一种基于深度学习的热带气旋路径集成预报方法

Consensus forecast method for a tropical cyclone track based on deep learning

本文提出了一种基于深度学习的热带气旋(tropical cyclone,TC)路径集成预报方法。该方法以长短期记忆深度网络为模型构架,利用前4个时刻(24 h,间隔6 h)及当前时刻的TC路径记录,以及由不同环境因素所计算的方向预报因子作为输入,分别直接预报和间接(通过预报移动速度)预报路径,集成两者预报结果实现时效为24 h的TC路径预报。试验部分使用不同环境因素所对应方向预报因子进行预报,进而探究在该模型中影响TC路径预报的环境因素。结果表明,经纬向风场所计算的方向预报因子对模型预报性能提升较为明显,而海表温度、高度的方向预报因子对性能提升相对较小。此外,将不同方向预报因子的预报模型进一步集成,可以提升预报精度。上述结果验证了本文所提出的方向预报因子、集成方法在TC路径预报问题中的有效性。

This paper proposes a tropical cyclone(TC)track forecast method based on Long Short-Term Memory(LSTM)networks as the backbone.Inputs include the TC historical data(24 h,6 h interval)and the current track,as well as the direction predictors(DP)calculated by different environmental factors.The model TC-track forecasts the track directly,while the model TC-velocity forecasts the track indirectly by the velocity.The models are integrated based on the consensus model,which forecasts the TC track after 24 hours.Through experiments,this paper compares the models driven by different DPs to explore the influence of the environmental factors on the TC track.The results reveal that the DP of the wind field has a more obvious effect on the forecast model performance,while DPs of the temperature and height fields at the sea surface have a relatively small effect.Moreover,the forecast accuracy can be further im-proved by integrating the selected forecast models of different DPs.The above results verify the feasibility and effec-tiveness of the proposed DP and consensus forecast method for the TC track forecast.