer, which greatly improves the compliance rate of the generated scheme and eliminates the need for subsequent manual correction. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the forward mapping model structure according to an embodiment of the present invention; Figure 3 This is a diagram of the conditional variational autoencoder architecture according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0019] like Figure 1 As shown, this embodiment of the invention provides a parametric design method for building envelope based on forward mapping and reverse generation, including the following steps.

[0020] Step S1: Construct the construction rule set of the enclosure structure. The construction rule set includes the allowable material set for each level of the enclosure structure and the material compatibility matrix of adjacent levels.

[0021] The purpose of this step is to lay the data foundation and rule constraints for subsequent forward mapping modeling and reverse design generation.

[0022] In this embodiment, the materials database is built using the MySQL 8.0 relational database management system and deployed on a rack server equipped with an Intel Xeon Silver 4214 processor, 64GB DDR4 ECC memory, and a 4TB NVMe solid-state drive. The operating system is Ubuntu 22.04 LTS. The materials database table is named envelope_material and contains the following fields: material_id is an INT type primary key with an auto-increment constraint set; material_name is of type VARCHAR and stores the Chinese name of the material. thermal_conductivity is of type DOUBLE and stores the thermal conductivity in units of W / (m·K); density is of type DOUBLE, representing storage density in kg / m³. `specific_heat` is of type DOUBLE and stores the specific heat capacity in J / (kg·K). solar_absorptance is a DOUBLE type that stores the solar radiation absorption coefficient, which is dimensionless. visible_transmittance is of type DOUBLE and stores the visible light transmittance, which is dimensionless. longwave_emissivity is of type DOUBLE and stores the longwave emissivity, which is dimensionless. min_thickness is of type DOUBLE and stores the minimum selectable thickness in mm; `max_thickness` is of type `DOUBLE` and stores the maximum selectable thickness in mm. thickness_step is of type DOUBLE and stores the thickness change step size in mm.

[0023] The database contains 35 pre-set materials, including cement mortar, paint layer, reinforced concrete, aerated concrete blocks, coral sand concrete, expanded polystyrene board, extruded polystyrene board, rock wool, gypsum board, ordinary transparent float glass, three different types of single-silver Low-E glass, two types of double-silver Low-E glass, air gap, and argon gap. Each material is assigned a unique material_id from 1 to 35.

[0024] The rule set is constructed and stored as a JSON file on the server's file system, named `construction_rules.json`. This file defines two top-level keys: `opaque_envelope` and `transparent_envelope`. The value corresponding to `opaque_envelope` contains a `layers` array, whose elements are, in order, four objects: `outer_finish`, `insulation`, `structural_core`, and `inner_finish`. Each object contains an `allowed_materials` array and a `thickness_range` object. The value corresponding to `transparent_envelope` contains two sub-objects: `double_glazing` and `triple_glazing`, defining the layer order of double-layer and triple-layer glass, respectively. Compatibility constraints are stored as a two-dimensional adjacency matrix under the `compatibility_matrix` key. The matrix has dimensions 35 x 35, where a value of 1 indicates allowed adjacency and 0 indicates prohibited adjacency. For example, the permissible material set for the exterior finish of a non-transparent building envelope includes cement mortar and paint; the permissible material set for the insulation layer includes expanded polystyrene board, extruded polystyrene board, and rock wool; the permissible material set for the structural layer includes reinforced concrete, aerated concrete blocks, and coral sand concrete; and the permissible material set for the interior finish layer includes cement mortar and gypsum board. In the compatibility matrix, paint is marked as permissible adjacent to expanded polystyrene board, extruded polystyrene board, and rock wool; cement mortar is marked as permissible adjacent to aerated concrete blocks; and rock wool is marked as prohibited from direct adjacent to reinforced concrete. Those skilled in the art can determine the complete compatibility matrix based on current common building construction practices. The construction rule set also defines the layer order constraints of the building envelope from the outside to the inside and the thickness range constraints for each layer.

[0025] This embodiment also constructs a performance evaluation index set, which includes five indices: heat transfer coefficient K, solar heat gain coefficient SHGC, visible light transmittance VLT, shading coefficient SC, and thermal inertia index D. The material physical parameters in the aforementioned material database, together with the hierarchical order and thickness range in the construction rule set, constitute the material construction combination space of the enclosure structure. The performance evaluation index set provides a unified quantitative dimension of thermal performance for each construction scheme in this material construction combination space.

[0026] After the JSON file is read by the Python program, a mask vectorization conversion operation is performed. For the four layers of the non-transparent enclosure structure, a 35-length floating-point one-dimensional mask vector is created for each layer. The material number corresponding to the allowed material set of that layer is assigned a value of 1.0, and the remaining positions are assigned a value of -10,000. The compatibility constraint matrix is ​​directly loaded as a 35x35 two-dimensional mask tensor. These mask vectors and tensors are registered as non-trainable buffers in PyTorch during model initialization and are called with each batch of data in the decoder mask layer of step S4, participating in the masking operation as part of the forward propagation computation graph.

[0027] Step S2: Using the thickness and material codes of each layer of the building envelope as input and thermal parameters as output, train a neural network as a forward mapping model; determine the target thermal parameters according to the building performance requirements.

