A building performance prediction method and evaluation system based on a hybrid neural network

Abstract

The application relates to the technical field of intelligent buildings, and provides a building performance prediction method based on a hybrid neural network, which comprises the following steps: real-time grabbing of material price data; the hybrid neural network model comprises: a price feature conversion unit, which converts the material price data into price features; a first long short-term memory network, which extracts energy consumption time features from historical energy consumption data; a second long short-term memory network, which extracts climate time features from site climate data; a three-dimensional convolutional neural network, which extracts three-dimensional space features of a building from three-dimensional volume data; a splicing module, which obtains comprehensive feature representation; a fully connected module, which extracts nonlinear features from the comprehensive feature representation; and an output module, which carries out regression processing on the nonlinear features to obtain energy consumption, cost and full life cycle carbon emissions of the building. The application further discloses a building performance evaluation system based on the hybrid neural network. The application improves the precision and real adaptability of building multi-performance prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent building technology, and in particular to a method and evaluation system for predicting building performance based on hybrid neural networks. Background Technology

[0002] Buildings consume 40% of the world's energy, making effective control of building energy consumption and carbon emissions a key direction for the development of near-zero energy buildings. Meanwhile, accurate estimation of building costs is crucial for building operation. Therefore, precise prediction of building energy consumption, carbon emissions, and costs is essential.

[0003] Existing technologies have yielded research on multi-objective optimization methods for buildings, including energy consumption, carbon emissions, and costs. For example, Chinese patent CN118551669A discloses a multi-objective energy-saving optimization method for building envelopes based on climate prediction. This patent uses simulation parameters and results to determine a sample set. A building performance prediction model based on a backpropagation neural network is trained using this sample set, and the model predicts building energy consumption, carbon emissions, thermal discomfort time, and costs. Simulation parameters include site climate data, envelope design parameters, and indoor thermal environment. This patent combines site climate data for multi-performance prediction. However, its sample data consists of simulation results, not measured data, and does not closely reflect reality, failing to incorporate actual material prices, leading to errors in the predicted performance. Furthermore, building performance parameters such as carbon emissions, cost, and energy consumption are closely related to the material properties of building components and the building's spatial characteristics. However, this patent does not incorporate the building's spatial characteristics and the material properties of its components into its performance prediction; therefore, the prediction results still have room for improvement. Summary of the Invention

[0004] The present invention aims to at least solve the technical problems in existing building performance prediction methods, where the sample data is simulated rather than measured data, and the spatial characteristics of the building and the material properties of each component are not incorporated, so that the building performance prediction results have room for optimization and improvement. The present invention provides a building performance prediction method and evaluation system based on a hybrid neural network.

[0005] To achieve the above-mentioned objectives of the present invention, according to a first aspect of the present invention, the present invention provides a building performance prediction method based on a hybrid neural network, comprising: real-time acquisition of material price data from multiple building material supplier websites; acquisition of historical energy consumption data, site climate data, and BIM model data of the building; extraction of three-dimensional volume data of the building from the BIM model data; and prediction of the building's energy consumption, cost, and life-cycle carbon emissions using a pre-trained hybrid neural network model based on the material price data, historical energy consumption data, site climate data, and three-dimensional volume data. The hybrid neural network model includes: a price feature conversion unit for converting material price data into price features; a first long short-term memory network for extracting energy consumption time features from historical energy consumption data; a second long short-term memory network for extracting climate time features from site climate data; a three-dimensional convolutional neural network for extracting three-dimensional spatial features of the building from the three-dimensional volume data; a stitching module for stitching together energy consumption time features, climate time features, price features, and three-dimensional spatial features to obtain a comprehensive feature representation; a fully connected module including one or more cascaded fully connected layers for extracting nonlinear features from the comprehensive feature representation; and an output module for performing regression processing on the nonlinear features to obtain the building's energy consumption, cost, and life-cycle carbon emissions.

[0006] To achieve the above-mentioned objectives of the present invention, according to a second aspect of the present invention, the present invention provides a building performance evaluation system based on a hybrid neural network, used to implement the building performance prediction method based on a hybrid neural network described in the first aspect of the present invention, comprising:

[0007] The first data acquisition module retrieves material price data in real time from multiple building material supplier websites; the second data acquisition module acquires historical energy consumption data, site climate data, and BIM model data; the third data acquisition module extracts the building's 3D volume data from the BIM model data; the prediction module uses a pre-trained hybrid neural network model to predict the building's energy consumption, cost, and life-cycle carbon emissions based on material price data, historical energy consumption data, site climate data, and 3D volume data. The hybrid neural network model includes: a price feature conversion unit to convert material price data into price features; a first long short-term memory network to extract energy consumption time features from historical energy consumption data; a second long short-term memory network to extract climate time features from site climate data; a 3D convolutional neural network to extract the building's 3D spatial features from the 3D volume data; a stitching module to stitch together energy consumption time features, climate time features, price features, and 3D spatial features to obtain a comprehensive feature representation; a fully connected module including one or more cascaded fully connected layers for extracting nonlinear features from the comprehensive feature representation; and an output module that performs regression processing on the nonlinear features to obtain the building's energy consumption, cost, and life-cycle carbon emissions.

[0008] The beneficial technical effects of this invention are as follows: This invention introduces material price data crawled in real time from multiple building material supplier websites to predict building costs, fully considering the impact of material price fluctuations on building costs and closely reflecting reality; it uses measured historical energy consumption data to extract energy consumption time features through a first long short-term memory network, while simultaneously utilizing a second long short-term memory network to obtain climate time features, which can capture the variation patterns of building performance (especially energy consumption) under different seasons and meteorological conditions, improving prediction accuracy and real-world adaptability; in three-dimensional volume data, cubic element units are mapped and associated with material types and material physical properties, so that the three-dimensional convolutional neural network can better extract the geometric spatial features of the building and material distribution, making the cost and carbon emission predictions of the hybrid neural network model more accurate and more realistically adaptable; through the integration of multi-source data and intelligent optimization technology, this invention achieves high-precision and highly realistically adaptable multi-performance prediction of buildings, improving the prediction and optimization capabilities of building design costs, energy consumption, and carbon emissions throughout the entire life cycle. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a preferred embodiment of the building performance prediction method based on a hybrid neural network according to the present invention.

