Qu et al. (2025) Accurate tropical cyclone intensity forecasts using a non-iterative spatiotemporal transformer model

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2025
• Date: 2025-12-06
• Authors: Hongyu Qu, Hongxiong Xu, Lin Dong, Chunyi Xiang, Gaozhen Nie
• DOI: 10.1038/s41612-025-01279-3

Research Groups

• National Meteorological Centre, Beijing, China
• State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, China Meteorological Administration, Beijing, China

Short Summary

This study introduces TIFNet, a non-iterative spatiotemporal transformer model, to generate accurate 5-day tropical cyclone (TC) intensity forecasts. TIFNet integrates high-resolution global forecasts with historical TC evolution, consistently outperforming operational numerical models, especially during rapid intensification and weakening events.

Objective

• To develop an artificial intelligence (AI)-based model capable of generating accurate and consistent long-range (up to 5 days) tropical cyclone intensity forecasts, specifically addressing the challenges of predicting rapid intensification (RI) and rapid weakening (RW) events where traditional models underperform.

Study Configuration

• Spatial Scale: Primarily the Western North Pacific (WNP) basin, with a preliminary assessment in the Eastern North Pacific (ENP). Input data cropped to a 17.75° × 17.75° domain, represented by 72 × 72 grid points at 0.25° resolution, centered on the TC.
• Temporal Scale: Forecasts generated for lead times up to 120 hours (5 days) at 6-hour intervals. The model uses 24 hours of pre-initialization data and 120 hours of post-initialization forecast fields. Training data spans 1990–2022, with real-time evaluation on independent cases from 2023–2024.

Methodology and Data

• Models used:
  • TIFNet: A transformer-based encoder–decoder network with a non-iterative design, spatiotemporal patch embedding, positional and temporal encoding, divided space-time attention, cross-attention, multi-head self-attention, and a fully connected layer.
  • Baselines: ECMWF Integrated Forecasting System (IFS), NCEP Global Forecast System (GFS), CMA Typhoon Model (TYM), CMA Official Forecasts (CMA_OF), and SAFNet (an AI-based model).
• Data sources:
  • CMA best-track dataset: Six-hourly records of TC positions and 2-minute maximum sustained wind speed over the WNP (1949–present), used for pretraining.
  • CMA real-time typhoon position and intensity analysis dataset: Near-real-time TC center and intensity (Dvorak analysis), used for fine-tuning and real-time evaluation.
  • ERA5 reanalysis dataset (ECMWF): Hourly global atmospheric fields at approximately 31 km horizontal resolution (1990–2022), including zonal/meridional winds, temperature, geopotential height, relative humidity (7 pressure levels), sea surface temperature, and 6 surface variables, used for pretraining.
  • ECMWF Integrated Forecasting System (IFS) forecasts: Surface variables at 0.125° resolution and upper-air variables at 0.25° resolution (up to 240 h lead time), bilinearly interpolated to 0.25°, used for fine-tuning (2000–2022) and real-time evaluation (2023–2024).

Main Results

• TIFNet consistently achieved the lowest Mean Absolute Error (MAE) across all forecast lead times (24–120 hours) compared to operational numerical models.
• At 24 hours, TIFNet reduced MAE by 57% against IFS, over 25% against GFS and TYM, and 12% against CMAOF. This advantage persisted, with MAE reductions of 29% (IFS), 21% (TYM), 8% (GFS), and 13% (CMAOF) at 120 hours.
• TIFNet maintained consistently low and stable Mean Error (ME) values, avoiding the substantial negative biases observed in several baseline models.
• The model demonstrated superior or comparable performance across all TC intensity categories, with its advantage becoming more pronounced for stronger systems, including Super Typhoons.
• In rapid intensification (RI) cases, TIFNet reduced forecast error (MAE) by 27–48% relative to operational baselines. For rapid weakening (RW) events, MAE reductions ranged from 6–57%.
• Case studies of high-impact TCs (e.g., DOKSURI) confirmed TIFNet's ability to accurately reproduce complex intensity evolutions, including multiple RI episodes and rapid decay, outperforming operational models.
• The non-iterative prediction design was crucial for capturing steep intensity gradients characteristic of RI and SuperTY events, as iterative variants showed degraded peak-intensity prediction.

Contributions

• Introduction of TIFNet, a novel non-iterative spatiotemporal transformer model for TC intensity forecasting, which generates coherent 5-day intensity trajectories in a single forward pass, mitigating cumulative error propagation.
• Demonstrated superior forecast skill over state-of-the-art operational numerical models (ECMWF IFS, NCEP GFS, CMA TYM, CMA_OF) and existing AI-based models (SAFNet) across all forecast horizons and intensity categories.
• Achieved significant improvements in forecasting rapid intensity changes (RI and RW), a long-standing challenge in TC prediction, by preserving the resolution of intensity fluctuations.
• Successfully integrated high-resolution global forecast fields with historical cyclone evolution, bridging the gap between data-driven models and real-time operational applications.
• Provided a scalable and transferable framework with preliminary evidence of cross-basin applicability, laying a foundation for next-generation global AI weather models and enhancing predictability of extreme events under climate change.

Funding

• National Key R&D Program of China (grant 2023YFC3008005)
• National Natural Science Foundation of China (grants 42375015 and 42192554)
• Typhoon Scientific and Technological Innovation Group of the China Meteorological Administration (grant CMA2023ZD06)
• Basic Research Fund of CAMS (grant 2023Z020)
• S&T Development Fund of CAMS (grants 2024KJ018 and 2024KJ022)
• State Key Laboratory of Severe Weather Meteorological Science and Technology (grant 20250ZA18)

Citation

@article{Qu2025Accurate,
  author = {Qu, Hongyu and Xu, Hongxiong and Dong, Lin and Xiang, Chunyi and Nie, Gaozhen},
  title = {Accurate tropical cyclone intensity forecasts using a non-iterative spatiotemporal transformer model},
  journal = {npj Climate and Atmospheric Science},
  year = {2025},
  doi = {10.1038/s41612-025-01279-3},
  url = {https://doi.org/10.1038/s41612-025-01279-3}
}