一种基于大语言模型的跨模态多源数据台风路径预测方法及系统
By constructing a cross-modal multi-source dataset, a semantically guided cross-modal association mechanism, and an LSTM low-rank fine-tuning mechanism, the shortcomings of existing technologies in typhoon path prediction are addressed. This enables accurate prediction of the long-term trajectory and short-term abrupt changes of typhoon paths, enhancing the model's feature representation ability and prediction accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-17
AI Technical Summary
Existing typhoon track prediction methods suffer from insufficient forecast accuracy due to limitations in initial field sensitivity, computational resource constraints, weak physical constraints, simplistic multimodal meteorological field fusion methods, and limited ability to model long-term dependence and abrupt changes in typhoon processes.
We construct a cross-modal multi-source dataset, including historical typhoon trajectory data, ERA5 reanalysis data, and Himawari satellite data. Through a semantically guided cross-modal association mechanism and a low-rank fine-tuning mechanism of LSTM, we train a large language model that integrates physical information to achieve accurate prediction of the long-term trajectory and short-term path changes of typhoons.
It significantly enhances the physical meaning and expressive power of features, achieves precise alignment between semantic modalities and physical spatiotemporal features, accurately grasps the long-term trajectory of typhoons and keenly captures short-term path changes, ensuring the stability and convergence effect of model training.
Smart Images
Figure CN122172351B_ABST