[0010] Figure 2 This is a schematic diagram of the structure of a hybrid neural network model in a preferred embodiment of the present invention;

[0011] Figure 3 This is a structural block diagram of the evaluation system in a preferred embodiment of the present invention;

[0012] Figure 4 This is a flowchart illustrating a specific application scenario of the building performance prediction method based on a hybrid neural network according to the present invention. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0015] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0016] The building performance prediction method based on hybrid neural networks provided by this invention can be executed by at least one of the following electronic devices: a server, a terminal, or other electronic devices that can be configured to execute the method provided in the embodiments of this application. In other words, this building performance prediction method based on hybrid neural networks can be executed by software or hardware installed on a terminal device or a server device. The software can be a blockchain platform. The server includes, but is not limited to: a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be a standalone server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0017] This invention provides a building performance prediction method based on hybrid neural networks. In a preferred embodiment, such as... Figure 1 As shown, the method includes:

[0018] Step S1: Grab material price data in real time from multiple building material supplier websites; obtain historical energy consumption data, site climate data, and BIM model data for the building.

[0019] In this embodiment, target building material supplier websites are selected, preferably but not limited to China Building Materials Network and Wanlong Building Materials Network. Web scraping technology is used periodically to crawl webpage data from these target building material supplier websites. By analyzing the webpage structure, key data fields stored in HTML tags are identified, thereby obtaining price data for various types of building materials. In one example of material price data scraping, the Python Scrapy framework is used to automatically scrape price information from webpages. This information is typically contained in specific HTML elements on product detail pages (such as...). Within the label<spanclass="price"> The crawling process includes handling the website's pagination logic to ensure a complete dataset is collected from each material product category. Material price data is numerical.

[0020] In this embodiment, the building's historical energy consumption data is not limited to the total electricity consumption of the building's air conditioning system, lighting, and other equipment over multiple past time periods, and can be collected through the building's energy monitoring system (such as the main electricity meter). The historical energy consumption data is stored in time series form, representing the building's energy consumption changes over different time periods, such as daily, weekly, and monthly energy consumption trends.

[0021] In this embodiment, climate factors have a significant impact on building energy consumption and carbon emissions. Site climate data includes temperature, humidity, solar radiation, etc., at the building's geographical location. Site climate data can be obtained through meteorological APIs or meteorological monitoring systems. Recording site climate data at different time periods forms time series data, which serves as input to capture the impact of climate conditions on building performance (such as energy consumption).

[0022] In this embodiment, a BIM model of the building is pre-built. The BIM model is usually stored in a standard format (such as IFC or Revit format) to obtain BIM model data. BIM model data includes geometry, material type, location, physical properties, etc.

[0023] Step S2: Extract the three-dimensional volume data of the building from the BIM model data.

[0024] In this embodiment, in one example, the processing procedure of step S2 is preferably, but not limited to, voxelizing the building's BIM model, converting the 3D building design into a voxel grid, with a preset grid size, discretizing the BIM model with a grid of the same size, and using the obtained voxel grid data as 3D volume data to express the building's 3D spatial structure in detail. In another example, the method for obtaining the building's 3D volume data is: projecting the building's BIM model into multiple 2D views (such as floor plans and elevations), and then inputting the multiple views as 3D volume data into a 2D CNN, or directly inputting them into a 3D CNN to utilize the 3D information.

[0025] Step S3: Using a pre-trained hybrid neural network model, predict the building's energy consumption, cost, and life-cycle carbon emissions based on material price data, historical energy consumption data, site climate data, and three-dimensional volume data. For example... Figure 2 As shown, the hybrid neural network model includes:

[0026] The price feature transformation unit is used to convert material price data into price features. This unit primarily standardizes the price data for various material types, such as linearly mapping all prices to a specific numerical range to facilitate subsequent feature extraction and processing. The price feature is a one-dimensional vector, including the material code for each material type and its standardized price value.

[0027] The First Long Short-Term Memory (LSTM) network is used to extract temporal features of energy consumption from historical energy consumption data. The LSTM network learns and predicts the temporal changes in historical energy consumption data to obtain energy consumption trends.

[0028] The Second Long Short-Term Memory (LSTM) network is used to extract temporal climate features from site climate data. The LSTM network learns and predicts climate, enabling hybrid neural network models to learn how changes in site climate over time affect building performance, such as cost, energy consumption, and life-cycle carbon emissions.

[0029] A 3D convolutional neural network (CNN) is used to extract the 3D spatial features of a building from 3D volume data. The 3D CNN, a type of CNN network, extracts 3D spatial features that characterize the geometric spatial features of the 3D voxel mesh and material distribution.

[0030] The price feature transformation unit, the first long short-term memory network, the second long short-term memory network, and the three-dimensional convolutional neural network can be executed in parallel to improve processing efficiency.

[0031] The stitching module stitches together energy consumption time features, climate time features, price features, and 3D spatial features to obtain a comprehensive feature representation. The 3D spatial feature output is a multidimensional feature tensor. The stitching module executes the following:

[0032] 1) Flatten the three-dimensional spatial features and convert them into a one-dimensional vector, as shown below:

[0033]

[0034] in, Representing three-dimensional spatial features, The function indicates that the value of a row or column is automatically calculated based on the total number of elements in x.

