基于时空特征的热带气旋强度预测方法

热带气旋是一种极具破坏力的天气系统,我国每年都深受其带来的灾害困扰. 目前热带气旋的强度预报业务以统计预报方法为主,通过利用气候持续因子对热带气旋未来的强度建立回归模型,不仅需要进行复杂的特征选择,而且缺乏对周围环境信息的利用,预报精度多年以来都未能有显著提升. 提出了一个能同时提取时序特征与空间特征的热带气旋强度预测网络,针对环境场物理量因子对热带气旋的影响,使用卷积层学习其空间信息,结合循环神经单元对热带气旋的历史时间序列进行建模,实现端到端的预测输出. 在对西北太平洋的热带气旋样本进行测试后,结果表明该24 h强度预测网络显著优于上海台风研究所公布的相应时段其他预报方法,故可作为一种新的智能预测模型,为预报员提供有价值的客观参考.

An Tropical Cyclone Intensity Prediction MethodBased on Spatial-Temporal Features

Tropical Cyclone(TC)is a destructive weather system,which causes disasters every year in China. At present,researchers usually develop statistical forecast methods for TC intensity forecast. They use the climatic persistence factors to develop a regression model for the future intensity of TC. Such model,however,needs a complex procedure of feature selection and lacks the use of information in the surrounding environment. So the forecast accuracy has not been significantly improved over the recent years. This paper proposes a TC intensity prediction model that can extract spatial and temporal features simultaneously. As for the influence of environmental physical factors on TC,the convolutional layer is used to learn its spatial information,and the recurrent neural unit is used to model the historical time series of tropical cyclone to achieve an end-to-end prediction. Having tested the TC samples in the Northwest Pacific,the results show that our spatial-temporal intensity prediction network is superior to other forecast methods published by Shanghai Typhoon Institute(STI)in the corresponding period,so it can be used as a new intelligent prediction model to provide valuable objective reference for forecasters.