[0028] The purpose of this step is to establish a surrogate model that can accurately map thermal parameters based on any construction scheme conforming to the set of construction rules, replacing the traditional time-consuming physical simulation calculations. Since walls have a certain heat storage function, there is a quantitative deviation between the simulation results directly modeled using the heat transfer coefficient K value and the results of the integrated material-to-structure simulation. Taking the fifteen-story tropical office building in this embodiment as an example, the annual energy consumption obtained from the integrated material-to-structure simulation is 7191.55 GJ, while the annual energy consumption obtained directly from the K value simulation is 6940.66 GJ, a difference of approximately 250.89 GJ. This embodiment resolves this deviation through a two-stage training strategy for the forward mapping model. In the pre-training stage, analytical heat transfer formulas are used to establish a basic mapping from material construction parameters to steady-state thermal parameters; in the fine-tuning stage, the model is transferred using the year-round dynamic simulation data from building energy consumption simulation software, enabling the model output to automatically absorb the systematic deviation caused by neglecting the wall heat storage effect in the steady-state assumption, transforming it into calibrated equivalent thermal parameters that match the year-round dynamic energy consumption performance. Taking the heat transfer coefficient as an example, for the same structural design, there is a nonlinear deviation between the steady-state heat transfer coefficient K calculated by the analytical formula and the calibrated equivalent heat transfer coefficient output by the fine-tuned model, which depends on the specific structural design. This deviation is automatically fitted by supervised learning during the fine-tuning process. The calibrated equivalent thermal parameters are used in conjunction with the building performance surrogate model trained on building energy consumption simulation software data in step S3. The two models are aligned with the same simulation benchmark, thereby eliminating the energy consumption prediction deviation caused by parameter simplification in the end-to-end design process.

[0029] In this embodiment, the pre-training dataset for the forward mapping model was generated on a deep learning workstation configured with an AMD EPYC 7543 32-core processor, 256GB of memory, and an NVIDIA RTX A6000 48GB GPU. A Python 3.9 script was written to connect to a MySQL materials database via the SQLAlchemy library and to parse the construction rule JSON file using the json library. For non-transparent enclosure structures, the script performed Latin hypercube sampling within the allowable material sets and thickness ranges of the four layers. For double- and triple-glass configurations of transparent enclosure structures, K values ​​were generated according to the EN673 standard calculation method, and SHGC and VLT were calculated using spectral data from the International Glass Database. The sampling operation was performed 100,000 times, generating a CSV file containing 100,000 samples, named pretrain_dataset.csv, which was randomly split into training and validation sets at an 8:2 ratio.

[0030] The forward mapping model is built using the PyTorch 2.0 framework as the ForwardMappingNet class. This model is a deep neural network structure, containing one input layer, four hidden layers, and one output layer.

[0031] The input layer dynamically adapts the number of nodes based on the enclosure structure type: The non-transparent enclosure structure uses twelve nodes, of which four nodes receive integer material codes, four nodes receive thickness values ​​normalized to the 0-1 range after minimum and maximum normalization, and four nodes receive material density values ​​normalized to the 0-1 range after minimum and maximum normalization. The double-glazed structure in the light-transmitting enclosure uses nine nodes, of which three nodes receive integer material codes, three nodes receive thickness values ​​normalized to the 0-1 range after minimum and maximum normalization, and three nodes receive material density values ​​normalized to the 0-1 range after minimum and maximum normalization. The three-layer glass in the light-transmitting enclosure structure uses fifteen nodes, of which five nodes receive integer material codes, five nodes receive thickness values ​​normalized to the 0-1 range after minimum and maximum normalization, and five nodes receive material density values ​​normalized to the 0-1 range after minimum and maximum normalization.

[0032] The material density values ​​of each layer are obtained from the material database based on the material code. After min-max normalization, they are used as continuous features input into the network, so that the material density can be used as a differentiable leaf node in the computation graph to participate in gradient calculation.

[0033] Material encoding is mapped to a 4-dimensional continuous vector via a shared entity embedding layer. This embedding layer is an `nn.Embedding` instance with a vocabulary size of 36 and an embedding dimension of 4. Index 0 is reserved as a padding marker, corresponding to encodings of 1 to 35 for 35 actual materials. The 16-dimensional material feature vector embedded in the non-transparent enclosure structure is concatenated with 4 thickness values ​​and 4 density values ​​to form a 24-dimensional vector, which is then input to the first hidden layer. The 12-dimensional material feature vector embedded in the double-layered glass enclosure structure is concatenated with 3 thickness values ​​and 3 density values ​​to form an 18-dimensional vector, which is then input to the first hidden layer. The 20-dimensional material feature vector embedded in the triple-layered glass enclosure structure is concatenated with 5 thickness values ​​and 5 density values ​​to form a 30-dimensional vector, which is then input to the first hidden layer. The first hidden layer automatically adapts to the input vectors of different dimensions. The four hidden layers are, in order, `nn.Linear` layers with 128, 64, 32, and 16 neurons respectively. Each fully connected layer is followed by an nn.BatchNorm1d batch normalization layer, an nn.Dropout layer, and an nn.ReLU activation function, with the dropout probability set to 0.1. The output layer is an nn.Linear layer with 5 neurons, corresponding to the heat transfer coefficient K, solar heat gain coefficient SHGC, visible light transmittance VLT, shading coefficient SC, and thermal inertia index D. Figure 2 As shown, the overall structure of the model consists of an input layer that receives material encoding, thickness and density information, four hidden layers that gradually extract features, and finally an output layer that outputs five thermal parameters.

[0034] The training of the forward mapping model is divided into a pre-training phase and a fine-tuning phase.

