An artificial intelligence-based urban design noise optimization method and system

Abstract

An artificial intelligence-based urban design noise optimization method and system includes: preprocessing building form and noise measurement data, constructing a basic data set after noise simulation verification; generating a three-dimensional urban model and a network of measurement points through a hierarchical parameterized framework, calculating the structure-borne noise level, and verifying it in real time in conjunction with planning constraints, providing a structured design space, basic noise data, and feasibility assurance; constructing a two-stage multilayer perceptron network to process the acoustic impact of local building clusters and global urban noise patterns, respectively, and achieving high-precision noise prediction through training; deploying a large language model as an optimization agent, and outputting the optimal solution through multiple rounds of iterative optimization and evaluation. Compared with traditional technologies, this invention effectively improves noise prediction accuracy. Parametric modeling provides a structured design space, real-time verification ensures feasibility, and the use of a large language model for optimization enables efficient iteration, significantly improving efficiency and reducing labor costs compared to traditional methods.

Description

Technical Field

[0001] The present invention relates to artificial intelligence, urban planning and design, and environmental engineering technology, and in particular to an artificial intelligence-based urban design noise optimization method and system. Background Art

[0002] During the urban design and planning stages, designers often lack effective tools to systematically consider noise mitigation factors. Traditional design methods rely primarily on the designer's experience and intuition, lacking scientific quantitative analysis and optimization methods. While some acoustic simulation software, such as CadnaA and SoundPLAN, exists, these tools are computationally complex and time-consuming, requiring professional assistance to interpret the results and provide noise reduction recommendations, making them difficult to integrate into rapidly iterative design workflows. For efficiency and cost reasons, designers often wait until the design is complete to conduct acoustic performance evaluations. If noise issues are discovered at this stage, modifying the design is both costly and difficult.

[0003] Traditional urban noise control methods rely primarily on passive interventions, such as installing noise barriers, using sound-absorbing materials, and implementing traffic controls. While these methods can alleviate noise problems to a certain extent, they suffer from high costs and impact on the urban landscape. While existing research confirms the relationship between urban form and noise transmission, a systematic approach to reducing noise hazards through urban design has yet to be established.

[0004] A review of existing technical literature reveals that while current machine learning noise prediction methods have improved computational efficiency to a certain extent, they still have shortcomings in terms of prediction accuracy and scope of applicability. First, most methods use traditional regression models or simple neural network architectures, failing to fully consider the complexity and spatial correlation of acoustic phenomena in high-density cities. Second, these prediction methods typically exist as standalone prediction tools, lacking systematic integration with the urban design process. They are unable to provide designers with directly applicable optimization guidance and are insufficient to support large-scale design optimization, limiting their application in actual engineering projects. Existing optimization methods primarily utilize traditional evolutionary algorithms, particle swarm optimization algorithms, and simulated annealing algorithms. While these methods can find good solutions, they suffer from the following drawbacks: First, they require extensive parameter tuning and empirical setup, lacking adaptability; second, they struggle to integrate domain knowledge and design experience, resulting in limited optimization efficiency; and third, they lack natural language interaction, making them inconvenient for designers to use.

[0005] It should be noted that the information disclosed in the above background technology section is only used to understand the background of this application, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0006] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide an urban design noise optimization method and system based on artificial intelligence.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] An artificial intelligence-based urban design noise optimization method comprises the following steps:

[0009] S1. Data preprocessing: Collect building form data and noise measurement data, verify the data through noise simulation to determine the background model and acoustic parameters, and build a basic data set;

[0010] S2. Parametric Modeling and Constraint Processing: A hierarchical parametric framework is established to generate a 3D city model and a network of measurement points. The structure-borne noise level of the 3D city model is calculated using the measurement point network. A real-time verification mechanism is implemented using multiple planning constraints on the 3D city model. This provides a structured design space, basic noise level data, and feasibility assurance for urban design noise optimization.

[0011] S3. Urban Noise Prediction Model Construction and Training: The 3D urban model data generated in step S2 is standardized and feature engineered to construct a training dataset. A two-stage multilayer perceptron network is constructed based on the training dataset. The first stage perceptron network processes the acoustic impact of local building clusters through parallel group-specific subnetworks. The second stage perceptron network captures global urban noise patterns through a layered dense network. The model is trained using an optimized training strategy to achieve high-precision noise prediction.

[0012] S4. Large language model optimization: Deploy a large language model as an optimization agent. Based on the structural noise level of the 3D urban model generated in step S2, an initial set of design solutions is generated through a prompt engineering framework. Combined with the urban noise prediction model constructed in step S3, a multi-objective performance evaluation is performed. After multiple rounds of iterative optimization and convergence judgment, the optimal solution is output.

[0013] An artificial intelligence-based urban design noise optimization system, including:

[0014] Data preprocessing module: collects building form data and noise measurement data, verifies the data through noise simulation to determine the background model and acoustic parameters, and constructs a basic data set;

[0015] Parametric Modeling and Constraint Processing Module: This module establishes a hierarchical parametric framework to generate a 3D city model and a network of measurement points. The noise level of the 3D city model is calculated using the network of measurement points. A real-time verification mechanism is implemented using various planning constraints on the 3D city model. This module provides a structured design space, basic noise level data, and feasibility assurance for noise optimization in urban design.

[0016] Urban noise prediction model construction and training module: This module standardizes and performs feature engineering on the generated 3D urban model data to construct a training dataset. A two-stage multilayer perceptron network is constructed based on the training dataset. The first-stage perceptron network processes the acoustic impact of local building clusters through parallel group-specific subnetworks. The second-stage perceptron network captures global urban noise patterns through a layered dense network. The model is then trained using an optimized training strategy to achieve high-precision noise prediction.

[0017] Large language model optimization: Deploy a large language model as an optimization agent. Based on the generated 3D city model and noise level, an initial set of design solutions is generated through a prompt engineering framework. Combined with the constructed urban noise prediction model, a multi-objective performance evaluation is performed. After multiple rounds of iterative optimization and convergence judgment, the optimal solution is output.

[0018] The present invention has the following beneficial effects:

[0019] This invention provides an artificial intelligence-based urban design noise optimization method and system, effectively addressing the lack of systematic consideration of noise mitigation in urban design. Traditional noise simulation methods are inefficient and lack integration with the design process. Evaluations are often conducted only after the design is complete, making it difficult to provide scientific guidance to designers. This method, through the collaborative design of deep learning neural networks, large language models, and parametric modeling, systematically incorporates noise mitigation factors in the early stages of urban design. Its core modules include a parametric urban form generation and constraint verification module, a two-stage multilayer perceptron (MLP) neural network prediction module, and a large language model intelligent optimization module. Together, these three modules form a complete AI-enhanced urban design optimization technology system, completing a technical chain from raw data input to intelligent optimization design solution output.

