| 摘要: |
| 在全球气候变化背景下，越来越多的地区面临着热带气旋的威胁。因此，准确预测热带气旋的轨迹变化对于气象预警和灾害管理至关重要。然而，传统的基于深度学习的热带气旋预测方法在建模热带气旋的时空相关性方面存在局限。为此，提出了一种新的深度时空融合网络——DSTFN(Deep Spatio-TemporalFusion Network)模型，以提高对热带气旋轨迹的预测精度和稳定性。构建了有效融合 ConvNeXt(ConvolutionalNext) 模 型 和 门 控 循 环 单 元 的 CaConvNeXt - GRU(Convolutional Block Attention Module Integrated withConvNeXt and Gated Recurrent Unit)模型，以提取热带气旋三维时序数据中的复杂非线性时空特征。同时，引入了卷积块注意力模块，以自动聚焦不同等压面对热带气旋影响更大的特征。此外，设计了分阶段的训练策略，通过依次进行预训练、联合训练和整体训练实现了不同模块的有效融合。为了评估所设计的方法，在国际气候管理最佳路径档案和第五代大气再分析数据集上进行了大量实验。实验结果证明，在预测未来24 h的热带气旋轨迹时，相比于现有的基于深度学习的热带气旋轨迹预测模型，DSTFN模型的平均预测误差降低了约13.71 km。 |
| 关键词: 热带气旋 路径预测 DSTFN模型 CaConvNeXt-GRU模型 时空序列预测 |
| DOI：10.16032/j.issn.1004-4965.2024.077 |
| 分类号: |
| 基金项目: |
|
| Tropical Cyclone Track Prediction Method Based on DSTFN Model |
|
FANG Wei1,2,3,4, DU Juan1, QI Meihan1, HU Pengyu5
|
|
1. School of Computer Science/Engineering Research Center of Digital Forensics of Ministry of Education, Nanjing University of Information Science & Technology, Nanjing 210044, China;2. Key Laboratory of Transportation Meteorology of China Meteorological Administration, Nanjing Joint Institute for Atmospheric Sciences, Nanjing 210041, China;3. Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology (CICAEET), Nanjing University of Information Science & Technology, Nanjing 210044, China;4. Provincial Key Laboratory for Computer Information Processing Technology, Soochow University, Suzhou, Jiangsu 215000, China;5. Shanghai Vocational College of Agriculture and Forestry, Shanghai 201699, China
|
| Abstract: |
| In the context of global climate change, more and more regions are facing the threat of tropical cyclones. Therefore, accurate prediction of changes in the tracks of tropical cyclones is essential for meteorological warning and disaster reduction. However, existing tropical cyclone prediction methods based on deep learning have limitations in modeling the spatio-temporal correlation of tropical cyclones. In the present study, we proposed a new deep spatio-temporal fusion network (DSTFN) model to improve the prediction accuracy and stability of tropical cyclone tracks. We developed the CaConvNeXt-GRU model, which effectively integrated the ConvNeXt model and the gated recurrent unit, to extract complex nonlinear spatio-temporal features in the 3D time series data of tropical cyclones. Meanwhile, the convolutional block attention module was introduced to automatically focus on the features that were affected more heavily by different isobaric surfaces on tropical cyclones. Moreover, we designed a staged training strategy to realize the effective integration of different modules through pre-training, joint training, and overall training. To evaluate the proposed model, we conducted extensive experiments on the International Best Track Archive for Climate Stewardship (IBTrACS) and the ERA5 dataset. Overall, in predicting tropical cyclone tracks for the next 24 hours, the DSTFN model reduced the average prediction error by about 13.71 km compared to existing tropical cyclone track prediction models based on deep learning. |
| Key words: tropical cyclone path prediction DSTFN model CaConvNeXt-GRU model spatio-temporal series prediction |