[0035] 2) The stitching module flattens the three-dimensional spatial features to obtain a one-dimensional vector, then stitches it together with the energy consumption time features, climate time features, and price features to form a longer one-dimensional feature vector. This feature vector is denoted as the comprehensive feature representation, which can simultaneously contain spatial and temporal information. The stitching module's processing can be represented as follows:

[0036]

[0037] in, This represents the energy consumption time characteristic of the output of the first long short-term memory network. This represents the climatic temporal characteristics of the output of the second long short-term memory network. Indicates price characteristics, This represents the 3D spatial features of a building extracted by a 3D convolutional neural network. Flatten the shape.

[0038] A fully connected module comprises one or more cascaded fully connected layers for extracting nonlinear features from a comprehensive feature representation. A fully connected module may include one or more cascaded fully connected layers, each of which can be viewed as a linear transformation followed by a nonlinear activation function. This helps the model capture complex nonlinear relationships. The processing of a fully connected module can be represented as follows:

[0039]

[0040] in, Indicates nonlinear characteristics, , These represent the weight matrix and bias matrix of the fully connected module, respectively. This represents the activation function of the fully connected module, preferably, but not limited to, the ReLU function.

[0041] The output module performs regression processing on the nonlinear characteristics to obtain the building's energy consumption, cost, and carbon emissions over its entire life cycle.

[0042] The output module can use a regression model to predict the building's current energy consumption, cost, and life-cycle carbon emissions. The output module may include a fully connected layer and an activation function processing unit.

[0043] In this embodiment, preferably, to improve prediction accuracy, a material price data cleaning step is also included: using Python's Pandas library to clean the scraped data, including removing non-numeric characters, handling missing values, removing duplicates, and standardizing the data format for subsequent analysis and model training.

[0044] In a preferred embodiment, to facilitate the simultaneous extraction of the geometric spatial features and material distribution characteristics of the building by the three-dimensional convolutional neural network, and to simultaneously map the material and physical properties to voxel units, step S2 involves extracting the three-dimensional volume data of the building from the BIM model data, including:

[0045] Step S21 involves discretizing the building's BIM model into multiple cubic elements using voxelization. In one example, the same cubic element size is used to discretize the building's BIM model into multiple cubic elements.

[0046] Step S22: Extract the material type of each cubic element from the BIM model data, and map the material code and material physical property vector corresponding to the material type to the cubic element.

[0047] Material type and physical properties are key parameters for calculating building costs, energy consumption, and carbon emissions. Mapping them to cubic cells helps 3D convolutional neural networks accurately extract building geometric spatial features and material distribution characteristics.

[0048] Material types are typically represented by discrete categorical variables, which can be expressed using integer codes. Each material type is assigned a unique identifier (i.e., a material code).

[0049]

[0050] in, It is an integer corresponding to the center point position of the cubic element. The material type in the example. A value of 0 could represent air. 1 represents concrete. 2 represents steel, etc.

[0051] Each type of material corresponds to a material physical property vector, which includes, but is not limited to, material density, material thermal conductivity, material specific heat capacity, and material carbon emission factor.

[0052] Material density is a physical property that describes the relationship between a material's mass and volume, and can be expressed by the following formula:

[0053]

[0054] Where ρ represents the material density at the center point (x, y, z) of the cubic element, and the unit is usually kg / m³. 3 .

[0055] The thermal conductivity of a material is the ratio of the heat flux per unit area of ​​the material to the temperature difference across the material, expressed by the following formula:

[0056]

[0057] in, Indicates the position of the center point. The thermal conductivity of the cubic elemental material is expressed in W / (m·K).

[0058] Specific heat capacity of a material, measured in J / (kg·K), is used in thermal performance analysis, such as evaluating the material's thermal buffering capacity in energy consumption optimization.

[0059] The carbon emission factor (e) of materials c As numerical data, it can be input into the model after standardization.

[0060] In this embodiment, each physical property in the material physical property vector can be stored in the corresponding cubic element in the form of a floating-point number.

[0061] Step S23: Represent each cubic element as a three-dimensional array, wherein each three-dimensional array includes the coordinates of the center point of the cubic element, the material code mapped to the cubic element, and the material physical property vector. Preferably, to reduce data storage redundancy and improve model processing efficiency, the three-dimensional array in the three-dimensional volume data is represented as a sparse matrix:

[0062]

[0063] in, Represents cubic element units The coordinates of the center point Represents cubic element units Mapped material type encoding, Represents cubic element units The mapped material physical property vector.

[0064] This implementation not only reduces redundant data by using sparse matrices, but also speeds up data reading and processing.

[0065] Step S24: Combine the three-dimensional array of all cubic elements to obtain the three-dimensional volume data of the building.

[0066] In this embodiment, to further reduce data storage and improve data processing efficiency, when storing the three-dimensional volume data of a building, the region formed by multiple consecutive cubic elemental units with the same material physical property vector is compressed and represented as follows:

[0067]

[0068] in, and These represent the coordinates of the center points of the two boundary cubic elements of a region formed by multiple consecutive cubic elements with the same material physical property vector. In one example, and These can represent the coordinates of the center points of the cube elements at the bottom left and top right vertices of the aforementioned regions (either cuboid or cube regions). and It can also represent the coordinates of the center points of the cube elements at the top left and bottom right vertices of the aforementioned regions (either cuboid or cubic regions). Through compressed representation, voxel data of homogeneous materials or blank regions can be efficiently stored, optimizing storage overhead and improving computational performance.

[0069] In this embodiment, to improve the mapping rate, the aforementioned material codes and material physical property vectors can be associated and mapped with various parts of the building (such as walls, floors, roofs, etc.). This can be achieved through the following steps and mapping relationships: Information about various parts of the building, such as walls, floors, roofs, and columns, is extracted from the BIM model. Each part corresponds to a specific geometric region and typically contains its material type information in the BIM model. Based on the part material information stored in the BIM model, the material type (such as concrete, steel) can be mapped to the corresponding material code by querying the material database. The material code is stored in each voxel element so that 3D CNNs can extract these features. Based on the material type code, the material's physical property vectors (such as density, thermal conductivity, carbon emission factor, etc.) are associated. This physical property data can be obtained from the material property database of the BIM model and mapped to the corresponding cubic voxel elements. For example, assuming the wall is made of concrete, all cubic voxel elements within the wall area are assigned the physical property data of concrete.