[0035] During the pre-training phase, the `torch.optim.Adam` optimizer was used, with an initial learning rate of 0.001, a weight decay factor of 0.0001, a batch size of 512, and a training epoch of 200. The learning rate scheduler used was `torch.optim.lr_scheduler.StepLR`, with a step size of 50 epochs and a decay factor of 0.5. The loss function was `nn.MSELoss` (mean squared error). During pre-training, construction schemes were generated through Latin hypercube sampling under the constraints of the construction rule set, and the corresponding thermal parameters were calculated using analytical heat transfer formulas as training samples. After pre-training, the model parameter dictionary was saved to the `forward_pretrained.pth` file.

[0036] The heat transfer analytical formula used in this embodiment includes the wall heat transfer coefficient. The calculation formula is shown below.

[0037] ; In the formula, The internal surface heat transfer resistance is taken as 0.13 m²·K / W; The external surface heat transfer resistance is taken as 0.04 m²·K / W; For the first Layer material thickness; For the first Thermal conductivity of the layer material; This represents the total number of floors. From 1 to Values.

[0038] The solar heat gain coefficient (SHGC) is calculated using two formulas: one for single-pane glass and one for multi-pane glass. The formula for single-pane glass is as follows: ; In the formula, The total solar transmittance of a single-layer glass pane; The solar absorption coefficient of a single-layer glass; The solar heat gain coefficient of double- and triple-glazed systems is calculated according to EN673 standard, using the following formula: ; In the formula, The total solar transmittance of the glass system; Let be the solar absorption coefficient of the j-th glass layer; The internal flow fraction of heat absorbed by the j-th layer of glass and transferred into the room; n This represents the number of glass layers.

[0039] The total solar transmittance of the glass system is calculated by superimposing multiple reflections and transmissions based on the optical parameters of each glass layer; the internal flow fraction is calculated based on the thermal resistance of each glass layer and the thermal resistance of the interlayer gas.

[0040] Thermal inertia index D The calculation formula is shown below: ; ; ; In the formula, For the first j Thermal resistance of the layer material; For the first j The heat storage coefficient of the layer material; T The temperature fluctuation period is set to 86400 seconds. For the first j Specific heat capacity of the layer material, expressed in J / (kg·K); For the first j Density of layer material.

[0041] During the fine-tuning phase, EnergyPlus 23.1 software was used to generate the fine-tuning dataset. A parametric baseline model of a fifteen-story tropical office building was built on a workstation configured with Grasshopper for Rhino 7 and the Ladybug Tools 1.5 plugin. The building faces due south, with a standard floor plan that is rectangular, 45m long, 18m wide, and 3.6m high, with a total building area of ​​12,150 square meters. Three types of variables were selected: external wall insulation thickness, roof insulation thickness, and window type encoding. One thousand combinations were generated using Sobol sequence sampling. For each combination, an IDF input file was generated using Ladybug Tools. An 8760-hour annual energy consumption simulation was then performed using the EnergyPlus command-line interface, and the steady-state thermal parameters corresponding to the building envelope were extracted from the output SQLite database.

[0042] The program fine-tunes the loading of pre-trained model weights, sets the requires_grad attribute of the first two hidden layers closest to the input layer to False to freeze parameters, and only updates the parameters of the last two hidden layers closest to the output layer and the output layer.

[0043] In the fine-tuning phase, the loss function adds a physical consistency gradient penalty term to the mean squared error. This physical consistency gradient penalty term is defined separately for each type of enclosure structure. The forward mapping model automatically identifies the enclosure structure type based on the input enclosure structure layer material code and applies the corresponding penalty term for constraint. The expression for the physical consistency gradient penalty term for non-transparent enclosure structures is: ; The expression for the physical consistency gradient penalty term of the light-transmitting enclosure structure is: ; In the formula, λ is the penalty coefficient, which is set to 0.01 in this embodiment; j For building envelope hierarchy index; n The number of layers in a non-transparent building envelope; The number of glass substrate layers in a light-transmitting enclosure structure; The number of gas interlayers in a light-transmitting enclosure structure; This refers to the thickness of the insulation layer; The heat transfer coefficient K is the output of the deep neural network; The solar heat gain coefficient (SHGC) value is the output of the deep neural network. Visible light transmittance output by a deep neural network; For the first j Glass substrate thickness; For the first j Interlayer thickness of gas layers; The thermal inertia index output by a deep neural network; For the first j Layer material density; For the first j Layer material thickness.

[0044] When the heat transfer coefficient K increases with the thickness of the insulation layer, the penalty term for the opaque building envelope is activated. According to the principles of heat transfer, an increase in insulation layer thickness should decrease the value of K. When the thermal inertia index decreases with the increase of material density or thickness, the penalty term for the opaque building envelope is activated. According to the principles of heat storage, an increase in material density or thickness should increase the thermal inertia index. The above-mentioned physical consistency gradient penalty term for the opaque building envelope only applies a gradient constraint on the heat transfer coefficient for the insulation layer thickness. This is because the thickness variation range of the exterior finish layer, structural layer, and interior finish layer is usually narrow due to building construction requirements, and their contribution to the heat transfer coefficient is much smaller than that of the insulation layer. Adding additional gradient constraint terms with a limited sample size for fine-tuning can easily lead to gradient conflicts. Therefore, applying constraints only to the insulation layer, which has the most significant impact, is sufficient to meet the physical consistency requirements.