[0020] Ultimately, this method achieved high-precision urban noise prediction through a two-stage MLP network architecture combined with a data structure centered on measurement points. This structure conforms to the principles of acoustic propagation in complex urban spaces. Compared with the simulation results of the industrial noise simulation software CadnaA, its prediction accuracy exceeded the accuracy requirements of a Class 1 sound level meter, providing a reliable foundation for efficient design optimization. At the same time, this method applied the powerful reasoning capabilities of large language models to urban design optimization for the first time, constructing an intelligent optimization strategy. Compared with the defects of traditional evolutionary algorithms that require a large amount of parameter debugging and lack interpretability, large language model optimization has significant advantages such as strong adaptability, good interpretability, the ability to integrate domain knowledge, and support natural language interaction. It can also automatically adjust the optimization strategy according to the characteristics of the problem, significantly improving optimization efficiency and effectiveness.

[0021] In terms of efficiency, the AI ​​prediction model implemented by this invention achieves an order of magnitude improvement compared to traditional simulation methods, shortening the time for a single prediction from hours to seconds, significantly reducing time costs. Furthermore, this method, for the first time, achieves full-process intelligence for urban design optimization, integrating parametric morphology generation, noise prediction, and solution optimization. From problem understanding, solution generation, performance evaluation, optimization reasoning, to result interpretation, the entire process possesses intelligent features. This intelligent process can be integrated into a design software platform or run independently (its efficiency is not limited by the design software platform when running independently), opening up a new technical path for AI applications in the field of urban design and can be widely used in urban planning and design, real estate development, and other fields.

[0022] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the process of an urban design noise optimization method based on artificial intelligence according to an embodiment of the present invention.

[0024] Figure 2 This is a framework diagram of the urban design noise optimization system based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present invention and its application.

[0026] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0027] This invention provides an artificial intelligence-based urban design noise optimization method and system, addressing the inability of traditional technologies to systematically consider noise mitigation in urban design. Traditional noise simulation methods are inefficient and lack integration with the design process. They are typically evaluated after the design is complete, failing to provide scientific design guidance for designers. This invention utilizes a parametric urban morphology generation and constraint verification module, a two-stage multilayer perceptron (MLP) neural network prediction module, and a large language model intelligent optimization module. By combining a two-stage multilayer perceptron network architecture with a measurement-point-centric data structure, this method achieves high-precision urban noise prediction. An intelligent large language model optimization strategy enables efficient multi-objective design optimization. By organically integrating the prediction model with parametric modeling and the large language model, this method achieves noise reduction design optimization without the need for traditional simulation software or the input of professional acousticians, thereby systematically reducing noise hazards in the early stages of urban design.

[0028] See Figure 1 , an embodiment of the present invention provides an urban design noise optimization method based on artificial intelligence, comprising the following steps:

[0029] Step S1, data preprocessing: collect building form data and noise measurement data, verify the data through noise simulation to determine the background model and acoustic parameters, and construct a basic data set.

[0030] In some embodiments, building geometric parameters and data on different types of noise sources are collected; the reliability of acoustic parameters is verified through noise simulation and the background model and acoustic parameters are saved; the collected data is verified through noise simulation software to confirm the reliability of acoustic parameters and save the background model and acoustic parameters; and a basic data set including building morphology data, noise measurement data and verified acoustic parameters is constructed.

[0031] Step S2, parametric modeling and constraint processing: A hierarchical parametric framework is established to generate a 3D city model and a measurement point network. The noise level of the 3D city model is calculated using the measurement point network. A real-time verification mechanism is implemented using multiple planning constraints on the 3D city model, providing a structured design space, basic noise level data, and feasibility assurance for urban design noise optimization.

[0032] In some embodiments, a parametric generator is established to construct a three-dimensional city model in a three-dimensional design platform; an adaptive measurement point grid is deployed to match the acoustic analysis, and the grid resolution is dynamically adjusted according to the building density and spatial complexity; through multiple types of planning constraints such as floor area ratio, building coverage ratio, height limit, building spacing, greening rate, etc., the three-dimensional city model is automatically verified in real time when parameters change to ensure the feasibility of the design.

[0033] Step S3: Urban Noise Prediction Model Construction and Training: The 3D urban model data generated in step S2 is subjected to standardization processing (e.g., conversion of measurement point center coordinates to building coordinates) and feature engineering to construct a training dataset. A two-stage multilayer perceptron network is constructed based on the training dataset. The first-stage perceptron network processes the acoustic impact of local building clusters through parallel group-specific subnetworks. The second-stage perceptron network captures global urban noise patterns through a layered dense network. The model is then trained using an optimized training strategy to achieve high-precision noise prediction.

[0034] In some embodiments, the three-dimensional city model data generated in step S2 is converted into the relative coordinates of the buildings using the measurement point center coordinate system, and the building rotation angle is converted into a trigonometric function; the converted coordinates and building geometric features are Z-score standardized (the target noise value retains the original decibel unit) to construct a training data set; a two-stage multi-layer perceptron network is constructed based on the training data set, in which the first-stage perceptron network independently processes the acoustic characteristics of each building group through parallel group-specific sub-networks, and the second-stage perceptron network fuses all group output features through a layered dense network to capture the global urban noise pattern; through a weighted loss function and a dynamic weight allocation strategy, higher weights are assigned to data groups in a predetermined low sample size range (i.e., with a low sample size) to balance the data distribution; and a dynamic learning rate decay strategy with an early stopping mechanism is used for model training.

[0035] Step S4, large language model optimization: Deploy the large language model as an optimization agent. Based on the 3D city model and noise level generated in step S2, generate an initial set of design solutions through the prompt engineering framework. Combined with the urban noise prediction model constructed in step S3, a multi-objective performance evaluation is performed. After multiple rounds of iterative optimization and convergence judgment, the optimal solution is output.

[0036] In some embodiments, the large language model deployment adopts a hybrid expert architecture, integrating expert modules in urban planning, acoustic environment and economic fields; calling a vectorized professional knowledge base through a retrieval enhancement generation mechanism; and using prompt engineering to generate a diverse set of initial solutions that meet the constraints of floor area ratio, height and density.