[0070] In a preferred embodiment, since the size of the cubic element directly affects the accuracy of the mesh and the computational complexity of the subsequent model, the size of the cubic element is set using an adaptive method for each region. Therefore, step S21, discretizing the building's BIM model into multiple cubic elements through voxelization processing, includes:

[0071] Step S211: Set the 3D geometric bounding box that is circumscribed to the BIM model. The 3D geometric bounding box can consist of the outer border of at least one stacked cube or cuboid, and the number and size of the stacked cubes or cuboids are determined according to the geometry of the BIM model. The occupied space area of ​​the BIM model is estimated through the geometric bounding box to determine the overall extent of the cube grid.

[0072] Step S212: Divide the three-dimensional geometric bounding box into multiple spatial regions.

[0073] In one example, the size of a spatial region is more than 10 times the size of a cubic element. A 3D geometric bounding box can be uniformly or non-uniformly divided into multiple spatial regions, whose geometry can be a cube, cuboid, or irregular shape.

[0074] Step S213: Determine the geometric complexity and material property change rate of each spatial region, and calculate the size of the cubic element in each spatial region based on the geometric complexity and material property change rate. The higher the geometric complexity and the greater the material property change rate within a spatial region, the smaller the size of the cubic element. The preferred size of the cubic element is, but not limited to, the diagonal length or the side length of the cube.

[0075] Step S214: Discretize each spatial region into one or more cubic elements based on the size of the cubic element in each region. The discretization process can refer to existing Marching Cubes algorithms or uniform meshing algorithms. The Marching Cubes algorithm is mainly used to decompose complex surface geometry (such as curved building facades) into multiple cubes.

[0076] This implementation allows for the use of smaller cubic units in complex geometric regions and spatial areas with significant variations in material properties, thereby improving the precision of feature extraction and the accuracy of building performance prediction. Generally, the cubic unit volume can be set between 1 cm³ and 10 cm³ units, depending on the required precision.

[0077] In this embodiment, to remove redundant cubic element units in step S21 and improve the accuracy of the obtained cubic element units, it is preferable to further include:

[0078] Determine whether each cubic element is covered by the geometry of the building, and discard the cubic elements that are not covered by the geometry of the building.

[0079] In this embodiment, if the center point of a cube element is located within the geometry of the building, the cube element is considered to be covered by the building, and the coverage flag is set to 1. If the center point of a cube element is not located within the geometry of the building, the cube element is considered not to be covered by the building, and the coverage flag is set to 0.

[0080] In a preferred embodiment, the size of the cubic element within each spatial region is calculated according to the following formula. :

[0081]

[0082] in, Indicates the geometric complexity of each spatial region; This represents the rate of change of material properties in each spatial region. Represents the diagonal length or side length of a cubic element; This represents an auxiliary constant, greater than or equal to 1 cm. , This represents the sum of the surface areas of the components of the BIM model contained in each spatial region. This represents the sum of the volumes of the BIM model components housed within each spatial region. Units are If a spatial region is located inside a component in the BIM model, but does not contain any surface of that component, then The value equals 0 if a spatial region contains the partial surfaces of two or more components. It is equal to the ratio of the sum of the areas of the local surfaces of all the components it contains to the sum of the volumes of the portions of all the components it contains within that spatial region. , This represents the number of physical properties in the material's physical property vector, where s represents the index of each physical property in the vector. This represents the maximum value of the s-th physical attribute within the spatial region. This represents the minimum value of the s-th physical property within the spatial region.

[0083] The above formula enables the use of smaller cubic elements in complex geometric regions and spatial regions with large variations in material properties, thereby improving the accuracy of three-dimensional volume data representation.

[0084] In a preferred embodiment, to facilitate subsequent simplification of discrete operations, step S213 includes:

[0085] Step a: Determine the geometric complexity and material property variation rate for each spatial region.

[0086] A building's BIM model can be viewed as composed of multiple components, with each spatial area exhibiting geometric complexity. This can be the number of components present within the spatial region. The formula for calculating the rate of change of material properties for each spatial region is:

[0087] ;

[0088] in, Indicates the quantity of different types of materials within a spatial region; This represents the number of physical properties in the material's physical property vector, where s represents the index of each physical property in the vector. This represents the maximum value of the s-th physical attribute within the spatial region. This represents the minimum value of the s-th physical property within the spatial region.

[0089] Step b: Calculate the cubic element size of each spatial region based on its geometric complexity and the rate of change of material properties. The cubic element size of each spatial region can be represented by the number of cubic elements included in the discretized spatial region. This means that, under the condition that the spatial region is the same size, A larger value indicates that the size of the cubic element in that spatial region is smaller. A smaller value indicates a larger cubic element size in that spatial region. The number of cubic elements... for:

[0090]

[0091] In the above steps, Indicates the AND operation. The value closest to a power of 2. For example, if... The value is 7.6, so the closest power of 2 is 8. ;like The value is 2.6, then .

[0092] In a preferred embodiment, in order to improve the prediction accuracy of the hybrid neural network model,

[0093] The training process of a hybrid neural network model includes:

[0094] Step B1: Obtain material price data for different time periods, site climate data, historical energy consumption data for multiple buildings, and BIM model data for multiple buildings.

[0095] Specifically, material price data, site climate data, and historical energy consumption data for multiple buildings are acquired for each time period. The acquired material price data, site climate data, and historical energy consumption data for multiple buildings undergo preprocessing such as cleaning and normalization.

[0096] Step B2: Extract 3D volume data based on the BIM model data of different buildings to obtain different 3D volume data; obtain corresponding 3D volume data for each building.

[0097] Step B3: Combine the three-dimensional volume data of each building with the material price data, historical energy consumption data and site climate data of the building at different time periods to obtain samples for each building at different time periods, and set corresponding labels for the samples to obtain the sample dataset.