[0045] The light-transmitting enclosure penalty term is activated when the heat transfer coefficient K or solar heat gain coefficient SHGC increases with increasing glass substrate thickness. According to heat transfer and optics principles, increasing glass substrate thickness should decrease both K and SHGC values. The light-transmitting enclosure penalty term is also activated when the visible light transmittance increases with increasing glass substrate thickness. According to optics principles, increasing glass substrate thickness should decrease visible light transmittance. Furthermore, the light-transmitting enclosure penalty term is activated when the visible light transmittance changes with increasing interlayer thickness. According to optics principles, changes in interlayer thickness should not alter visible light transmittance. All partial derivatives are analytically calculated after each forward propagation using PyTorch's `torch.autograd.grad` function. The gradient penalty term for physical consistency of the light-transmitting enclosure structure does not include the gradient constraint of the heat transfer coefficient K on the thickness of the gas interlayer. The reason is that the heat transfer mechanism of the gas interlayer involves the coupling of conduction, convection and radiation. When the thickness of the gas interlayer exceeds the critical value, the enhanced convection will lead to a reversal in the heat transfer coefficient. There is no global monotonic relationship between K and the thickness of the gas interlayer, so a unidirectional gradient penalty cannot be applied.

[0046] Fine-tuning was performed using the Adam optimizer with a learning rate of 0.0001, a batch size of 100, and a training epoch of 50. After fine-tuning, the model achieved a determination coefficient R² of 0.998 on the validation set.

[0047] In this embodiment, a building performance proxy model is first trained. This proxy model includes one energy consumption proxy model and three comfort proxy models, all trained on the same NVIDIA RTX A6000 GPU workstation. The training dataset for the energy consumption proxy model is generated using EnergyPlus 23.1. The parameter space is defined to include twelve dimensions: exterior wall heat transfer coefficient, roof heat transfer coefficient, exterior window heat transfer coefficient, exterior window solar heat gain coefficient, exterior window visible light transmittance, shading coefficient of shading components, north-facing window-to-wall ratio, south-facing window-to-wall ratio, east-facing window-to-wall ratio, west-facing window-to-wall ratio, building shape coefficient, and number of building floors. Five thousand sample points are generated using Latin hypercube sampling within the value range of each dimension. Each sample point corresponds to a building IDF model file. The multiprocessing module of Python is used to call EnergyPlus in parallel to perform a year-round simulation.

[0048] Annual operating energy consumption is calculated as the sum of the annual electricity consumption of air conditioning, lighting, and equipment systems, normalized to the building area, and expressed in kWh / m². Thermal comfort hours are defined as the cumulative number of hours when the indoor operating temperature falls within the tropical climate adaptability comfort range defined by ASHRAE Standard 55. Humidity comfort hours are defined as the cumulative number of hours when the indoor relative humidity is between 30% and 70%. Light comfort hours are defined as the cumulative number of hours when the natural illuminance on the work surface is between 300 and 3000 lux and the glare index does not exceed 22.

[0049] The energy consumption proxy model is defined as the EnergyProxyNet class, employing a fully connected feedforward network structure. The input layer has 12 nodes, hidden layers have 256, 128, and 64 neurons respectively, with the activation function being nn.LeakyReLU, and the output layer has 1 node. Training uses the mean squared error loss function, with an initial learning rate of 0.001 using the Adam optimizer, a batch size of 1024, and 500 training epochs. After training, the MAPE (Mean Average Performance) was validated at 2.3% on the 15% reserved test set. The three comfort proxy models have the same network structure as the energy consumption proxy model. All four models were exported in TorchScript format for use by the optimization algorithm.

[0050] In this embodiment, the multi-objective evolutionary algorithm employs the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy, implemented using the Geatpy 2.7 evolutionary algorithm library. The objective function is defined as shown in the following four equations.

[0051] ; ; ; ; In the formula, The decision vector is composed of the thermal parameters of the building envelope; Annual operating energy consumption; , , The first The thermal comfort indicator function, humidity comfort indicator function, and light comfort indicator function are set to 1 if the comfort conditions are met, and 0 otherwise. Values ​​range from 1 to 8760. Each comfort objective is negative to transform it into a minimization problem.

[0052] The decision variables are 10-dimensional vectors, including exterior wall K, roof K, exterior window K, SHGC, VLT, SC, and the window-to-wall ratio for each of the four orientations. The population size is set to 200, the maximum number of generations to 300, the crossover probability to 0.9, and the mutation probability to 0.1. Simulated binary crossover is used as the crossover operator, and the distribution index... Set to 20. The mutation operator uses multinomial mutation, with a distribution exponent of 20. Set it to 20. The selection operator uses tournament selection, with a tournament size of 2.

[0053] In each generation of fitness assessment, each individual is decoded into a decision vector. The thermal parameter components in this decision vector are directly input into the energy consumption proxy model and the comfort proxy model to calculate four performance indicators. The role of the forward mapping model in the optimization process is to define the feasible range of values ​​for the decision variables: before optimization begins, forward mapping is performed on all legal construction schemes in the material database in batches, and the output value range of each thermal parameter is statistically analyzed. The upper and lower bounds of this value range are used as constraints on the value range of each decision variable in NSGA-II. Non-dominated ranking divides all individuals in the population into frontier levels according to Pareto dominance relationships. Individuals with a non-dominated level of 1 constitute the first frontier, individuals with a non-dominated level of 2 constitute the second frontier, and so on. Within the same frontier, individuals are ranked using crowding distance to maintain population diversity. The formula for calculating crowding distance is as follows: ; In the formula, For the first The crowding distance of each individual; , According to the number After sorting the target, the first The and the first The individual in the first Function values ​​on each target; and These represent the maximum and minimum values ​​of the objective at the current frontier, respectively. The above equation sums the four objective functions; a larger crowding distance indicates a sparser solution space around the individual, and thus a higher priority for preservation in environmental selection.