[0037] In some embodiments, the multi-objective performance evaluation specifically includes: inputting solution parameters into a prediction model to obtain multi-location noise distribution data; calculating comprehensive indicators of noise control, economic benefits and regulatory compliance; and performing weighted scoring, ranking and constraint violation screening through a large language model.

[0038] In some embodiments, the multiple rounds of iterative optimization specifically include: identifying the solution optimization direction based on the performance evaluation results; generating an improved solution set through solution component exchange and geometric transformation; maintaining a historical optimal solution library and dynamically updating prompt engineering strategies.

[0039] In some embodiments, the convergence judgment specifically includes: calculating the change in the Euclidean distance of the design variable vector of continuous iterations; determining convergence when the change is lower than a preset threshold; and verifying the repeatability of the results through multiple (e.g., 3-5) independent optimizations.

[0040] In some embodiments, the final output specifically includes: optimal solution building parameters verified through multiple rounds of iterations; a multi-objective performance indicator evaluation report and constraint satisfaction status; and a complete decision record including the solution optimization trajectory.

[0041] An embodiment of the present invention further provides an artificial intelligence-based urban design noise optimization system, comprising:

[0042] Data preprocessing module: collects building form data and noise measurement data, verifies the data through noise simulation to determine the background model and acoustic parameters, and constructs a basic data set;

[0043] Parametric Modeling and Constraint Processing Module: This module establishes a hierarchical parametric framework to generate a 3D city model and a network of measurement points. The noise level of the 3D city model is calculated using the network of measurement points. A real-time verification mechanism is implemented using various planning constraints on the 3D city model. This module provides a structured design space, basic noise level data, and feasibility assurance for noise optimization in urban design.

[0044] Urban noise prediction model construction and training module: This module performs standardization processing (using the measurement point center coordinate system to transform building coordinates) and feature engineering on the generated 3D urban model data to construct a training dataset. A two-stage multilayer perceptron network is constructed based on the training dataset. The first-stage perceptron network processes the acoustic impact of local building clusters through parallel group-specific subnetworks. The second-stage perceptron network captures global urban noise patterns through a layered dense network. The model is then trained using an optimized training strategy to achieve high-precision noise prediction.

[0045] Large language model optimization: Deploy a large language model as an optimization agent. Based on the generated 3D city model and noise level, an initial set of design solutions is generated through a prompt engineering framework. Combined with the constructed urban noise prediction model, a multi-objective performance evaluation is performed. After multiple rounds of iterative optimization and convergence judgment, the optimal solution is output.

[0046] This paper proposes an AI-based urban design noise optimization method and system. Through the collaborative design of a two-stage multilayer perceptron neural network, a large language model, and parametric modeling, it constructs an AI-enhanced urban design noise optimization system. High-precision noise prediction is achieved by combining a two-stage MLP network with a measurement point-centric data structure. Intelligent multi-objective optimization is achieved through the reasoning capabilities of the large language model. Parametric modeling provides a structured design space, forming a full-process intelligent optimization mechanism from data processing to solution output. Experimental results demonstrate that the urban noise prediction accuracy achieved by this method meets the standards of a Class 1 sound level meter, and its efficiency is improved to seconds compared to traditional simulations. The model boasts strong adaptability and good interpretability, and it achieves intelligent design throughout the entire design process, opening up new avenues for AI applications in urban design.

[0047] The following further describes specific embodiments of the present invention and experimental verification.

[0048] An artificial intelligence-based urban design noise optimization method comprises the following steps:

[0049] Data preparation phase: Collect urban building morphology data and noise measurement data, verify the reliability of noise simulation software, and clarify the background model and acoustic parameters required for noise simulation.

[0050] Parametric modeling and constraint processing phase: Establish a hierarchical parametric modeling framework, intelligently generate 3D models and measurement grids based on the building complex organization strategy, implement an intelligent constraint verification mechanism, and provide a structured design space and feasibility guarantee for the optimization process;

[0051] Urban noise prediction model construction and training phase: Data is standardized and feature-engineered to construct a high-quality training dataset. Based on the data grouping results, a two-stage MLP network architecture is constructed. The first stage uses parallel group-specific subnetworks to address the impact of local building clusters. The second stage uses a layered dense network to capture global noise patterns. High-precision noise prediction is achieved through optimized loss function design and intelligent model training.

[0052] Large language model intelligent optimization stage: Build an intelligent optimization agent and design a prompt engineering framework. Through intelligent reasoning and solution generation, iterative optimization and strategy adjustment, convergence judgment and result output, achieve multi-objective design optimization based on natural language understanding and reasoning capabilities.

[0053] In some embodiments, an AI-based urban design noise optimization method and system primarily includes a data preprocessing phase, a parameterized morphology generation phase, a neural network model training phase, and a large language model optimization phase. The large language model optimization phase can be further divided into a problem modeling phase, an evaluation and screening phase, a convergence judgment phase, and an iterative optimization phase.

[0054] The data preprocessing phase includes the following steps:

[0055] Step 1: Collect urban building form data within a certain radius (e.g., 200 meters) around the target plot for design. Building form data includes geometric parameters such as building height, width, length, rotation angle, and location coordinates. This data can generally be obtained from data sources such as GIS, map data, and remote sensing data.

[0056] Step 2: Conduct noise measurement within the building measurement range. The noise measurement data includes the noise values ​​at different measurement points, which are distributed on the building facade, base, and ground, as well as the type of noise source and the equivalent noise value of the noise source. The specific data structure is defined as follows:

[0057] Noise measurement point dataset D N = {(X1,Y1, Z1,N1), (X2,Y2, Z2,N2), ..., (X m ,Y m , Z m ,N m )}, where (X i ,Y i , Z i ,N i ) represents the coordinates of the i-th noise measurement point and the measured noise value.

[0058] Noise source measurement dataset

[0059] If the noise source is a point sound source, the dataset format is:

[0060] D NPS = {(X1,Y1, Z1,NPS1),(X2,Y2, Z2,NPS2), ..., (X n ,Y n , Z n ,NPS n )}, where (X i ,Y i ,Z i ,NPS i ) represents the coordinates of the i-th point noise source and the noise value of the point noise source.

[0061] If the noise source is a line source, the dataset format is:

[0062] D NLS = {(L1, NLS1),(L2, NLS2), ..., (L n ,NLS n )}, where (L i , NLS i) represents the i-th line noise source and the corresponding noise value.

