Predicting Urban Surface Roughness Aerodynamic Parameters Using Random Forest
Predicting Urban Surface Roughness Aerodynamic Parameters Using Random Forest
复制标题
DOI：
10.1175/jamc-d-20-0266.1
复制
发表时间：
2021-05
影响因子：
3
通讯作者：
G. Duan;T. Takemi
中科院分区：
文献类型：
--
作者：
G. Duan;T. Takemi
表面粗糙度气动参数z 0（粗糙度长度）和d（零平面位移高度）对Monin-Obukhov相似理论的精度至关重要。在传统框架内推导改进的城市冠层参数化（UCP）方案在数学上仍然具有挑战性。目前的研究探讨了机器学习（ML）算法的潜力，随机森林（RF），作为传统的UCP计划的补充。采用大涡模拟和集合采样，结合对流层风廓线的非线性最小二乘回归，建立了约4.5 × 103个样本的空气动力学参数和形态统计数据集，为ML模式的训练提供了条件。虽然对d的预测不如Kanda等人之后的UCP好，Z 0的性能是显著的。RF算法还将z 0和d分类为具有出色性能评分的类别：总体钟形分布得到了很好的预测，并且±0.5σ类别（即，38%百分位数）被充分捕获（对于z 0为37.8%，对于d为36.5%）。在形态特征中，平均和最大建筑物高度（分别为Have和Hmax）对z 0和d的预测有显著影响。一个可能违反直觉的结果是，建筑物高度可变性的重要性要小得多。可能的原因进行了讨论。特征重要性分数可用于识别表面空气动力特性的影响因素。本文的研究结果对中尺度数值模拟中基于ML的UCP的发展具有一定的指导意义。
The surface roughness aerodynamic parameters z0 (roughness length) and d (zero-plane displacement height) are vital to the accuracy of the Monin–Obukhov similarity theory. Deriving improved urban canopy parameterization (UCP) schemes within the conventional framework remains mathematically challenging. The current study explores the potential of a machine-learning (ML) algorithm, a random forest (RF), as a complement to the traditional UCP schemes. Using large-eddy simulation and ensemble sampling, in combination with nonlinear least squares regression of the logarithmic-layer wind profiles, a dataset of approximately 4.5 × 103 samples is established for the aerodynamic parameters and the morphometric statistics, enabling the training of the ML model. While the prediction for d is not as good as the UCP after Kanda et al., the performance for z0 is notable. The RF algorithm also categorizes z0 and d with an exceptional performance score: the overall bell-shaped distributions are well predicted, and the ±0.5σ category (i.e., the 38% percentile) is competently captured (37.8% for z0 and 36.5% for d). Among the morphometric features, the mean and maximum building heights (Have and Hmax, respectively) are found to be of predominant influence on the prediction of z0 and d. A perhaps counterintuitive result is the considerably less striking importance of the building-height variability. Possible reasons are discussed. The feature importance scores could be useful for identifying the contributing factors to the surface aerodynamic characteristics. The results may shed some light on the development of ML-based UCP for mesoscale modeling.