基于机器学习方法的热带气旋加强机理与预报技术研究

The scale of harm of the tropical cyclone (TC) is closely related with the TC intensity. How to better understand the process of the TC intensity change for improving the TC intensity forecast skill, is one of the most concerned issues in the international atmospheric science community. In order to deepen the understanding of the TC intensity change, and to enhance the TC intensity forecast ability, this project plans to apply modern machine learning methods to the analysis of the mechanism of the TC intensity change. Firstly, nonparametric regression and kernel principle component analysis methods are jointly applied, by learning from data of underlying oceanic surface, large-scale environment fields, the TC intensity and structure, to analyze factors influencing the dynamic and thermodynamic processes of the TC intensity change and their combined nonlinear effects. Secondly, gradient boosting decision trees approach is applied to evaluate the individual importance of dynamic and thermodynamic factors to the TC intensification, and the sensitivity of the TC intensification rate to these factors. Based on the above results, an additive model of these factors will be built, to serve as a tool to reveal qualitatively the mechanism influencing the TC intensity change, and an analysis and forecast model that can estimate quantitatively the TC intensification as well. Finally, a prototype of new TC intensity forecast model, with both a sound physical basis and an application prospect, will be developed and operationally tested.

热带气旋（后称TC）的危害程度与其强度直接相关。如何理解TC的强度变化过程，提高TC的强度预报水平是国际大气科学界最为关注的科学问题之一。为了加深对TC强度变化的认识，提升TC强度预报能力，本项目拟将现代机器学习方法应用于TC强度变化机理分析。首先综合运用非参数回归和核主成分分析方法，从海洋下垫面、大尺度环境场、TC涡旋自身强度和结构等方面，通过“学习”数据的方式来分析影响TC强度变化的动力和热力过程以及多影响因子的非线性综合效应；进一步应用梯度提升决策树方法，评估热动力因子对TC增强的重要性，以及TC强度的增强率对不同热动力因子敏感性，并建立影响因子的加性模型，使之成为既能定性解释影响TC强度变化的机理，又能定量估计TC增强的分析和预报模型。在此基础上，发展成为既具有物理基础，又具有业务应用前景的新型TC强度预报模型，并开展运行试验。

尽管近年高分辨率数值预报模式对TC强度预报的精度有所提高，但受限于动力模式对控制TC强度变化的动力和物理过程的认知和描述，误差仍然很大。本项目围绕TC强度和结构变化的内部热动力机理、影响TC强度变化的环境热动力因子和利用机器学习方法结合物理约束进行TC强度预报开展了广泛的研究。通过将现代机器学习方法引入TC强度变化机理分析，研究不同热动力因子及多因子综合效应对TC的增强的重要性及TC的增强率对不同热动力因子敏感性，并将其定量地考虑到模型中。此模型既能定性地解释影响TC强度变化的动力和热力学机理，又能定量地估计不同动力和热力过程对TC强度变化的相对贡献及综合影响。在上述研究基础上发展了具有物理约束与机器学习方法相结合的、具有业务应用前景的新的TC强度预报方法，个例验证试验表明在3-7天TC强度预报优势明显，优于目前现有的官方公布的预报水平。该模型以其理论基础扎实，算法快捷高效在气象导航航线优选，海上台风灾害评估，台风风险评估以及海洋气象预报业务中都会发挥科技支撑作用。