[0063] The parameters of the line sound source are expressed as: L i = {X 1a ,Y 1a , Z 1a ,X 1b ,Y 1b , Z 1b , …, X 1j ,Y 1j , Z 1j}, where (X ij , Y ij , Z ij ) represents the coordinates of the jth node on the i-th line sound source.

[0064] If the noise source is a surface sound source, the dataset format is:

[0065] D NSS = {(S1, NSS1),(S2, NSS2), ..., (S n , NSS n )}, where S i Represents the i-th surface sound source, NSS i Indicates the corresponding noise value.

[0066] The parameters of the surface sound source are expressed as: S i = {X 1a ,Y 1a , Z 1a ,X 1b ,Y 1b , Z 1b , …, X 1j ,Y 1j , Z 1j}, where (X ij , Y ij , Z ij ) represents the coordinates of the jth endpoint of the i-th surface sound source.

[0067] Step 3: Use the measured noise data to verify the reliability of the noise simulation software and determine the parameters required for noise simulation. Input the surrounding urban building form data and noise source data into the simulation software (such as CadnaA, SoundPLAN, Predictor LimA, or other commercial software). Arrange noise simulation points corresponding to the noise measurement points. Set simulation parameters based on data such as building materials, usage, local climate, and noise standards, and conduct noise simulation. Compare the simulation results with the measured results to verify the reliability of the simulation software, and save the background model and parameters required for the simulation.

[0068] The parametric morphology generation phase consists of the following steps:

[0069] Step 4: In urban design, buildings are typically designed in groups, with each group having uniform geometric parameters. After the designer completes the initial building cluster layout, they parametrically represent the spatial relationships of the initially designed building groups on the target plot. They also define the range of variation and constraints for each parameter (such as building coverage ratio and volume ratio) based on urban land use regulations and design requirements.

[0070] The specific data structure is defined as follows:

[0071] Building Form Dataset D B = {B1, B2, ..., B n}, where B i Represents the geometric parameters of the i-th building complex.

[0072] The building complex parameters are expressed as: B i = {BD1,BD2,…,BD n , h i , w i , l i ,θ i}, where BD j is the jth building in the building complex. j For example, (x j , y j , z j ) is the center coordinate of the building bottom surface (m). The unified geometric parameters of all buildings in the building complex are: h i is the building height (m), w i is the building width (m), l i is the building length (m), θ i The counterclockwise rotation angle (degrees) of the building's long side relative to due east.

[0073] Step 5: Establish a parametric modeling framework. Develop a parametric urban morphology generator to generate a 3D building model based on the data structure and parameter variation range of the initial building complex designed by the designer in Step 4. The generator can run independently or be deployed within a 3D design software platform (e.g., using Rhinoceros / Grasshopper to generate the 3D model). Simultaneously, a corresponding grid of noise measurement points, such as facade measurement points, base measurement points, and ground measurement points, is deployed based on the building morphology layout to ensure that the measurement point distribution matches the acoustic analysis requirements. The measurement point deployment strategy utilizes an adaptive grid generation method, dynamically adjusting the measurement point density based on building density and spatial complexity, increasing the number of measurement points in densely built areas and appropriately reducing the number of measurement points in open areas.

[0074] Step 6: Implement constraint processing. This system integrates automatic verification of various planning constraints, including floor area ratio, building coverage ratio, height restrictions, non-overlapping between buildings and roads, building spacing requirements, greening requirements, fire safety requirements, and sunshine hours requirements. A real-time verification mechanism automatically checks each constraint when parameters change, providing timely feedback on constraint violations.

[0075] In step 7, a large number of different building layouts are generated within the parameter variation range. The 3D model of each building layout is input into the noise simulation software. Combined with the background model and parameters saved in step 3, the noise values ​​at designated measurement points in the space (such as the building facade, base, and ground grid) are calculated. The building facade measurement grid is typically set 1 meter from the building surface, with the grid size determined by the building's floor height and the distance between two rooms (or the main windows of a typical two-family residence). The base and ground grids are set 4 meters above the surface, with the plane grid typically measuring 10m x 10m.

[0076] The neural network model training phase includes the following steps:

[0077] Prepare training data:

[0078] Step 8: For each building form layout, standardize the building form data described in step 4. Using the measurement point center coordinate system, convert all building coordinates into relative coordinates with the measurement point as the origin to eliminate the influence of absolute position.

[0079] For the measurement point P = (x, y,z), the building complex B i The jth building BD j The relative coordinates are converted to:

[0080] x' j =x p -x j

[0081] y' j =y p -y j

[0082] z' j =z p -z j

[0083] Its rotation angle θ i Converted to trigonometric function, sin(θ i ) and cos(θ i )

[0084] In step 9, the input features are Z-score normalized. Normalization parameters are calculated for different types of features. The target noise value remains in the original decibel unit without normalization, so that the model can directly learn to predict the true noise value. The normalization parameters are stored in HDF5 format files.

[0085] Coordinate feature normalization:

[0086] x' i = (x i - μ x ) / σ x

[0087] y' i = (y i - μ y ) / σ y

[0088] z' i = (z i - μ z ) / σ z

[0089] Geometric feature standardization:

[0090] h' i = (h i - μ h ) / σ h

[0091] w' i = (w i - μ w ) / σ w

[0092] l' i = (l i - μ l ) / σ l

[0093] Where μ and σ represent the mean and standard deviation of each parameter respectively

[0094] Step 10: Construct a training dataset. The standardized building morphology parameters are used as input features, and the noise simulation values ​​of the corresponding measurement points are used as target outputs to form a dataset for neural network training.

[0095] For the measurement point P, the data format is:

[0096] N R = {B'1,B'2,…B' n , N p} where B' i B' is the relative coordinate of the morphological layout of building group i relative to point P. i ={(x' 11 , y' 11 , z' 11 ),(x' 12 , y' 12 , z' 12 ),…,(x' ij , y' ij , z' ij ), h' i , w' i , l' i , sin(θ i ), cos(θ i )}, N p is the simulated noise value corresponding to point P.

[0097] Training the neural network:

[0098] Step 11: Construct a two-stage multilayer perceptron neural network architecture. This network architecture design fully considers the dual characteristics of urban acoustics, namely the influence of local building clusters and global propagation patterns. The first stage utilizes parallel cluster-specific subnetworks, each dedicated to processing the acoustic impact of a specific building cluster on the measurement point. The second stage utilizes a layered dense network to process the global urban noise pattern and integrate the local contributions of all clusters.

