A Data-Driven Regional Model for Skillful Medium-Range Typhoon Prediction

Abstract

Accurate prediction of tropical cyclones remains a major challenge for both numerical weather prediction and emerging artificial intelligence weather prediction systems. While recent global AI models have demonstrated strong skill in large-scale circulation prediction, they often struggle to represent the mesoscale structures critical for tropical cyclone intensity and precipitation. Here we develop the Hybrid Intelligent Typhoon System (HITS), a regional AI forecasting framework for medium-range typhoon prediction over the Asia-Pacific region, trained on a newly constructed 9-km high-resolution typhoon reanalysis dataset. The model combines regional autoregressive prediction with large scale dynamical constraints from the state-of-the-art ECMWF Artificial Intelligence Forecasting System (AIFS), allowing it to remain dynamically consistent with the evolving large-scale circulation while resolving mesoscale structures. HITS is further extended with a structure-aware perceptual training strategy (HITS-LPIPS) that improves the representation of convective and typhoon rainband structures. Experiments show that the hybrid framework substantially improves precipitation structure and typhoon intensity forecasts compared with both purely autoregressive regional AI models and standalone AI downscaling approaches. In particular, HITS-LPIPS reduces intensity errors by up to 47.8% relative to AIFS at a 72 hour lead time and produces a near-unbiased wind-pressure relationship for simulated typhoons. These results demonstrate that dynamically constrained regional AI systems provide a promising pathway for improving medium-range typhoon prediction.

Figures (6)

• Figure 1: Lead-time dependence of forecast skill for atmospheric variables across different experiments. Panels show verification against HiRes as a function of lead time from 6 to 120 h, evaluated over the testing period from June to November 2025. Rows (a1--a4), (b1--b4), (c1--c4), and (d1--d4) show the root mean square error (RMSE) of temperature, zonal wind, meridional wind, and specific humidity at 250, 500, 850, and 925 hPa, respectively. Specific humidity is given in g kg$^{-1}$. The last row (e1--e4) shows surface variables, including 2 m temperature, 10 m zonal wind, 10 m meridional wind, and mean sea level pressure (MSLP). Black, blue, red, green, and gray curves denote CTL, HITS, HITS-LPIPS, ISTM, and AIFS, respectively.
• Figure 2: A typical summer convective precipitation event and evaluation of precipitation forecast skill across different models in 2025. Columns show forecast lead times of 6 h, 12 h, 18 h, 24 h, 30 h, and 36 h, respectively, initialized at 0000 UTC on 11 August 2025. Rows correspond to composite radar reflectivity forecasts from (a1--a6) HiRes, (b1--b6) HITS-LPIPS, (c1--c6) HITS, (d1--d6) CTL, and (e1--e6) ISTM. Panels (f1--f4) present the fractions skill score (FSS) of composite radar reflectivity as a function of lead time (6--120 h) for thresholds above 10 dBZ, 20 dBZ, 30 dBZ, and 40 dBZ, respectively. The FSS values are averaged over June--December 2025, comparing CTL (grey), HITS (blue), HITS-LPIPS (red), and ISTM (green). Higher FSS indicates improved spatial agreement of reflectivity exceedance patterns.
• Figure 3: Comparison of composite radar reflectivity forecasts for Typhoon Danas (2025) across different experiments. Rows correspond to (a1--a6) HiRes, (b1--b6) HITS-LPIPS, (c1--c6) HITS, (d1--d6) CTL, (e1--e6) ISTM, and (f1--f6) AIFS-TCW. Columns show forecast lead times of 12 h, 18 h, 24 h, 30 h, 36 h, and 96 h from forecasts initialized at 0000 UTC on 5 July 2025.
• Figure 4: Spatial distribution of typhoon tracks and forecast errors in 2025. a, CMA best tracks for all typhoon testing cases in 2025 over the western North Pacific, with track segments colored by intensity category (TD, TS, STS, TY, STY and SuperTY). The black rectangle denotes the regional model domain used in this study. The inset shows the mean track errors as a function of forecast lead time for different experiments. b, Mean track errors (km) as a function of lead time from 0 to 120 h, comparing HITS-LPIPS, HITS, ISTM, AIFS and CTL. c, Mean intensity errors (m s$^{-1}$) over the same forecast range. Numbers in parentheses along the x-axis indicate the sample size at each lead time. Lower values indicate better forecast performance.
• Figure 5: Relationship between minimum sea-level pressure and maximum wind speed in observations and different forecast experiments. Scatter plots show the relationship between minimum sea-level pressure ($P_{\min}$) and maximum surface wind speed ($V_{\max}$) for a, CMA best-track observations, and forecasts from b, AIFS, c, ISTM, d, CTL, e, HITS, and f, HITS-LPIPS for all typhoons in 2025. The red curve in all panels denotes the empirical pressure--wind relationship fitted from the CMA best-track data. Bias values (m s$^{-1}$), defined as the mean difference between modelled $V_{\max}$ and the fitted curve, are annotated in each forecast panel.