[0054] After 300 iterations, all individuals with a non-dominant level of 1 in the final generation are extracted to form the Pareto front solution set.

[0055] The multi-attribute decision ranking adopts the Top-Side Distance Method (TOPSIS). In this embodiment, the weights of the four objectives are set to be equal, i.e. Designers can adjust the weights of each objective based on the specific project's emphasis on energy saving or comfort. The TOPSIS decision module first performs a reciprocal positive transformation on the annual operating energy consumption column, converting all objectives into positive indicators where larger values ​​are better. Then, it performs range standardization, as shown in the following formula: ; In the formula, For the first The first scheme is in the Standardized values ​​for each target; For the first i The first scheme is in the k The function values ​​after positive transformation for each objective. When the maximum and minimum values ​​of an objective in the Pareto front solution set are equal, the standardized value of that objective column is uniformly set to 0.5.

[0056] After standardization, the positive ideal solution Z+ is defined as the vector composed of the maximum values ​​of each column, i.e., Z+ is equal to a vector of all 1s, and the negative ideal solution Z- is defined as the vector composed of the minimum values ​​of each column, i.e., Z- is equal to a vector of all 0s. The Euclidean distances between each solution and the positive ideal solution Z+ and the negative ideal solution Z- are calculated. The formula for calculating the relative closeness is as follows: ; In the formula, For the first i The relative similarity of the two solutions; For the first i The weighted Euclidean distance between each scheme and the negative ideal solution Z-; For the first i The weighted Euclidean distance between the proposed solution and the ideal solution Z+.

[0057] Based on relative proximity Sort the solutions from largest to smallest, select the top 5 as the candidate optimal solution set, and output the K objective value, SHGC objective value and VLT objective value for use in step S3. The closer the value is to 1, the closer the solution is to the positive ideal solution and the farther it is from the negative ideal solution, indicating better overall performance.

[0058] After TOPSIS sorts and outputs the candidate optimal solution set, the decision variables in each candidate solution are divided into three categories according to their parameter properties. The first category is the construction parameters of the light-transmitting building envelope, including the heat transfer coefficient K of the exterior windows, the solar heat gain coefficient SHGC, and the visible light transmittance VLT. These parameters are used for the reverse screening of the light-transmitting building envelope. The second category is the building design parameters, including the roof heat transfer coefficient, the shading coefficient SC, and the window-to-wall ratio for each orientation. These parameters correspond to the overall building design decision and are directly adopted by the designers based on the TOPSIS sorting results without the need for reverse identification network processing. The third category is the construction parameters of the non-light-transmitting building envelope, which only includes the target value of the exterior wall heat transfer coefficient K. This parameter serves as the conditional input of the conditional variational autoencoder and is used to reversely generate wall construction schemes.

[0059] Step S3: For non-transparent enclosure structures, the target thermal parameters are used as conditions and input into the trained conditional generation model. The decoding end of the conditional generation model performs autoregressive decoding layer by layer from the outside to the inside of the enclosure structure. The decoding output of each layer generates an allowable mask based on the allowable material set and generates a dynamic mask based on the material compatibility matrix determined in the previous layer. The allowable mask and the dynamic mask are applied to the original logical value of the current layer's material category prediction branch output, so that the output probability of materials that do not belong to the allowable material set or are incompatible with the materials in the previous layer is zero, thus obtaining the candidate construction scheme. For transparent enclosure structures, under the constraints of the construction rule set, the legal material and thickness combinations are traversed, and the combination with the smallest deviation from the target thermal parameters is selected as the candidate construction scheme.

[0060] In this embodiment of the invention, a conditional variational autoencoder is used as the conditional generation model. For example... Figure 3 As shown, the overall architecture of the conditional variational autoencoder includes conditional input, encoder, latent space variables, decoder, construction rule mask layer, and alternative construction scheme set.

[0061] In this embodiment, the conditional variational autoencoder is defined as the InverseGenerationVAE class in PyTorch. The encoder subnetwork consists of two fully connected layers and a latent spatial parameter output layer. The input vector is composed of the target thermal parameter conditional vector. and the vector of the actual construction scheme It is assembled from multiple parts. In this embodiment, the conditional variational autoencoder is used for the inverse generation of the opaque enclosure structure. Conditional vector The dimension is 3. For non-transparent enclosure structures, its components are set as the target value, zero value, and zero value of the external wall K, i.e. The latter two components are set to zero because the solar heat gain coefficient and visible light transmittance of a non-transparent wall are essentially close to zero. Construction scheme vector s After entity embedding encoding and cThe concatenation process results in a total input dimension of 15. The material embedding dimension in the CVAE encoder is 2, while the material embedding dimension in the forward mapping model is 4. These different embedding dimensions are used to suit their respective model training requirements. The first fully connected layer has an input dimension of 15 and an output dimension of 256, followed by a LeakyReLU activation function and a LayerNorm layer for normalization. The second fully connected layer has an input dimension of 256 and an output dimension of 128, followed by a LeakyReLU activation function and a LayerNorm layer for normalization. The latent space parameter output layers are two parallel fully connected layers, outputting the mean vector and the log-variance vector, respectively, both with a dimension of 32. The reparameterization operation samples noise from a standard normal distribution during training and scales it using the mean and standard deviation to obtain the latent variables. The mean vector is used directly during inference.