[0099] The first stage network architecture definition: For an input containing G building clusters, the first stage contains G parallel subnetworks: SubNet_dg: R dg →R 32 , where dg is the characteristic dimension of the g-th building complex.

[0100] An example of a single sub-network structure, where the number of network layers and neurons is adjusted according to the complexity of the data:

[0101] Layer 1: Dense(128) + ReLU + BatchNorm + Dropout + L2 regularization

[0102] Layer 2: Dense(64) + ReLU + BatchNorm +Dropout +L2 regularization

[0103] Layer 3: Dense(32) + ReLU

[0104] Group feature extraction formula: f g = SubNet_g(B g ), F1 = Concat([f 1 , f 2 , ..., f G ]), where F1∈R 32G This is the output of the first stage.

[0105] Second stage network architecture definition: DenseNet: R 32G →R 1

[0106] Layered dense network structure, the number of network layers and neurons is adjusted according to the complexity of the data:

[0107] Layer 1: Dense(512) + ReLU + BatchNorm + Dropout

[0108] Layer 2: Dense(256) + ReLU + BatchNorm + Dropout

[0109] Layer 3: Dense(128) + ReLU + BatchNorm + Dropout

[0110] Layer 4: Dense(64) + ReLU + BatchNorm + Dropout

[0111] Layer 5: Dense(32) + ReLU + BatchNorm + Dropout

[0112] Layer 6: Dense(1)

[0113] Final noise prediction formula: N_pred = DenseNet(F1)

[0114] Step 12: To address the data imbalance problem in which some noise intervals in the training data have fewer samples, a weighted mean square error loss function is designed to assign higher weights to the noise intervals with low sample sizes, thereby ensuring the prediction accuracy of the model within the entire noise range.

[0115] Mathematical definition of weighted loss function:

[0116] L_weighted_mse = (1 / n) ×Σ i =1 n [w i × (y i _true - y i _pred) 2 ]

[0117] Weight allocation rules (taking the noise range <55.0 dB with a small sample size as an example):

[0118] w i = {

[0119] 2.5, if y i _true< 55.0 dB (low noise sample)

[0120] 1.0, if y i _true≥ 55.0 dB (high noise sample)}

[0121] in:

[0122] n is the number of samples in the batch

[0123] y i _true is the true noise value of the i-th sample (dB)

[0124] y i _pred is the predicted noise value of the i-th sample (dB)

[0125] w i is the weight coefficient of the i-th sample

[0126] Threshold = 55.0 dB

[0127] Low noise weight low_weight=2.5

[0128] High noise weight high_weight=1.0

[0129] Loss function implementation algorithm:

[0130] Algorithm: WeightedMSELoss

[0131] Input: y_true, y_pred, threshold=55.0, low_weight=2.5, high_weight=1.0

[0132] Output: weighted_loss

[0133] 1. Calculate the squared error: squared_error = (y_true - y_pred) 2

[0134] 2. Create a weight mask:

[0135] weights = {low_weight if y_true[i] < threshold else high_weight for iin range(n)}

[0136] 3. Apply weights: weighted_squared_error = weights × squared_error

[0137] 4. Calculate the average loss: weighted_loss = mean(weighted_squared_error)

[0138] 5. Return weighted_loss

[0139] Step 13: Use the Adam optimizer with a learning rate of 1×10 -3 , and adopts a dynamic learning rate decay strategy. An early stopping mechanism is used during training, stopping training when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting.

[0140] The large language model optimization phase includes the following steps:

[0141] Local deployment of large language models

[0142] Step 14: Build a large language model as the core intelligent optimization engine. Prepare a single high-performance server, install the deep learning framework, deploy an inference framework (such as Ollama or vLLM), load and configure the target large language model (such as Llama, Qwen, Deepseek, etc.), and test the model's inference performance and response speed.

[0143] Step 15: Build the Mixed-of-Experts (MoE) architecture. Design the MoE architecture and define the number and division of labor of the expert networks. For example, define experts in urban planning, architectural design, urban acoustics, and economics. Train or fine-tune each expert module, implement an expert routing mechanism, and automatically select experts based on task type. Integrate the expert networks into the main model.

[0144] Step 16: Build a professional knowledge base. Collect urban planning-related regulations, standards, and case studies; organize architectural design specifications, technical manuals, and design cases; compile urban acoustic environment data, simulation models, and evaluation criteria; preprocess and vectorize knowledge documents; construct a vector database (such as Chroma or Pinecone); and establish a knowledge indexing and retrieval mechanism.

[0145] Step 17: Integrate the RAG system. Install and configure the RAG framework (such as LangChain and LlamaIndex), connect the vector database to the LLM engine, implement semantic search capabilities, match relevant knowledge based on the query, design prompt word templates, integrate search results into the generation process, optimize the search strategy and relevance ranking algorithm, and test the accuracy of knowledge retrieval and generation quality.

[0146] In the intelligent optimization phase of the solution:

[0147] Step 18 randomly generates a set of initial design solutions within the range of building parameters. Leveraging the creative generative capabilities of the Large Language Model (LLM), a diverse set of initial design solutions is generated based on the problem description and constraints. Leveraging the innovative generative capabilities of the Large Language Model (LLM), the model draws on its extensive knowledge in urban planning, architectural design, acoustic engineering, and other fields, based on the problem description, while strictly adhering to planning constraints, to generate a set of initial solutions that meet basic constraints and exhibit optimization potential. Using core planning indicators such as floor area ratio, building height, and building coverage as primary constraints, the system automatically generates 100 design solutions that meet regulatory requirements through algorithmic optimization and creative combination. Each solution, while meeting statutory constraints, embodies distinct design concepts and spatial layout strategies, providing a rich initial sample space for subsequent optimization analysis and ensuring the comprehensiveness and feasibility of design exploration.

[0148] Example of prompt words:

[0149] You are a professional urban planning designer. Please generate 100 initial urban design plans based on the following basic constraints:

[0150] Basic restrictions:

[0151] - Floor area ratio range: [specific value, such as 2.5-4.5]

[0152] - Building height range: [specific value, such as 80 meters to 120 meters]

[0153] - Building coverage range: [specific value, such as 20%-40%]

[0154] - Green space ratio minimum: [specific value, such as 30%]

[0155] - Land area: [specific value, such as 10 hectares]

[0156] - Other constraints: [road setbacks, firefighting distances, etc.]

