Scholarlink

<p>Simulation of urban wind environments is essential for urban planning, pollution mitigation, and renewable energy applications. However, the high computational cost of high-fidelity computational fluid dynamics (CFD) methods limits their practical deployment in real urban contexts. To overcome this challenge, we propose a Fourier Neural Operator (FNO) model for predicting urban wind fields across different wind directions and city layouts, trained on velocity data generated by large-eddy simulations (LES) with the CityFFD solver. The results demonstrate that the FNO achieves accuracy comparable to CFD while reducing the per-frame wall-clock time from 2.210 s with CityFFD to 0.006 s on an NVIDIA V100 GPU, corresponding to an approximately 370x speedup. To further mitigate GPU memory constraints, a patch-based training strategy is introduced, which partitions the wind field into smaller spatial blocks, enabling the FNO to capture localized flow dynamics more effectively. In addition, incorporating signed distance function (SDF) data provides critical building-geometry information, thereby improving boundary recognition, enhancing prediction realism, and strengthening overall model generalizability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Urban wind field prediction using Fourier Neural Operators across different wind directions and cities

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Cheng Chen,
  • Geng Tian,
  • Shaoxiang Qin,
  • Senwen Yang,
  • Dingyang Geng,
  • Dongxue Zhan,
  • Jinqiu Yang,
  • David Vidal,
  • Liangzhu Leon Wang

摘要

Simulation of urban wind environments is essential for urban planning, pollution mitigation, and renewable energy applications. However, the high computational cost of high-fidelity computational fluid dynamics (CFD) methods limits their practical deployment in real urban contexts. To overcome this challenge, we propose a Fourier Neural Operator (FNO) model for predicting urban wind fields across different wind directions and city layouts, trained on velocity data generated by large-eddy simulations (LES) with the CityFFD solver. The results demonstrate that the FNO achieves accuracy comparable to CFD while reducing the per-frame wall-clock time from 2.210 s with CityFFD to 0.006 s on an NVIDIA V100 GPU, corresponding to an approximately 370x speedup. To further mitigate GPU memory constraints, a patch-based training strategy is introduced, which partitions the wind field into smaller spatial blocks, enabling the FNO to capture localized flow dynamics more effectively. In addition, incorporating signed distance function (SDF) data provides critical building-geometry information, thereby improving boundary recognition, enhancing prediction realism, and strengthening overall model generalizability.