基于融合TC-WREM 模型的热带气旋大风半径估算研究

Research on the Estimation of Tropical Cyclone Gale Radius Based on a Fusion TC-WREM Model

关键词:  热带气旋  大风半径估算  卷积神经网络模型  多层感知器神经网络模型  融合 TC-WREM 模型  西北太平洋

Key words:  tropical cyclone  wind radius estimation  Convolutional Neural Network  multi-layer perceptron net  fusion TC-WREM model  western North Pacific

摘要:

利用 2001—2020年美国联合台风警报中心(JTWC)热带气旋(Tropical Cyclone，TC)最佳资料数据集和静止气象卫星云图，建立了基于多层感知器神经网络模型(Multi-Layer Perceptron，MLP)和卷积神经网络(Convolutional Neural Network，CNN)融合的 TC 大风半径估算模型(TC Wind Radii Estimation Model，TC-WREM)。该模型利用 MLP和 CNN分别对 TC属性数据和卫星云图中与 TC大风半径相关联的核心特征进行预提取，最终通过融合 TC-WREM 模型开展大风半径估算。融合的 TC-WREM 模型能实现对 TC 属性数据和卫星云图底层特征的深度客观挖掘，较单独的MLP和CNN模型的估算误差降低7%~24%。以TC近地面8级大风半径(R8)估算为例，针对 2021 年台风“烟花”的独立样本估算检验显示分象限 R8 估算平均绝对误差(MeanAbsolute Error，MAE)分别为 39、33、40 和 51 km，均值为 41 km，误差中位值约 40 km，优于业务估算精度(为大风半径的 25%～40%)及西北太平洋和大西洋同类研究估算结果。由于融合 TC-WREM 模型的输入为易获取的TC属性数据和静止气象卫星云图，因此该模型易于在业务中进行推广，从而可改善国内 TC大风半径估算模型缺乏的现状。

Abstract:

In this paper, a fusion TC Wind Radii Estimation Model (TC-WREM) that combines a Multi- Layer Perceptron net (MLP) and a Convolutional Neural Network (CNN) is established by using the Tropical Cyclone (TC) best track data set and the static satellite cloud images. This model utilizes MLP and CNN to pre-extract the core features associated with TC wind radius from TC attribute data and satellite cloud images and ultimately performs gale wind radius estimation. The fused TC-WREM model in this study can achieve deep and objective mining of TC attribute data and underlying features of satellite cloud images, whose estimation error is reduced by about 7%-24% compared to individual MLP and CNN models. Taking the estimation of 17.2 m·s-1 wind radius of TC In-fa in 2021 as an example, the fused TC- WREM model has higher estimation accuracy than the independent MLP and CNN model. Independent sample testing shows that the mean absolute estimation error in 4 quadrants is 39, 33, 40, and 51 km, respectively, with an average of 41 km, respectively, which is superior to that of other similar research. The fused TC-WREM model is advantageous due to its utilization of easily obtainable TC attribute information and geostationary meteorological satellite cloud images as inputs. This makes it suitable for operational use and addresses the current lack of domestic TC gale radius estimation models.