[0098] In this implementation, each building can correspond to multiple samples, but each sample spans a different time period. Different buildings within the same time period correspond to different samples. A sample includes material price data, site climate data, historical energy consumption data, and three-dimensional volume data for a given time period. Each sample is labeled with three true values: the true value of building cost, the true value of building energy consumption, and the true value of carbon emissions throughout the building's life cycle.

[0099] The material price data from this sample is input into professional cost estimation software (such as Glodon). This software provides detailed cost estimates for the construction phase, including material, labor, and machinery costs. The construction cost obtained after calculating these detailed cost data using the software is considered the true value of the construction cost. The material price at time point t is:

[0100]

[0101] Where, p i and q i These are the unit price and quantity of the i-th material, respectively.

[0102] The true value of building energy consumption is obtained by using the actual energy consumption at the next moment in the corresponding time period of the sample. Based on the building BIM model data and actual energy consumption data, and in accordance with relevant carbon emission calculation standards, such as the "Guidelines for Building Carbon Emission Calculation," these standards comprehensively calculate the carbon emissions of the building throughout its entire process from construction to operation to demolition, based on the carbon emission factors of different materials and energy uses. This carbon emission amount is then used as the true value of carbon emissions. This approach improves the accuracy and usability of the hybrid neural network model.

[0103] Step B4: Construct the network structure of the hybrid neural network model.

[0104] Step B5 involves iteratively training the constructed hybrid neural network model using the sample dataset until the training stopping condition is met. During iterative training, the network parameters in the hybrid neural network model are optimized through backpropagation based on the gradient of the loss function. The loss function... for:

[0105]

[0106] in, Represents the network parameters of a hybrid neural network model; , , These represent the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. This represents the mean squared error between the cost output by the hybrid neural network model during training and the true value of the building cost in the sample labels; This represents the mean square error between the energy consumption output by the hybrid neural network model during training and the true value of building energy consumption in the sample labels. This represents the mean squared error between the total life-cycle carbon emissions output by the hybrid neural network model during training and the true total life-cycle carbon emissions in the sample labels.

[0107] In a preferred embodiment, to optimize the design of new building projects and the renovation of existing buildings, and to achieve multi-objective optimization, the method further includes optimizing the design parameters of new buildings based on the NSGA-III algorithm, including:

[0108] Step A1: Initialize a population containing N solutions, each representing a set of design parameters for a new building; set a target performance set, including target energy consumption, target cost, and target life-cycle carbon emissions, where N is a positive integer; each set of design parameters is finite but not limited to the building's geometric features, number of floors, orientation, etc. The size of N is preferably, but not limited to, 500. Each solution represents a potential building design scheme.

[0109] Step A2 involves performing crossover and mutation operations on the population to generate a progeny population, which contains N solutions. The progeny population and the population are then merged to obtain a candidate population. The candidate population contains 2N solutions. Binary crossover (crossover rate of 0.9) and multi-point mutation (mutation rate of 0.1) are used as genetic operations to maintain genetic diversity.

[0110] Step A3 divides the candidate population into several different non-dominated layers based on Pareto-dominated non-dominated ordination. In each generation, individuals are selected to enter the next generation by comparing non-dominated ordination and crowding, ensuring the diversity and quality of solutions.

[0111] Step A4: Calculate the distance between each solution and the target performance set in each non-dominated layer, and select the N solutions that are closest to the target performance set from multiple non-dominated layers as a new population.

[0112] Step A5: Repeat steps A2-A4 until the iteration stopping condition is met, then proceed to step A6. The iteration stopping condition is preferably, but not limited to, reaching the maximum number of iterations.

[0113] Step A6: Select the solution that is closest to the target performance set from the new population as the Pareto optimal solution, and output the design parameters of the Pareto optimal solution.

[0114] In each generation of a genetic algorithm, the solution set evolves, eventually generating a set of Pareto optimal solutions. Pareto optimal solutions represent different trade-offs where an improvement in any objective (cost, energy consumption, or carbon emissions) may worsen other objectives. NSGA-III improves the breadth and accuracy of the solution space search by optimizing algorithm parameters such as population size, crossover probability, and mutation probability.

[0115] In this embodiment, preferably, step A4, calculating the distance between each solution and the target performance set in each non-dominated layer, includes:

[0116] Step A41: Obtain current material price data and site climate data.

[0117] Step A42: Generate BIM model data for the new building based on the design parameters of each solution. Use energy simulation software to obtain the current energy consumption simulation data of the new building based on the BIM model data of the new building, and use the energy consumption simulation data as the historical energy consumption data of the new building.

[0118] Step A43: Extract the three-dimensional volume data of the new building from the BIM model data of the new building;

[0119] Step A44: Input the three-dimensional volume data, material price data, site climate data and historical energy consumption data of the new building into the pre-trained hybrid neural network model to obtain the predicted energy consumption, cost and life cycle carbon emissions of the new building. Combine the predicted energy consumption, cost and life cycle carbon emissions of the new building into the performance set of each solution.

[0120] Step A45: Calculate the distance between the performance set of each solution and the target performance set. Specifically, calculate the ratio of the absolute value of the difference between the predicted energy consumption of the new building and the target energy consumption of the target performance set to the target energy consumption, as the energy consumption ratio; calculate the ratio of the absolute value of the difference between the predicted cost of the new building and the target cost of the target performance set to the target cost, as the cost ratio; calculate the ratio of the absolute value of the difference between the predicted life-cycle carbon emissions of the new building and the target life-cycle carbon emissions of the target performance set to the target life-cycle carbon emissions, as the carbon emission ratio; sum the cost ratio, energy consumption ratio, and carbon emission ratio to obtain the total distance between the performance set of each solution and the target performance set. This distance represents the degree of closeness between the performance set of each solution and the target performance set.