[0062] The decoder subnetwork adopts a layer-by-layer autoregressive decoding structure, with latent variables as input. and target thermal parameter condition vector The layers are concatenated, with a dimension of 35. The first shared fully connected layer has an input dimension of 35 and an output dimension of 128, followed by a LeakyReLU activation function and a LayerNorm normalization layer. The second shared fully connected layer has an input dimension of 128 and an output dimension of 256, followed by a LeakyReLU activation function and a LayerNorm normalization layer. Independent decoding units are set up for the four layers of the building envelope from the outside to the inside: exterior finish layer, insulation layer, structural layer, and interior finish layer. Each decoding unit contains a material category prediction linear layer and a thickness regression linear layer. The first exterior finish layer decoding unit directly receives the output of the second shared fully connected layer. The second insulation layer decoding unit receives the concatenated vector of the output of the second shared fully connected layer and the decoding result of the first layer as input. The third structural layer decoding unit receives the concatenated vector of the output of the second shared fully connected layer and the decoding results of the first two layers as input. The fourth interior finish layer decoding unit receives the concatenated vector of the output of the second shared fully connected layer and the decoding results of the first three layers as input. Through this cascading conditional transmission mechanism, the decoding of subsequent layers can make full use of the material and thickness information already determined in the preceding layers, thereby generating a construction scheme that is physically compatible between layers.

[0063] Set a construction constraint mask layer at the output of each material category. For each layer of the enclosure structure, predefine a material allowance mask vector. The length of this vector is equal to the number of selectable materials at the corresponding layer. Allowed materials are assigned a 1 at their corresponding positions, while prohibited materials are assigned a 0. During layer-by-layer decoding, a predefined 35x35 material compatibility matrix is ​​dynamically queried based on the material types determined in the previous layer to generate the compatibility mask vector for the current layer. In this case, the positions corresponding to materials incompatible with those already determined in the previous layer are set to 0, and the remaining positions are set to 1. The material allowable mask vector is then used. Compatibility mask vector The original logical value vector output by the current layer material category prediction branch The element-wise multiplication operation is performed as shown in the following expression: ; In the formula, This represents an element-wise multiplication operation. After the multiplication operation, the logical value corresponding to an invalid material is set to zero. In the specific implementation, to ensure that the probability of an invalid material after the softmax function calculation is strictly zero rather than approaching a uniform distribution, the above mask vector... and At the logical level, it takes values ​​of 1 and 0, compared to the original logical value vector. Element-wise multiplication sets the logic value corresponding to illegal materials to zero. In engineering implementation, to avoid non-zero outputs from the softmax function for zero-value logic, this embodiment implements the product operation as an additive masking method: the mask value for allowed materials is set to 0.0, and the mask value for prohibited materials is set to negative infinity. The basic allowed material mask and the dynamic compatibility mask are added element-wise to obtain the final mask vector. Then, the original logic value vector and the final mask vector are added element-wise, and the result is fed into the softmax function to calculate the probability distribution of each material, thereby forcing the output probability of illegal materials to zero. The output value of the thickness regression branch is mapped to the actual thickness range of each layer through inverse normalization, that is, the standardized value of the network output in the 0-1 interval is linearly mapped according to the minimum and maximum thickness of the material in that layer, restoring it to the actual physical thickness.

[0064] During the training phase, the decoder employs a teacher-forced strategy: during layer-by-layer decoding, the conditional input of each decoding unit uses the true material category and thickness values ​​from the training samples for that layer, rather than the prediction results from the previous layer, thus ensuring stable backpropagation of gradients during training. The loss for the material category prediction branch is calculated using the cross-entropy loss function, and the loss for the thickness regression branch is calculated using the mean squared error loss function; the sum of these two functions constitutes the reconstruction error term. During the inference phase, the conditional input of each decoding unit uses the actual prediction results from the previous layer after being filtered by the mask layer.

[0065] The training loss function consists of three terms, and its mathematical expression is shown below: ; ; In the formula, For the first i Reconstruction error of each sample; For the firsti The true construction scheme vector of each sample; The first generation generated for decoder reconstruction i A vector of construction schemes for each sample; It is a conditional vector; It is a forward mapping model; Let KL divergence be a metric. The posterior distribution for encoder parameterization; It follows a standard normal prior distribution; The weighting coefficients for the KL divergence regularization term; The weighting coefficient for the physical consistency check item; This represents the number of samples in the training batch. This represents the first three thermal parameters output by the forward mapping model; J The number of layers in the building envelope; The cross-entropy loss function; For the first i The first sample j The actual material category label of the layer; The first predicted by the decoder j Probability distribution of layer material categories; For the first i The first sample j The actual thickness value of the layer; The first predicted by the decoder j Layer thickness value.

[0066] The first term is the reconstruction error term, which is composed of the cross-entropy loss of the material category and the mean square error loss of the thickness, measuring the distance between the generated scheme and the real scheme; the second term is the KL divergence regularization term, which constrains the posterior distribution of the latent space to be close to the standard normal prior distribution, with a weight coefficient set to 0.5; the third term is the physical consistency verification term, which inputs the constructed scheme vector generated by the decoder into the forward mapping model to obtain the predicted value of the thermal parameters, extracts the first three terms of the predicted value and calculates the sum of squared deviations with the target thermal parameters in the condition vector and includes them in the loss calculation, with a weight coefficient set to 1.0; the training dataset is randomly selected from the pre-training dataset in step S2, with 20,000 samples.

