Subseasonal forecasting of extreme heavy rainfall in extratropical regions remains a significant challenge for dynamical models and data-driven methods. This study investigates the predictability of the record-breaking rainfall event that struck Beijing from 25 to 29 July 2025, based on eight operational dynamical models and the AI-based “FENGSHUN” system. Results show that conventional full-member multi-model ensembles (MMEs) significantly underestimated the rainfall intensity, losing probabilistic skill beyond a 5-day lead time. This failure is attributed to the “noise” generated by large ensemble members of light or no rainfall forecast, which dilutes the signal of rainfall extreme. Here we propose a “noise-free” ensemble scheme, designed to preserve high-impact weather signals. Unlike traditional ensemble techniques that aim to minimize mean error, this approach filters out members with daily rainfall forecast below 10 mm day⁻¹ (approximately the local 90^th percentile). Meanwhile, we impose a confidence constraint derived from false-alarm analysis, requiring at least 16% of the ensemble members to exceed this threshold. Application suggests that the forecast skill of extreme rainfall event in Beijing can be greatly extended to subseasonal timescale by two weeks. The improvement stems from the scheme’s ability to selectively isolate members that correctly resolve key tropical-extratropical interactions. The source of subseasonal predictability is mainly from impacts of Rossby wave train from the Barents-Kara Sea, the subsequent coupling between the North China cold vortex and the western Pacific subtropical high.

Rainfall extremes, Subseasonal forecast, Ensemble scheme