[0157] Design requirements:

[0158] 1. Strictly comply with all basic restrictions

[0159] 2. Generate 100 different design solutions

[0160] 3. The proposal should reflect diverse design concepts:

[0161] - High-density compact layout

[0162] - Low-density dispersed layout

[0163] - Mixed-function cluster layout

[0164] - Linear development model

[0165] - Enclosed courtyard layout

[0166] - Modern minimalist style

[0167] - Fusion of traditional culture

[0168] - Eco-first design

[0169] Variation range of building parameters:

[0170] - The range of variation of each parameter that affects the building's form and layout

[0171] - Building Type: Residential / Commercial / Office / Mixed

[0172] - Layout type: point type / plate type / enclosed type / group type

[0173] Output format:

[0174] Project number: [S001-S100]

[0175] Design concept: [Briefly describe the design theme]

[0176] Key parameters:

[0177] - Floor Area Ratio: [value]

[0178] - Building Height: [value]

[0179] - Construction density: [value]

[0180] - Green space ratio: [value]

[0181] Layout features: [Spatial organization]

[0182] Functional configuration: [Proportion of each functional area]

[0183] Please ensure that each scenario meets the constraints and is unique.

[0184] Step 19: Perform a performance evaluation of the proposed scheme. The generated design parameters are input into the trained neural network model to predict noise levels. Other performance indicators are calculated simultaneously, and constraints are determined to determine if any are violated. Proposals that violate these constraints are removed. Performance evaluation includes not only noise level prediction but also a comprehensive assessment of multiple metrics, such as building density and economic benefits. Multiple performance evaluation parameters are output, forming a complete scheme evaluation system.

[0185] Example: Solution performance evaluation function reference definition: F(x) = {f1(x), f2(x), f3(x)}

[0186] Where f1(x): noise level, f2(x): building density, f3(x): economic benefits

[0187] Noise level function f1(x): f1(x) = Σ i w i · L i (x)

[0188] Where: L i (x) = noise level at the ith measurement point (dB), w i = Weight coefficient of the i-th point

[0189] Building density function f2(x): f2(x) = FAR_target- FAR(x)

[0190] Where: FAR(x) = actual floor area ratio, FAR_target = target floor area ratio

[0191] Economic benefit objective function f3(x): f3(x) = Revenue(x) - Cost(x)

[0192] Revenue (x) = Project revenue = Building area × unit price × utilization rate

[0193] Cost(x) = Construction cost + Maintenance cost + Environmental governance cost

[0194] Step 20: Multi-objective optimization screening. Leveraging the powerful analytical capabilities of large language models, combined with appropriate prompting, the system conducts a comprehensive, multi-dimensional evaluation of numerous urban design proposals. The system comprehensively considers key indicators such as noise control, economic costs, and building density, using intelligent algorithms to quantitatively score and rank each proposal. Ultimately, the optimal set of proposals that meet the criteria is selected from a large pool of candidate proposals.

[0195] Example of prompt words:

[0196] You are a senior urban planning expert. Please conduct a comprehensive evaluation and analysis of the following urban design plans and select the 10 best ones:

[0197] Evaluation dimensions:

[0198] 1. Noise control (40% weight): Evaluate the effectiveness of the scheme in controlling urban noise

[0199] 2. Economic efficiency (40% weight): Analyze construction costs, maintenance costs, and return on investment

[0200] 3. Building density (weight 20%): Check whether it complies with relevant regulations and planning standards

[0201] Evaluation requirements:

[0202] - Give a score of 1-10 for each dimension and explain your reasons

[0203] - Calculate weighted total score

[0204] - Identify the strengths and weaknesses of the solution

[0205] - Make suggestions for improvement

[0206] - Finally give a conclusion on whether to recommend

[0207] Output format:

[0208] Project Name: [Project Name]

[0209] Ratings for each dimension: [Detailed ratings and reasons]

[0210] Weighted total score: [XX points]

[0211] Comprehensive evaluation: [Advantages, Disadvantages, Suggestions]

[0212] Recommended level: [A / B / C / D]

[0213] Please provide a professional evaluation of the urban design proposals provided based on the above criteria.

[0214] Step 21: Determine convergence. The Large Language Model (LLM) intelligently determines the convergence of the urban planning model optimization process. By analyzing the magnitude of changes in the design variable vectors between the current and previous iterations, the system accurately identifies whether the optimization algorithm has reached a stable state. When the change in variables between consecutive iterations is less than a preset threshold, the system determines that the optimization process has converged, avoiding ineffective waste of computing resources.

[0215] Example of prompt words:

[0216] You are a professional optimization algorithm analyst. Please use the following information to determine whether the urban planning model optimization process has converged:

[0217] Input data:

[0218] - Current round design variable vector: x_current = [X1, X2, ..., Xn]

[0219] - Previous round design variable vector: x_previous = [Y1, Y2, ..., Yn]

[0220] - Convergence threshold: ε = [threshold value]

[0221] Judging criteria:

[0222] 1. Calculate the change in the variable vector:

[0223]

[0224] 2. Compare the change with the threshold: Δ<ε

[0225] 3. Analyze change trends and stability

[0226] Output format:

[0227] Calculation of change:

[0228]

[0229] Threshold comparison: C ≥ ε means non-convergence, otherwise it means convergence

[0230] Convergence judgment: [converged / not converged]

[0231] Please conduct professional analysis of the data provided based on the above criteria.

[0232] Step 22: Generate Improvement Plans. Starting with the 10 selected basic urban design solutions, the system conducts in-depth analysis of noise simulation results, economic assessments, and building density data. Using intelligent algorithms, the system identifies optimization potential and improvement directions for each solution, automatically generating 90 innovative and improved solutions. These new solutions, while maintaining the original strengths, specifically address key issues such as noise control, cost optimization, and regulatory compliance. The final output is a set of 100 complete solutions, including the original 10.

[0233] Example of prompt words:

[0234] You are an experienced urban design expert. Please generate an improved urban design plan based on the following basic plan and analysis results:

[0235] Input information:

[0236] - Number of basic plans: 10

[0237] - Noise simulation results: [Noise distribution data for each solution]

[0238] - Economic analysis: [construction cost, maintenance cost, return on investment, etc.]

[0239] - Compliance analysis: [regulatory compliance, planning standards matching, etc.]

