Accurate short‐term prediction of typhoon 10‐m wind fields is crucial for early warning and risk reduction. We propose a lightweight spatiotemporal deep‐learning model that couples a convolutional neural network (CNN) for spatial features with a long short‐term memory (LSTM) network for temporal dynamics, augmented by squeeze‐and‐excitation (SE) channel attention and a multi‐branch feature fusion network (MBFN). Using ERA5 winds and China Meteorological Administration best‐track records over East Asia (2020–2023), the model ingests four hourly frames to predict the 10‐m wind field 1‐h ahead. Across root mean square error (RMSE), mean absolute error (MAE), and average wind speed error (AWSE), the approach consistently outperforms U‐Net, ConvLSTM, and Transformer baselines and better reconstructs high‐wind structures near typhoon cores; relative to a plain CNN–LSTM baseline, average RMSE and MAE decrease by 0.90% and 0.68% over 2020–2023. Ablation studies isolate the effects of SE and MBFN, evidencing robust generalization and computational efficiency suitable for near‐real‐time operations. A supplementary 6‐h experiment shows only modest, consistent increases across years—RMSE by 0.54% on average, MAE by 0.50%, and AWSE by 0.41%—indicating robustness at longer lead times.
ABSTRACT
PUBLICATION RECORD
- Publication year
2026
- Venue
Meteorological Applications
- Publication date
2026-01-01
- Fields of study
Not labeled
- Identifiers
- External record
- Source metadata
Semantic Scholar
CITATION MAP
EXTRACTION MAP
CLAIMS
- No claims are published for this paper.
CONCEPTS
- No concepts are published for this paper.
REFERENCES
Showing 1-21 of 21 references · Page 1 of 1
CITED BY
- No citing papers are available for this paper.
Showing 0-0 of 0 citing papers · Page 1 of 1