[0121] In this implementation, the NSGA-III multi-objective optimization algorithm can also generate a set of Pareto optimal solutions. These solutions provide the best trade-offs in terms of cost, energy consumption, and carbon emissions, helping decision-makers select the optimal design that meets the specific needs of a project. The generation of the Pareto front makes the optimization process more transparent, allowing decision-makers to intuitively understand the performance of different design schemes across multiple dimensions and make balanced choices based on project priorities such as budget, energy conservation, or environmental protection. This multi-objective optimization provides greater flexibility to architectural design, making the design process more efficient and controllable.

[0122] The optimized solution set will be presented through visualization technology to help decision-makers understand the performance of different building design schemes in terms of cost, energy consumption, and carbon emissions. The Matplotlib library in Python will be used to plot the dynamic changes in the optimization process, such as the decreasing trends of cost and energy consumption with the number of iterations. Interactive visualizations using Bokeh will show the performance of different design schemes across the three dimensions of cost, energy consumption, and carbon emissions. Furthermore, the optimization results will be integrated into Power BI to build an interactive dashboard that displays the performance metrics of each design scheme in real time and allows users to adjust and compare different schemes according to different priorities.

[0123] In this implementation, regarding visualization, all optimization results can be presented intuitively through visualization tools, including predictions and comparisons of key indicators such as building costs, energy consumption, and carbon emissions. Using interactive visualization platforms (such as Power BI or Tableau), management can track changes brought about by design adjustments in real time and make accurate decisions through data-driven approaches. This visualization function not only facilitates the work of architects and project managers but also improves communication efficiency with investors or other stakeholders, allowing every participant to clearly understand the optimization results and make informed choices.

[0124] From a long-term benefit perspective, the solution implemented in this way also has dynamic update capabilities. As the market environment changes, such as fluctuations in material prices, technological advancements, or policy changes, the system can flexibly adjust and optimize the process, ensuring that the design of the building project always possesses optimal economic efficiency and sustainability. Simultaneously, during the building's operation phase, the real-time input of relevant data can provide continuous optimization support for the building's long-term maintenance and energy management, ensuring that the project exhibits optimal benefits throughout its entire life cycle.

[0125] In summary, this technical solution, through multi-dimensional intelligent integration and optimization, ensures that building designs possess efficient cost control, energy consumption optimization, and carbon emission reduction capabilities throughout the initial planning, construction, and operation phases, meeting the requirements of green building and low-carbon development. This systematic and dynamically optimized solution provides a forward-looking and sustainable solution for building projects, significantly improving their economic benefits and environmental friendliness.

[0126] Figure 4 The diagram illustrates the specific process of the building performance prediction method based on hybrid neural networks provided by this invention in an application scenario.

[0127] This invention also discloses a building performance evaluation system based on a hybrid neural network, used to implement the aforementioned building performance prediction method based on a hybrid neural network, with reference to... Figure 3 As shown, it includes:

[0128] The first data acquisition module retrieves material price data in real time from multiple building material supplier websites;

[0129] The second data acquisition module acquires historical energy consumption data, site climate data, and BIM model data of the building.

[0130] The third data acquisition module extracts the three-dimensional volume data of the building from the BIM model data;

[0131] The prediction module uses a pre-trained hybrid neural network model to predict a building's energy consumption, cost, and life-cycle carbon emissions based on material price data, historical energy consumption data, site climate data, and 3D volume data. The hybrid neural network model includes:

[0132] The price feature conversion unit is used to convert material price data into price features;

[0133] The first long short-term memory network is used to extract energy consumption time features from historical energy consumption data;

[0134] The second long short-term memory network is used to extract climate temporal features from site climate data.

[0135] A 3D convolutional neural network is used to extract the 3D spatial features of a building from 3D volume data.

[0136] The splicing module splices together energy consumption time characteristics, climate time characteristics, price characteristics, and three-dimensional spatial characteristics to obtain a comprehensive feature representation;

[0137] A fully connected module includes one or more cascaded fully connected layers for extracting nonlinear features from a comprehensive feature representation;

[0138] The output module performs regression processing on the nonlinear characteristics to obtain the building's energy consumption, cost, and carbon emissions over its entire life cycle.

[0139] In this embodiment, the first data acquisition module and the second data acquisition module correspond to step S1 in the above-mentioned building performance prediction method based on hybrid neural networks, the third data acquisition module corresponds to step S2 in the above-mentioned building performance prediction method based on hybrid neural networks, and the prediction module corresponds to step S3 in the above-mentioned building performance prediction method based on hybrid neural networks. The specific execution process of each module will not be described in detail here.

[0140] This invention provides a building performance prediction method and evaluation system based on hybrid neural networks. Through multi-dimensional integrated optimization, it can significantly improve the overall performance of building design in terms of cost, energy consumption and carbon emissions throughout the entire life cycle.

[0141] Firstly, regarding cost optimization, the technical solution combines material market data obtained from Web Scraping with cost data from cost estimation software. Using 3D CNN neural networks to analyze and extract architectural design parameters, it can generate accurate construction cost predictions. Through continuously updated market data, the construction cost prediction results, adjusted for design changes, can also be dynamically output in real time. This allows project managers to make informed decisions during the design phase, ensuring strict control over the construction budget while maximizing savings on material and labor costs and improving economic efficiency.

[0142] In terms of energy consumption prediction and optimization, this solution incorporates measured energy consumption data and combines it with an LSTM neural network model to more accurately capture the actual time-series energy consumption of buildings under different climatic conditions. The solution can simulate energy use during future operational phases, helping to optimize the building's energy management system and providing solutions to reduce energy consumption, such as improving ventilation design, selecting insulation materials, or optimizing the operation and scheduling of energy equipment. This is particularly important for green building design, enabling buildings to exhibit higher energy efficiency during operation and reducing long-term energy consumption.

