Abstract: A warm-sector torrential rain under weak synoptic scale forcing occurred in Guangzhou on 7 May 2017. The precipitation process developed rapidly and locally, and the precipitation intensity is extremely high. Many operational numerical weather prediction models failed to forecast this storm. To study the model uncertainty in the forecast of this precipitation process, the Conditional Nonlinear Optimal Perturbation related to Parameters (CNOP-P) is adopted to select key physical parameters which can best represent the nonlinear error growth characteristics of the meso-micro scale system. A new model perturbation scheme CNOP-P-RP is constructed based on these key parameters. Convective-permitting ensemble prediction experiment is carried out based on the CMA-Meso model. Finally, the physical mechanism behind the influences of key parameters selected by CNOP-P on local convection in different stages is investigated. The result shows that the key parameters selected by CNOP-P are mainly related to vertical diffusion, auto-conversion from cloud to rain and conversion from other hydrometeors to raindrops. Compared with Stochastic Perturbed Parameterization Tendencies (SPPT) scheme which is widely utilized in operational ensemble prediction systems, the ensemble prediction experiment based on the CNOP-P-RP scheme is more skillful and reliable for probability forecast of precipitation and surface elements in this process. Further analysis shows the variation of piedmont temperature gradient and surface cold pool caused by the uncertainty of vertical diffusion plays an important role in convective triggering and rainstorm development. From 00:00 BT to 04:00 BT 7 May, the enhancement of vertical diffusion near the center of heavy precipitation in Huadu strengthened the vertical transport of heat, momentum and water vapor. The melting of snow and graupel particles is the main reason for the increase of precipitation, indicating that although the formation of raindrops is mainly caused by condensation near the top of boundary layer and collision of cloud water, the effect of ice particles cannot be ignored. From 04:00 BT to 08:00 BT 7 May, with the strengthening of water vapor transport and upward movement, a more active warm rain process dominated the increase of precipitation in the heavy precipitation center at Zengcheng. This study preliminarily proves the feasibility of the CNOP-P-RP scheme in describing the uncertainties in convection-permitting ensemble prediction systems, and provides some references for the improvement of warm-sector torrential rain forecast in South China.