基于双注意力机制的台风轨迹预测模型

台风轨迹的准确预测对于减少台风灾害及风险评估意义重大。本文提出了一种基于双注意力机制的台风轨迹预测模型(Dual-Attention-Encoder-Decoder),首先根据台风轨迹数据计算台风轨迹的变化曲率,将台风曲率序列与台风轨迹序列一同作为预测模型的特征输入,充分考虑了台风轨迹中隐藏的转向、偏折信息;然后构建双注意力机制增强的编码器-解码器网络(Encoder-Decoder)作为预测模型,利用特征注意力机制和时间注意力机制分别对模型输入和隐藏状态进行权重分配,能够学习输入特征和预测目标之间的关系,并且有效解决编码器-解码器结构对过长序列预测的性能下降问题,编码器和解码器均采用LSTM网络,能够存储长时间依赖并且收敛性好,不易发生梯度消失或爆炸;最后,本文使用1949—2017年中国气象局提供的西北太平洋台风最佳路径数据集,将DA-Encoder-Decoder模型与BP、SVR、LSTM、ELM等模型进行对比,分别对24 h、48 h、72 h台风轨迹进行预测。结果表明:DA-Encoder-Decoder模型的均方根误差和实际误差距离指标均优于其他四种预测方法,验证了本文方法的有效性。

Typhoon trajectory prediction model based on dual attention mechanism

Accurate prediction of typhoon trajectory is of great significance for typhoon disaster deduction and risk assessment. This paper proposes a typhoon trajectory prediction model based on the dual attention mechanism, Dual-AttentionEncoder-Decoder. First, the typhoon trajectory change curvature is calculated based on the typhoon trajectory data. The typhoon curvature sequence together with the typhoon trajectory sequence are serving as the feature input to prediction model, taking full account of the hidden steering and deflection information in the typhoon trajectory. Then an Encoder-Decoder network enhanced by the dual attention mechanism is built as a predictive model. Using the feature attention mechanism and time attention mechanism for weight distribution of model input and hidden state respectively, we can learn the relationship between input features and prediction targets, and effectively solve the problem of performance degradation of the encoderdecoder structure for the excessively long sequences prediction. The encoder and decoder use LSTM respectively. The network can store long-term dependence and has good convergence, and it is not easy to cause gradient disappearance or explosion. Finally, the best path data set of the Northwest Pacific typhoon provided by the China Meteorological Administration (CMA) from 1949 to 2017 is used to compare the DA-Encoder-Decoder with BP, SVR, LSTM, and ELM to predict the typhoon trajectory at 24h, 48h, and 72h respectively. The results show that the root mean square error and actual error distance indicators of DA-Encoder-Decoder model are better than the other four prediction methods, demonstrating the effectiveness of this method.