[0143] Meanwhile, the scheme adheres to China's carbon emission calculation standards, using a neural network model to accurately estimate the carbon emissions throughout the building's entire life cycle. This includes carbon emissions from material production and construction to energy consumption during operation and waste disposal during demolition. Based on this estimation, building designers can adjust material selection or design optimizations according to the carbon emission prediction results. For example, choosing low-carbon building materials or increasing the proportion of clean energy use can effectively reduce the environmental impact of building projects. This scheme also meets national requirements for low-carbon and green buildings, ensuring that building projects meet environmental standards from the design stage and contribute to global carbon reduction goals.

[0144] The key technical point of this invention is:

[0145] By integrating multi-source data and employing intelligent optimization techniques, this invention enhances the ability to predict and optimize building design costs, energy consumption, and life-cycle carbon emissions. It integrates various data sources, including building material market data obtained through Web Scraping, measured energy consumption data, Chinese carbon emission calculation standards, and building cost data generated by cost estimation software. By cleaning, standardizing, and integrating this data, the model can handle complex input data, providing robust data support for subsequent analysis and prediction.

[0146] This invention employs a hybrid neural network model that combines 3D CNN (3D Convolutional Neural Network) and LSTM (Long Short-Term Memory Network). The 3D CNN is used to process spatial features in architectural design, such as geometry and material usage; the LSTM is used to analyze the time-series characteristics of building energy consumption. This combination of models can more accurately predict building costs, energy consumption, and carbon emissions, especially when handling large-scale data and complex features.

[0147] The inclusion of measured energy consumption data is a significant feature of this invention. Unlike traditional simulation data, measured data reflects the actual energy consumption of a building during operation. By using this data, the model can more accurately capture fluctuations and trends in building energy consumption, significantly improving the realism and reliability of energy consumption predictions.

[0148] Meanwhile, this invention calculates the carbon emissions of a building throughout its entire life cycle based on China's current carbon emission calculation standards, combined with carbon emission factors and energy consumption data of building materials. This standardized process ensures that the predictions comply with national environmental protection policies and helps in designing buildings that better meet low-carbon and environmental standards.

[0149] In terms of cost calculation, this invention integrates data from cost estimation software. These software programs are widely used in the construction industry and can provide accurate construction cost estimates. By inputting this data into the model and combining it with design parameters, the overall cost of the building can be predicted more accurately. This is particularly crucial when dealing with dynamic market changes (such as fluctuations in material prices).

[0150] The application of the NSGA-III multi-objective optimization algorithm is another key point. This algorithm allows for the optimization of multiple objectives, such as building cost, energy consumption, and carbon emissions, generating a set of Pareto optimal solutions through a genetic algorithm. These solutions help decision-makers find the best balance between different optimization objectives, ensuring that the design scheme achieves an optimal combination in terms of economy, energy efficiency, and environmental protection.

[0151] To support decision-making, this invention uses visualization technology to display multi-dimensional optimization results, helping decision-makers quickly understand the advantages and disadvantages of different design options. Visualization makes the design process more transparent and also improves communication efficiency among different stakeholders. Ultimately, the entire optimization process is dynamic, continuously updated to adapt to changes in market conditions, material supply, and technology, ensuring that the architectural design remains optimal throughout its entire lifecycle.

[0152] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," "a preferred embodiment," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0153] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A building performance prediction method based on hybrid neural networks, characterized in that, include: Real-time material price data was scraped from multiple building material supplier websites; Acquire historical energy consumption data, site climate data, and BIM model data for the building; Extracting 3D volume data of a building from BIM model data; A pre-trained hybrid neural network model is used to predict a building's energy consumption, cost, and life-cycle carbon emissions based on material price data, historical energy consumption data, site climate data, and three-dimensional volume data. The hybrid neural network model includes: The price feature conversion unit is used to convert material price data into price features; The first long short-term memory network is used to extract energy consumption time features from historical energy consumption data; The second long short-term memory network is used to extract climate temporal features from site climate data. A 3D convolutional neural network is used to extract 3D spatial features of a building, including a 3D voxel mesh and material distribution, from 3D volume data. The splicing module splices together energy consumption time characteristics, climate time characteristics, price characteristics, and three-dimensional spatial characteristics to obtain a comprehensive feature representation; A fully connected module includes one or more cascaded fully connected layers for extracting nonlinear features from a comprehensive feature representation; The output module performs regression processing on the nonlinear characteristics to obtain the building's energy consumption, cost, and total life-cycle carbon emissions. The extraction of three-dimensional volume data of a building from BIM model data includes: The building's BIM model is discretized into multiple cubic elements through voxelization. Extract the material type of each cubic element from the BIM model data, and map the material code and material physical property vector corresponding to the material type to the cubic element. Each cubic element is represented as a three-dimensional array, where each three-dimensional array includes the coordinates of the center point of the cubic element, the material code mapped to the cubic element, and the material physical property vector. The three-dimensional volume data of the building is obtained by combining a three-dimensional array of all cubic elements; The process of discretizing the BIM model of a building into multiple cubic elements through voxelization includes: Set the 3D geometric bounding box that is connected to the BIM model; Divide the three-dimensional geometric bounding box into multiple spatial regions; Determine the geometric complexity and material property change rate of each spatial region, and calculate the cubic element size of each spatial region based on the geometric complexity and material property change rate of each spatial region; Each spatial region is discretized into one or more cubic elements based on the size of the cubic element in each region.

2. The building performance prediction method based on a hybrid neural network as described in claim 1, characterized in that, Also includes: Determine whether each cubic element is covered by the geometry of the building, and discard the cubic elements that are not covered by the geometry of the building.

3. The building performance prediction method based on a hybrid neural network as described in claim 2, characterized in that, The size of the cubic unit cell in each spatial region is calculated using the following formula. : ； in, Indicate the geometric complexity of each spatial region; This represents the rate of change of material properties in each spatial region.

