Urban Flow Field Prediction with RANS-to-LES CNN Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods struggle to accurately and efficiently predict wind speed distributions in urban areas with multiple buildings, and existing neural network approaches lack sufficient accuracy.
Innovation Solution
A flow field prediction device utilizing a Convolutional Neural Network (CNN) that integrates analysis results from a first computation method like RANS with more accurate results from a second method like LES, enhancing prediction accuracy through supervised learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CFD (Computational Fluid Dynamics) is used to analyze wind environment, then analysis accuracy is improved, but computation time increases significantly
Solution Approach 1:
The system performs preliminary action by pre-computing high-accuracy LES results for various building configurations and storing them as teaching data. During actual prediction, the neural network retrieves and processes this pre-computed data, avoiding the need to perform time-consuming LES computations in real-time while maintaining high prediction accuracy.
Solution Approach 2:
The system creates a copying approach by using neural networks to learn the relationship between simplified input data (RANS results) and high-accuracy reference data (LES results). The network copies the complex physical relationships encoded in LES computations into a faster predictive model, allowing rapid inference without re-executing the expensive CFD simulations.
2Loss of time
If simple analysis methods or neural networks are used, then computation time is reduced, but prediction accuracy is insufficient
Solution Approach 1:
The system introduces an intermediary approach by using RANS (Reynolds-Averaged Navier-Stokes) simulations as an intermediate computation level. RANS provides sufficiently detailed flow field information at much lower computational cost than LES. The neural network then acts as a mediator that translates these intermediate RANS results into accurate predictions by learning the mapping relationship from teaching data, achieving both speed and accuracy.
Solution Approach 2:
The system applies parameter changes by transforming the input data from raw RANS flow field parameters into processed features that capture essential flow characteristics. The neural network learns to map these transformed parameters to accurate LES-like predictions, effectively changing the representation of physical parameters to bridge the gap between computational efficiency and prediction accuracy.
3Loss of time
If neural networks are used to predict wind speed distributions, then computation time is reduced, but accuracy for city blocks with multiple buildings is insufficient
Solution Approach 1:
The system segments the complex city block flow field problem into manageable components by processing plane data at different heights separately. The neural network predicts wind speed distributions for each height level independently, then combines these segmented predictions to produce the final three-dimensional flow field profile, improving accuracy for complex urban environments.
Solution Approach 2:
The system applies dimensionality change by transitioning from two-dimensional plan view analysis to three-dimensional vertical profile prediction. The neural network processes horizontal plane data and outputs vertical profiles of wind speed distributions, adding the height dimension to capture the complex three-dimensional flow patterns that occur in city blocks with multiple buildings of varying heights.
Data Source
AI summary
A flow field prediction device includes: an analysis result acquisition unit that acquires an analysis result based on a first computation method of analyzing at least an influence between buildings with respect to a flow field in which a city block including a plurality of buildings is set as a target region; and a prediction processing unit that inputs a new analysis result acquired by the analysis result acquisition unit to a learning model trained using teaching data that includes an analysis result based on a second computation method that is more accurate than the first computation method, as correct answer data for input data including at least the analysis result based on the first computation method, and outputs an output from the learning model as a prediction result of the flow field.