[0240] Build requirements:

[0241] 1. Generate 90 improved versions of the basic solution

[0242] 2. Improvements include but are not limited to:

[0243] - Noise control optimization (noise reduction facility layout, building material selection)

[0244] - Improved economic efficiency (cost control, maximum benefit)

[0245] - Improve compliance (regulatory adaptation, standard optimization)

[0246] - Functional layout optimization (spatial configuration, traffic organization)

[0247] - Enhanced environmental friendliness (greening configuration, ecological protection)

[0248] - Exchange, copy, mirror, and rotate components between different solutions

[0249] Design principles:

[0250] - Maintain the core concepts and advantages of the original solution

[0251] - Targeted solutions to identified problems and deficiencies

[0252] - Ensure the feasibility and practicality of improvement plans

[0253] - Consider the combined application of different improvement strategies

[0254] - Balance various indicators to avoid neglecting one while focusing on another

[0255] Output format:

[0256] Original proposal number: [A01-A10]

[0257] Improvement plan number: [B01-B100]

[0258] Key points for improvement: [Specific improvement measures]

[0259] Expected results: [Improvements in noise, economy, compliance, etc.]

[0260] Design parameters: [Adjustment of key design variables]

[0261] Please generate 90 innovative and practical improvement plans for the basic plan.

[0262] Step 23: Perform multiple rounds of iterative optimization. Repeat steps 19-22 to continuously refine the design through multiple rounds of iteration. In each round, the large language model learns from the previous round's optimization experience, adjusts its optimization strategy, and improves search efficiency. During the iteration process, a library of optimal solutions is maintained to record the best solutions.

[0263] Step 24 verifies the stability and consistency of the optimization results through multiple independent runs. Due to the randomness inherent in large language models, multiple runs are necessary to ensure the repeatability of the optimal solution. Typically, 3-5 independent optimization runs are performed, with the optimal solution selected as the final solution. The optimal design solution is output, including complete information such as detailed building parameter configuration, expected noise performance, and constraint satisfaction. A detailed log of the optimization process is also generated, providing designers with an optimization trajectory and decision-making basis.

[0264] Experimental results

[0265] In the experimental test, the method uses a two-stage MLP network architecture constructed by combining a data structure with the measurement point as the origin. The structure conforms to the acoustic propagation principle in complex urban spaces and provides a technical basis for efficient design optimization. In comparison with the simulation results of the industrial noise simulation software CadnaA, R 2 = 0.93, RMSE = 1.18 dB, which is a high-precision prediction that exceeds the accuracy requirement of a Class 1 sound level meter.

[0266] This method achieves an order of magnitude improvement in the efficiency of AI prediction models compared to traditional simulation methods, reducing the time for a single prediction from hours to seconds. In a test case, this method used MLP to simulate noise, with each solution taking 3-5 seconds to calculate and running automatically. Traditional methods, using CadnaA simulation software with simple environmental parameter settings, take around 30 minutes to simulate. CadnaA simulation time increases with increasing spatial complexity. Furthermore, the software's lack of an automated interface requires significant additional labor costs and time for large-scale simulations.

[0267] Furthermore, this method utilizes a designed large language model optimization strategy, applying the powerful reasoning capabilities of large language models to urban design optimization for the first time, achieving an intelligent optimization process. Compared to traditional evolutionary algorithms, which require extensive parameter tuning and lack interpretability, large language model optimization offers significant advantages, including strong adaptability, excellent interpretability, integration of domain knowledge, and support for natural language interaction. Furthermore, compared to traditional genetic / evolutionary algorithms, large language model optimization boasts greater intelligence and adaptability, automatically adjusting optimization strategies based on problem characteristics, significantly improving optimization efficiency and effectiveness.

[0268] In summary, the present invention provides an artificial intelligence-based urban design noise optimization method and system. Compared with the existing technology, the present invention has the following significant advantages:

[0269] Traditional technologies are unable to systematically consider noise mitigation in urban design. Traditional noise simulation methods are inefficient and lack integration with the design process. They are typically evaluated after the design is completed, failing to provide scientific design guidance for designers. The method proposed in this paper integrates advanced technologies such as deep learning, large language models, and parametric modeling to systematically embed noise mitigation factors in the early stages of urban design. Its core modules include a parametric urban form generation and constraint verification module, a two-stage multi-layer perceptron (MLP) neural network prediction module, and a large language model intelligent optimization module. Together, these three modules form a complete AI-enhanced urban design optimization technology system. Through four main stages: data preprocessing, neural network training, parametric modeling, and large language model intelligent optimization, it implements a complete technical chain from raw data input to intelligent optimized design solution output, achieving a technological breakthrough in urban design optimization. The method of the present invention achieves high-precision urban noise prediction by leveraging a two-stage multi-layer perceptron network architecture combined with a measurement-point-centered data structure. Through an intelligent large language model optimization strategy (including building an intelligent optimization agent, designing a prompt engineering framework, implementing intelligent reasoning optimization, and executing an iterative improvement process), and relying on natural language understanding and reasoning capabilities, it achieves efficient multi-objective design optimization. By organically combining the prediction model with parametric modeling and the large language model, it achieves noise reduction design optimization and efficient building form layout optimization without the need for traditional simulation software or the input of professional acousticians. This systematically reduces noise hazards in the early stages of urban design, addressing the technical shortcomings of traditional methods, such as poor computational efficiency and lack of system integration.

[0270] This invention, for the first time, intelligently integrates parametric morphology generation, noise prediction, and solution optimization throughout the entire urban design optimization process, from problem understanding and solution generation to performance evaluation, optimization reasoning, and result interpretation. This intelligent process can be integrated into a design software platform or run independently (in which case, its efficiency is not limited by the design software platform), opening up a new technical path for the application of AI in urban design.

[0271] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0272] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0273] An embodiment of the present invention further provides a processor, which executes a computer program and at least performs the method described above.

[0274] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory (Flash Memory), a magnetic surface memory, an optical disc or a read-only optical disc (CD-ROM); the magnetic surface memory can be a magnetic disk memory or a magnetic tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0275] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0276] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0277] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0278] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0279] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0280] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0281] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0282] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0283] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art will recognize that, without departing from the scope of the present invention, several equivalent substitutions or obvious variations can be made, and the performance or use of the same should be considered to fall within the scope of protection of the present invention.