[0067] For translucent enclosure structures, since their material combination space is much smaller than that of non-translucent enclosure structures, this embodiment uses an exhaustive screening method under construction rule constraints to replace the conditional variational autoencoder. Specifically, based on the double-glazed and triple-glazed hierarchical structures defined in the construction rule set, all legal combinations of glass types and gas interlayer types, as well as discrete thickness levels for each layer, are traversed. The heat transfer coefficient K, solar heat gain coefficient SHGC, and visible light transmittance VLT corresponding to each combination are calculated in batches using the EN673 standard analytical formula. From these, several combinations with the smallest deviations from the target values ​​of K, SHGC, and VLT output in step S3 are selected as candidate construction schemes for translucent enclosure structures. Since the material database in this embodiment contains seven glass types and two gas interlayer types, the number of effective combinations for double-glazed structures is in the hundreds, and for triple-glazed structures it is in the thousands. The analytical formula calculation speed is extremely fast, and the exhaustive screening can be completed in seconds. The selected candidate translucent enclosure structures are also input into the forward mapping model for forward verification.

[0068] Step S4: Input the alternative construction schemes into the forward mapping model for forward verification and then output the final recommended scheme.

[0069] The inference phase uses only the decoder and mask layer; the encoder does not participate in the inference process. The optimal thermal parameter target values ​​output from step S3 are constructed as a condition vector, and one hundred alternative construction schemes are generated by multiple random samplings of the latent variable z. Each alternative scheme is fed into the forward mapping model trained in step S2 for forward validation, calculating the absolute value of the relative deviation from the corresponding terms of the target thermal parameters determined in steps S2 and S3. Schemes with a relative deviation exceeding a 5% threshold for any indicator are discarded, and all validated schemes are retained as the final recommended construction scheme output. Each recommended scheme includes material numbers and thickness values ​​at each level, allowing designers to directly access the material database for complete physical parameter information.

[0070] At the engineering implementation level, this invention establishes a multi-parameter combined design model for the building envelope based on materials, structure, and architecture through secondary development in Revit. This model enables automated calculation of thermal parameters, carbon emissions, and costs of the building envelope, and allows for design scheme recommendations. The integrated model encapsulates the forward mapping model in step S2 and the conditional variational autoencoder in step S3 into a callable inference interface, embedded in the parametric design flow of the Revit plugin. After designers modify the building envelope construction parameters in the Revit environment, the plugin automatically calls the forward mapping model to calculate thermal parameters in real time, without needing to launch external simulation software. When designers input a target thermal performance range, the plugin automatically calls the conditional variational autoencoder to generate construction schemes that meet the target for selection. Through the above parameter simplification method, the original nearly 70 parameters are reduced to less than 20, significantly reducing model construction complexity and simulation time while maintaining simulation accuracy.

[0071] In summary, the parametric design method for building envelopes based on forward mapping and backward generation provided by this invention achieves millisecond-level accurate solution of all thermal parameters for any construction scheme through a forward mapping model. It automatically generates compliant alternative schemes with the optimal thermal parameter target value as a condition through a conditional variational autoencoder and verifies them through closed-loop. Through a multi-objective optimization module, it simultaneously balances energy consumption and multiple comfort objectives and outputs the thermal parameter target value with the best overall performance. This solves the problems of low efficiency in forward solution, reliance on manual trial and error in backward derivation, and difficulty in simultaneously balancing multiple conflicting objectives in existing building envelope design technologies.

[0072] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0073] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A parametric design method for building envelopes based on forward mapping and backward generation, characterized in that, Includes the following steps: Step S1: Construct a set of construction rules for the enclosure structure, which includes the set of permissible materials for each level of the enclosure structure and the material compatibility matrix of adjacent levels; Step S2: Using the thickness and material codes of each layer of the building envelope as input and thermal parameters as output, train a neural network as a forward mapping model; determine the target thermal parameters according to the building performance requirements. Step S3: For non-transparent enclosure structures, the target thermal parameters are used as conditions and input into the trained conditional generation model. The decoding end of the conditional generation model performs autoregressive decoding layer by layer from the outside to the inside of the enclosure structure. The decoding output of each layer generates an allowable mask based on the allowable material set, and generates a dynamic mask by querying the material compatibility matrix based on the material type determined in the previous layer. The allowable mask and the dynamic mask are applied to the original logical value of the current layer material category prediction branch output, so that the output probability of materials that do not belong to the allowable material set or are incompatible with the materials in the previous layer is zero, thus obtaining alternative construction schemes. For a light-transmitting enclosure structure, under the constraints of the construction rule set, legal material and thickness combinations are traversed, and the combination with the smallest deviation from the target thermal parameters is selected as the alternative construction scheme. The conditional generation model is a conditional variational autoencoder. During the training phase, the encoder concatenates the thermal parameter conditional vector with the true construction scheme vector and outputs the mean and log-variance of the latent space variables through a fully connected layer. The training loss function includes a reconstruction error term, a KL divergence regularization term, and a physical consistency check term. The physical consistency check term inputs the construction scheme vector generated by the decoder into the forward mapping model to obtain the predicted thermal parameter values, extracts the first three terms of the predicted thermal parameter values, calculates the sum of squared deviations between the first three terms and the thermal parameter conditional vector, and includes the sum of squared deviations in the loss calculation. The expression for the training loss function is: ； ； In the formula, For the first i Reconstruction error of each sample; For the first i The true construction scheme vector of each sample; The first generation generated for decoder reconstruction i A vector of construction schemes for each sample; It is a conditional vector; It is a forward mapping model; Let KL divergence be a metric. The posterior distribution for encoder parameterization; It follows a standard normal prior distribution; The weighting coefficients for the KL divergence regularization term; The weighting coefficient for the physical consistency check item; This represents the number of samples in the training batch. This represents the first three thermal parameters output by the forward mapping model; J The number of layers in the building envelope; The cross-entropy loss function; For the first i The first sample j The actual material category label of the layer; The first predicted by the decoder j Probability distribution of layer material categories; For the first i The first sample j The actual thickness value of the layer; The decoder predicts the first j Layer thickness value; Step S4: Input the alternative construction schemes into the forward mapping model for forward verification and then output the final recommended scheme.

