Revealing the Potential of Learnable Perturbation Ensemble Forecast Model for Tropical Cyclone Prediction

TL;DR

This study introduces FuXi-ENS, a learnable perturbation ensemble forecasting framework for tropical cyclones, and rigorously compares it against ECMWF-ENS using 2018 TC data. FuXi-ENS delivers improved track accuracy and tighter ensemble coherence, and better representation of large-scale circulation and moist thermodynamics, though it consistently underestimates TC intensity. The work highlights the potential of flow-dependent, AI-driven perturbations to enhance ensemble skill in TC forecasting and provides mechanistic insights into the dynamical and thermodynamical underpinnings of improved performance. It also discusses current limitations and proposes targeted improvements in loss functions, data resolution, and end-to-end optimization for extreme events, signaling a promising route for AI-based ensemble predictions of high-impact weather events.

Abstract

Tropical cyclones (TCs) are highly destructive and inherently uncertain weather systems. Ensemble forecasting helps quantify these uncertainties, yet traditional systems are constrained by high computational costs and limited capability to fully represent atmospheric nonlinearity. FuXi-ENS introduces a learnable perturbation scheme for ensemble generation, representing a novel AI-based forecasting paradigm. Here, we systematically compare FuXi-ENS with ECMWF-ENS using all 90 global TCs in 2018, examining their performance in TC-related physical variables, track and intensity forecasts, and the associated dynamical and thermodynamical fields. FuXi-ENS demonstrates clear advantages in predicting TC-related physical variables, and achieves more accurate track forecasts with reduced ensemble spread, though it still underestimates intensity relative to observations. Further dynamical and thermodynamical analyses reveal that FuXi-ENS better captures large-scale circulation, with moisture turbulent energy more tightly concentrated around the TC warm core, whereas ECMWF-ENS exhibits a more dispersed distribution. These findings highlight the potential of learnable perturbations to improve TC forecasting skill and provide valuable insights for advancing AI-based ensemble prediction of extreme weather events that have significant societal impacts.

Figures (21)

• Figure 1: ROCASS comparison between FuXi-ENS (green) and ECMWF-ENS (orange) for surface variables over 15-day forecasts with 6-hourly evaluation intervals. ROCASS are evaluated for the upper tercile (top panel) and lower tercile (bottom panel) of three surface variables: mean sea level pressure (MSL, left panel), 2m temperature (T2M, middle panel), and 10m wind speed (WS10M, right panel).
• Figure 1: Probabilistic forecast comparison for hurricane FLORENCE track prediction. Hurricane FLORENCE track probabilistic forecast distribution based on initialization at 12 UTC on September 3, 2018: (a) FuXi-ENS and (b) ECMWF-ENS. Blue lines represent ensemble mean tracks, black lines represent IBTrACS observed tracks, and gray lines represent ECMWF-HRES (deterministic forecast). Shading indicates TC strike probability (%). Date labels mark the temporal evolution of the track.
• Figure 2: ROCASS comparison between FuXi-ENS (green) and ECMWF-ENS (orange) for upper-level variables over 15-day forecasts with 6-hourly evaluation intervals. ROCASS are evaluated for the upper tercile (top panel) and the lower tercile (bottom panel) of three upper-level atmospheric variables: 500 hPa geopotential (Z500, left panel), 850 hPa temperature (T850, middle panel), and 850 hPa wind speed (WS850, right panel).
• Figure 2: Probabilistic forecast comparison for hurricane BUD track prediction. Hurricane BUD track probabilistic forecast distribution based on initialization at 00 UTC on June 10, 2018: (a) FuXi-ENS and (b) ECMWF-ENS. Blue lines represent ensemble mean tracks, black lines represent IBTrACS observed tracks, and gray lines represent ECMWF-HRES (deterministic forecast). Shading indicates TC strike probability (%). Date labels mark the temporal evolution of the track.
• Figure 3: Violin plots of TC track forecast errors in 5-day forecasts for ECMWF-ENS (purple) and FuXi-ENS (teal). (a), accumulated ensemble mean position error ($\mathrm{AccError}_{\mathrm{TC}}$, km). (b), accumulated ensemble spread ($\mathrm{AccSpread}_{\mathrm{TC}}$, km). (c), along-track error (AT, km). (d), cross-track error (CT, km). Violin plots show the distribution characteristics of each error metric across forecast lead times, with dashed horizontal lines indicating the median and interquartile ranges marked. Purple represents ECMWF-ENS forecasts, and teal represents FuXi-ENS forecasts. Asterisks indicate statistical significance levels ($*$$\mathrm{p} < 0.05$, $**$$\mathrm{p} < 0.01$, $***$$\mathrm{p} < 0.001$) based on Mann-Whitney U tests comparing the two ensemble systems.