Weighted multiscale nonlinear optimal initial‐perturbation strategy for convective‐scale ensemble forecasting

Abstract Adequately capturing the multiscale and nonlinear nature of initial uncertainties is a prerequisite for constructing convective‐scale ensemble prediction systems (EPSs). In this study, we propose a new initial‐perturbation strategy that weights larger‐ and smaller‐scale Conditional Nonlinear Optimal Perturbations (L‐CNOPs, >200 km; S‐CNOPs, ≤200 km), each designed to maximize nonlinear growth in its respective band. A positive–negative perturbation pair is generated by combining the L‐CNOP and S‐CNOP with a single random weight from a uniform distribution on the interval (0, 1). This perturbation pair is then added to the control analysis to create the two respective ensemble members. The method is evaluated over 24‐hour forecasts within the 3‐km China Meteorological Administration (CMA) convective‐scale EPS across 10 convective precipitation and two non‐precipitation cases from three regions of China. The results demonstrate that the initial‐perturbation strategy yields significant improvements during the first 18 forecast hours, even compared to the conventional dynamic downscaling approach, which also incorporates uncertainties in both initial and lateral‐boundary conditions. These improvements are reflected in enhanced probabilistic forecast skill for precipitation, particularly for heavy rainfall, evidenced by lower Brier Scores, a larger area under the Receiver Operating Characteristic (ROC) curve, and higher Fraction Skill values, along with reduced root‐mean‐squared error and increased ensemble spread across multiple pressure levels. The benefit of the strategy weakens after 18 hours, with supplementary experiments confirming that lateral‐boundary uncertainty becomes the case‐dependent dominant control on late‐stage forecast skill. These findings demonstrate that the proposed Weighted Multiscale CNOP method constitutes an effective strategy for enhancing convective‐scale EPS performance by better capturing multiscale initial uncertainties and improving probabilistic forecast skill.