4. A building performance prediction method based on a hybrid neural network as described in claim 1, 2, or 3, characterized in that, Also includes: Representing the three-dimensional array in the three-dimensional volume data as a sparse matrix: ； in, Represents cubic element units The coordinates of the center point Represents cubic element units Mapped material type encoding, Represents cubic element units Mapped material physical property vectors; The region formed by multiple consecutive cubic elemental units with the same material physical property vector is compressed and represented as: ； in, and These represent the coordinates of the center points of the two boundary cubic elements of a region formed by multiple consecutive cubic elements with the same material physical property vector.

5. A building performance prediction method based on a hybrid neural network as described in claim 1, 2, or 3, characterized in that, The training process of the hybrid neural network model includes: Acquire material price data, site climate data, historical energy consumption data of multiple buildings, and BIM model data of multiple buildings for different time periods; Three-dimensional volume data is extracted from BIM model data of different buildings to obtain different three-dimensional volume data. The three-dimensional volume data of each building is combined with the material price data, historical energy consumption data and site climate data of the building at different time periods to obtain samples for different time periods for each building, and corresponding labels are set for the samples to obtain sample datasets; Constructing the network structure of a hybrid neural network model; The constructed hybrid neural network model is iteratively trained using a sample dataset until the training stopping condition is met. During iterative training, the network parameters in the hybrid neural network model are optimized by backpropagation based on the gradient of the loss function. The loss function... for: ； in, Represents the network parameters of a hybrid neural network model; , , These represent the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. This represents the mean squared error between the cost output by the hybrid neural network model during training and the true value of the building cost in the sample labels; This represents the mean square error between the energy consumption output by the hybrid neural network model during training and the true value of building energy consumption in the sample labels. This represents the mean squared error between the total life-cycle carbon emissions output by the hybrid neural network model during training and the true total life-cycle carbon emissions in the sample labels.

6. The building performance prediction method based on a hybrid neural network as described in claim 5, characterized in that, It also includes optimization of design parameters for new buildings based on the NSGA-III algorithm, including: Step A1: Initialize a population containing N solutions, each solution representing a set of design parameters for the new building; set the target performance set, which includes target energy consumption, target cost, and target life-cycle carbon emissions, where N is a positive integer. Step A2: Perform crossover and mutation operations on the population to generate offspring populations. The offspring populations include N solutions. Merge the offspring populations and the population to obtain candidate populations. Step A3: Divide the candidate population into several different non-dominated layers based on the Pareto-dominated non-dominated ordination. Step A4: Calculate the distance between each solution and the target performance set in each non-dominated layer, and select the N solutions that are closest to the target performance set from multiple non-dominated layers as a new population; Step A5: Repeat steps A2-A4 until the iteration stopping condition is met, then proceed to step A6; Step A6: Select the solution that is closest to the target performance set from the new population as the Pareto optimal solution, and output the design parameters of the Pareto optimal solution.

7. The building performance prediction method based on a hybrid neural network as described in claim 6, characterized in that, In step A4, the distance between each solution and the target performance set is calculated in each non-dominated layer, including: Obtain current material price data and site climate data; Based on the design parameters of each solution, BIM model data of the new building is generated. Based on the BIM model data of the new building, energy consumption simulation software is used to obtain the current energy consumption simulation data of the new building. The energy consumption simulation data is used as the historical energy consumption data of the new building. Extract the three-dimensional volume data of the new building from the BIM model data of the new building; The three-dimensional volume data, material price data, site climate data, and historical energy consumption data of the new building are input into a pre-trained hybrid neural network model to obtain the predicted energy consumption, cost, and life-cycle carbon emissions of the new building. The predicted energy consumption, cost, and life-cycle carbon emissions of the new building are then combined to form the performance set of each solution. Calculate the distance between the performance set of each solution and the target performance set.

8. A building performance evaluation system based on a hybrid neural network, used to implement the building performance prediction method based on a hybrid neural network as described in any one of claims 1-7, characterized in that, include: The first data acquisition module retrieves material price data in real time from multiple building material supplier websites; The second data acquisition module acquires historical energy consumption data, site climate data, and BIM model data of the building. The third data acquisition module extracts the three-dimensional volume data of the building from the BIM model data; The prediction module uses a pre-trained hybrid neural network model to predict a building's energy consumption, cost, and life-cycle carbon emissions based on material price data, historical energy consumption data, site climate data, and 3D volume data. The hybrid neural network model includes: The price feature conversion unit is used to convert material price data into price features; The first long short-term memory network is used to extract energy consumption time features from historical energy consumption data; The second long short-term memory network is used to extract climate temporal features from site climate data. A 3D convolutional neural network is used to extract 3D spatial features of a building, including a 3D voxel mesh and material distribution, from 3D volume data. The splicing module splices together energy consumption time characteristics, climate time characteristics, price characteristics, and three-dimensional spatial characteristics to obtain a comprehensive feature representation; A fully connected module includes one or more cascaded fully connected layers for extracting nonlinear features from a comprehensive feature representation; The output module performs regression processing on the nonlinear characteristics to obtain the building's energy consumption, cost, and total life-cycle carbon emissions. The third data acquisition module extracts the three-dimensional volume data of the building from the BIM model data, including: The building's BIM model is discretized into multiple cubic elements through voxelization. Extract the material type of each cubic element from the BIM model data, and map the material code and material physical property vector corresponding to the material type to the cubic element. Each cubic element is represented as a three-dimensional array, where each three-dimensional array includes the coordinates of the center point of the cubic element, the material code mapped to the cubic element, and the material physical property vector. The three-dimensional volume data of the building is obtained by combining a three-dimensional array of all cubic elements; The process of discretizing the BIM model of a building into multiple cubic elements through voxelization includes: Set the 3D geometric bounding box that is connected to the BIM model; Divide the three-dimensional geometric bounding box into multiple spatial regions; Determine the geometric complexity and material property change rate of each spatial region, and calculate the cubic element size of each spatial region based on the geometric complexity and material property change rate of each spatial region; Each spatial region is discretized into one or more cubic elements based on the size of the cubic element in each region.

Images