2. The parametric design method for building envelope based on forward mapping and backward generation according to claim 1, characterized in that: In step S3, the length of the allowed mask is equal to the number of selectable materials in the corresponding layer. The corresponding position of the material belonging to the allowed material set is set to 1, and the corresponding position of the material not belonging to the allowed material set is set to 0. In the dynamic mask, the corresponding position of the material that is incompatible with the material type determined in the previous layer is set to 0, and the other positions are set to 1.

3. The parametric design method for building envelope based on forward mapping and backward generation according to claim 1, characterized in that: The training of the forward mapping model in step S2 is divided into a pre-training stage and a fine-tuning stage. In the pre-training stage, under the constraints of the construction rule set, a construction scheme is generated by Latin hypercube sampling, and the corresponding thermal parameters are calculated using the heat transfer analytical formula as training samples. The loss function is the mean square error. The fine-tuning stage utilizes data generated by building energy consumption simulation software to perform transfer learning on the forward mapping model. The loss function in the fine-tuning stage adds a physical consistency gradient penalty term to the mean square error.

4. The parametric design method for building envelope based on forward mapping and backward generation according to claim 3, characterized in that: The forward mapping model automatically identifies the type of enclosure structure based on the material codes of each layer of the input enclosure structure, and applies the corresponding physical consistency gradient penalty term accordingly. The expression for the physical consistency gradient penalty term of the non-transparent enclosure structure is: ； The expression for the physical consistency gradient penalty term of the light-transmitting enclosure structure is: ； In the formula, For the physical consistency gradient penalty term of the non-transparent enclosure structure; The physical consistency gradient penalty term for the light-transmitting enclosure structure; λ is the penalty coefficient; j For building envelope hierarchy index; n The number of layers in a non-transparent building envelope; The number of glass substrate layers in a light-transmitting enclosure structure; The number of gas interlayers in a light-transmitting enclosure structure; This refers to the thickness of the insulation layer; The heat transfer coefficient K value is output by the forward mapping model; The solar heat gain coefficient (SHGC) value is the output of the forward mapping model. This represents the visible light transmittance output by the forward mapping model. For the first j Glass substrate thickness; For the first j Interlayer thickness of gas layers; The thermal inertia index output by the forward mapping model; For the first j Layer material density; For the first j Layer material thickness.

5. The parametric design method for building envelope based on forward mapping and backward generation according to claim 3, characterized in that: The forward mapping model includes one input layer, four hidden layers, and one output layer. The input layer selects corresponding node configurations according to the type of enclosure structure. For non-transparent enclosure structures, twelve nodes are used, four of which receive integer material codes. These integer material codes are mapped into four-dimensional continuous vectors through a shared entity embedding layer and then concatenated with four minimum-maximum normalized thickness values ​​and four minimum-maximum normalized material density values ​​before being input to the first hidden layer. For translucent enclosure structures, nine nodes are used for double-layer glass and fifteen nodes are used for triple-layer glass, with the node allocation method being the same as for non-translucent enclosure structures. The material density values ​​at each level are obtained from the enclosure structure material database based on the material codes and used as continuous feature inputs. The number of neurons in the four hidden layers are 128, 64, 32, and 16, respectively. Each hidden layer is followed by a batch normalization layer and a random dropout layer, with the activation function being a modified linear unit. The output layer has 5 nodes.

6. The parametric design method for building envelope based on forward mapping and backward generation according to claim 5, characterized in that: The fine-tuning stage freezes the parameters of the first two hidden layers closest to the input layer in the forward mapping model, and only updates the parameters of the last two hidden layers closest to the output layer and the output layer.

7. The parametric design method for building envelope based on forward mapping and backward generation according to claim 1, characterized in that: Step S2 includes establishing a mapping relationship between thermal parameters and building performance indicators. Taking the building's annual operating energy consumption and indoor environmental comfort indicators as optimization objectives, a multi-objective evolutionary algorithm is used to search for the Pareto front non-dominated solution set in the thermal parameter space. After sorting by multi-attribute decision, the target thermal parameters are output.

8. The parametric design method for building envelope based on forward mapping and backward generation according to claim 7, characterized in that: The multi-objective evolutionary algorithm employs the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy, and the multi-attribute decision-making sorting uses the top-inferiority distance method (TOPSIS). The indoor environmental comfort indicators include thermal comfort satisfaction rate, humidity comfort satisfaction rate, and light comfort satisfaction rate. The multi-objective evolutionary algorithm optimizes by minimizing the entire objective function, the expression of which is: ； ； ； ； In the formula, The decision vector is composed of the thermal parameters of the building envelope; Annual operating energy consumption; , , The first The thermal comfort indicator function, humidity comfort indicator function, and light comfort indicator function for each hour are set to 1 if the comfort conditions are met, and 0 otherwise.

Images