Claims

1. An urban design noise optimization method based on artificial intelligence, characterized in that: The following steps are involved: S1. Data preprocessing: Collect building form data and noise measurement data, verify the data through noise simulation to determine the background model and acoustic parameters, and build a basic data set; S2. Parametric Modeling and Constraint Processing: A hierarchical parametric framework is established to generate a 3D city model and a network of measurement points. The noise level of the 3D city model is calculated using the measurement point network. A real-time verification mechanism is implemented using multiple planning constraints on the 3D city model. This provides a structured design space, basic noise level data, and feasibility assurance for urban design noise optimization. S3. Construction and training of urban noise prediction model: Standardization and feature engineering of the 3D urban model data generated in step S2 are performed to construct a training dataset; A two-stage multilayer perceptron network is constructed based on the training dataset. The first stage of the perceptron network processes the acoustic impact of local building clusters through parallel group-specific subnetworks. The second stage of the perceptron network captures the global urban noise pattern through a layered dense network. The model is trained by optimizing the training strategy to achieve high-precision noise prediction. In step S3: The three-dimensional city model data generated in step S2 is converted into the relative coordinates of the building using the measurement point center coordinate system, and the building rotation angle is converted into a trigonometric function; Perform Z-score normalization on the converted coordinates and building geometric features to construct a training dataset; A two-stage multilayer perceptron network was constructed based on the training dataset. In the first stage, the perceptron network independently processed the acoustic characteristics of each building group through parallel group-specific sub-networks. In the second stage, the perceptron network fused the output features of all groups through a layered dense network to capture the global urban noise pattern. By using a weighted loss function and a dynamic weight allocation strategy, higher weights are assigned to data groups in a predetermined low sample size range to balance the data distribution; Use dynamic learning rate decay strategy with early stopping mechanism for model training; S4. Large language model optimization: Deploy the large language model as an optimization agent. Based on the 3D city model and noise level generated in step S2, generate an initial set of design solutions through the prompt engineering framework. Combined with the urban noise prediction model constructed in step S3, perform multi-objective performance evaluation, and output the optimal solution after multiple rounds of iterative optimization and convergence judgment. The multi-objective performance evaluation includes: inputting solution parameters into a prediction model to obtain multi-location noise distribution data; calculating comprehensive indicators of noise control, economic benefits and regulatory compliance; and performing weighted scoring, ranking and constraint violation screening through a large language model.

2. The method according to claim 1, wherein Step S1 specifically includes: Collect data on building geometry and different types of noise sources; Verify the reliability of acoustic parameters through noise simulation and save the background model and acoustic parameters; Verify the collected data through noise simulation software, confirm the reliability of acoustic parameters and save the background model and acoustic parameters; Construct a basic dataset containing building form data, noise measurement data and verified acoustic parameters.

3. The method according to any one of claims 1 to 2, characterized in that Step S2 specifically includes: Establish a parametric generator to construct a 3D city model in a 3D design platform; Deploy an adaptive grid of measurement points matched to acoustic analysis, dynamically adjusting the grid resolution based on building density and spatial complexity; Through multiple types of planning constraints, real-time automatic verification of parameter changes in the 3D urban model is implemented to ensure design feasibility.

4. The method according to any one of claims 1 to 2, characterized in that In step S4: The large language model deployment adopts a hybrid expert architecture, integrating expert modules in urban planning, acoustic environment, and economics. Calling vectorized expertise base through retrieval-enhanced generation mechanism; Prompt engineering is used to generate a diverse set of initial solutions that meet floor area ratio, height and density constraints.

5. The method according to any one of claims 1 to 2, characterized in that The multiple rounds of iterative optimization in step S4 include: Identify solution optimization directions based on performance evaluation results; Generate improved solution sets through solution component exchange and geometric transformation; Maintain a database of historical optimal solutions and dynamically update engineering strategies.

6. The method according to any one of claims 1 to 2, characterized in that The convergence judgment in step S4 includes: Calculate the Euclidean distance change of the design variable vector in successive iterations; Convergence is determined when the change is lower than a preset threshold; The reproducibility of the results was verified through multiple independent optimizations.

7. The method according to any one of claims 1 to 2, characterized in that The final output of step S4 includes: The optimal architectural parameters verified through multiple rounds of iterations; Multi-objective performance indicator evaluation report and constraint satisfaction status; A complete decision record containing the optimization trajectory of the solution.

8. An artificial intelligence-based urban design noise optimization system, characterized by: include: Data preprocessing module: collects building form data and noise measurement data, verifies the data through noise simulation to determine the background model and acoustic parameters, and constructs a basic data set; Parametric Modeling and Constraint Processing Module: This module establishes a hierarchical parametric framework to generate a 3D city model and a network of measurement points. The noise level of the 3D city model is calculated using the network of measurement points. A real-time verification mechanism is implemented using various planning constraints on the 3D city model. This module provides a structured design space, basic noise level data, and feasibility assurance for noise optimization in urban design. Urban noise prediction model construction and training module: standardizes and performs feature engineering on the generated 3D urban model data to construct a training dataset; A two-stage multilayer perceptron network is constructed based on the training dataset. The first stage of the perceptron network processes the acoustic impact of local building clusters through parallel group-specific subnetworks. The second stage of the perceptron network captures the global urban noise pattern through a layered dense network. The model is trained by optimizing the training strategy to achieve high-precision noise prediction. in: The generated 3D city model data is converted into the relative coordinates of the building using the measurement point center coordinate system, and the building rotation angle is converted into a trigonometric function; Perform Z-score normalization on the converted coordinates and building geometric features to construct a training dataset; A two-stage multilayer perceptron network was constructed based on the training dataset. In the first stage, the perceptron network independently processed the acoustic characteristics of each building group through parallel group-specific sub-networks. In the second stage, the perceptron network fused the output features of all groups through a layered dense network to capture the global urban noise pattern. By using a weighted loss function and a dynamic weight allocation strategy, higher weights are assigned to data groups in a predetermined low sample size range to balance the data distribution; Use dynamic learning rate decay strategy with early stopping mechanism for model training; Large language model optimization: Deploy a large language model as an optimization agent. Based on the generated 3D city model and noise levels, an initial set of design solutions is generated through a prompt engineering framework. Combined with the constructed urban noise prediction model, a multi-objective performance evaluation is conducted. After multiple rounds of iterative optimization and convergence judgment, the optimal solution is output. The multi-objective performance evaluation includes: inputting solution parameters into the prediction model to obtain multi-location noise distribution data; calculating comprehensive indicators of noise control, economic benefits, and regulatory compliance; and performing weighted scoring, ranking, and constraint violation screening through the large language model.

Images