| 摘要: |
| 台风最大风速半径（RMW）是精准预测风暴增水的关键参数，目前文献中最大风速半径的经验公式种类繁多，存在地域适用性。本研究采用中国气象局热带气旋最佳路径数据集(CMA-BST Dataset)，分区域、分强度对西太平洋地区常用的9个RMW经验公式精度进行了评估。结果显示，不同公式精度差异显著，具有明显区域与强度依赖性。针对传统经验公式难以刻画多因子非线性耦合关系的缺陷，本研究构建多任务物理增强神经网络（MT-PINN）、LightGBM、随机森林三种机器学习模型，实现RMW与近中心最大风速（Vmax）协同预测。结果显示，多任务物理增强神经网络模型预测精度最高，RMW预测RMSE为3.00 km，MAPE为5.35%，R²为0.85，Bias为-0.11 km，物理一致性与泛化能力显著提升。 |
| 关键词: 最大风速半径 机器学习 台风风场模型 经验公式 台风预测 |
| DOI： |
| 分类号:P732 ??????????????????????????? |
| 基金项目:国家自然科学基金青年科学(C类)(No.52501348)、南京水利科学研究院中央级公益性科研院所基本科研业务费专项资金（Y225004） |
|
| Research on Calculation Method of Tropical Cyclone Radius of Maximum Wind Based on Multi-Task Physics-Enhanced Neural Network |
|
Yao Hongbin ¹, Zhang Weisheng ¹, Zhang Jinshan ¹, Yin Chengtuan ¹, Xiong Mengjie
|
|
The National Key Laboratory of Water Disaster Prevention,Nanjing Hydraulic Research Institute
|
| Abstract: |
| The radius of maximum wind (RMW) is a critical parameter in tropical cyclone (TC) wind field modeling and storm surge prediction, as it determines the radial location of peak wind speeds and strongly governs the spatial structure of the cyclone wind and pressure fields. However, existing empirical formulas for RMW estimation typically rely on linear or simple nonlinear relationships involving only one to three predictors—such as central pressure, latitude, and translation speed—and consequently suffer from pronounced regional dependence, insufficient stability, and limited generalization capability. Using the China Meteorological Administration Tropical Cyclone Best-Track Dataset (CMA-BST Dataset) from 1981 to 2020, a total of 21,669 valid tropical cyclone records were obtained after excluding tropical depressions and post-landfall samples. Nine widely used RMW empirical formulas, representing three predictor-combination categories, were systematically evaluated across four sub-regions (North China, East China, South China, and open ocean) and three intensity levels in the western North Pacific. The evaluation reveals considerable performance disparities: the Fang Genshen formula, which incorporates both central pressure and latitude, achieves the best overall performance (RMSE = 8.12 km, MAPE = 15.05%, R² = 0.63, Bias = ?1.74 km), while the Kato formula performs the worst (RMSE = 77.58 km, MAPE = 163.10%); the Lin Wei formula ranks second (RMSE = 8.40 km, MAPE = 15.88%, R² = 0.60). Regionally, Fang Genshen and Lin Wei maintain RMSE ≤ 8.51 km and R² ≥ 0.60 across all four sub-regions, with latitude-dependent formulas exhibiting markedly better adaptability at high latitudes where the Coriolis force is stronger, whereas single-parameter formulas produce substantially larger errors. Intensity-stratified analysis further confirms that Fang Genshen retains optimal performance from tropical storm to typhoon intensity levels (RMSE ≤ 8.78 km, MAPE ≤ 13.05%). To overcome the inherent limitations of empirical formulas in capturing complex nonlinear multi-factor coupling, three machine learning models were developed for joint prediction of RMW and the near-center maximum wind speed (Vmax): a multi-task physics-informed neural network (MT-PINN), LightGBM, and random forest (RF). All models use the same 26-dimensional input features—comprising 11 continuous variables including longitude, latitude, central pressure, environmental pressure, and movement characteristics, plus a 15-category satellite type encoding—and are trained on a standardized 7:1.5:1.5 data split. The MT-PINN model adopts a hard parameter-sharing architecture with a shared bottom network learning generic TC structural features and task-specific branches for RMW and Vmax, and embeds the gradient wind balance equation as a physics-informed constraint within the loss function to enhance physical consistency. On the independent test set, MT-PINN achieves the highest prediction accuracy, with RMW RMSE = 3.00 km, MAPE = 5.35%, R² = 0.85, and Bias = ?0.11 km; Vmax RMSE = 1.56 m/s, MAPE = 3.75%, and R² = 0.98. Compared with the optimal Fang Genshen formula, MT-PINN reduces RMW RMSE by 63.1% and MAPE by 64.5%, while increasing R² from 0.63 to 0.85. LightGBM achieves RMW RMSE = 3.08 km, MAPE = 5.48%, and R² = 0.556 (62.1% RMSE reduction), and random forest achieves RMSE = 3.09 km, MAPE = 5.52%, and R² = 0.552 (61.9% reduction). Wind field simulations for Typhoon Muifa (2212) and Typhoon Doksuri (2305) further demonstrate that the MT-PINN-based hybrid model, coupling machine learning RMW predictions with the Holland parametric wind field framework, yields significantly lower simulation errors than conventional empirical-formula-based approaches and more accurately reproduces nearshore wind field structures including peak wind speed distributions and radial decay patterns. These results confirm that integrating multi-task learning with physics-based constraints effectively improves the accuracy, physical consistency, and generalization capability of RMW estimation. The proposed MT-PINN framework provides a reliable parameterization scheme for engineering TC wind field models and storm surge hazard assessment, while LightGBM and random forest offer computationally efficient alternatives suitable for rapid engineering estimation. |
| Key words: radius of maximum wind machine learning typhoon wind field model empirical formula typhoon prediction |
