Severe mesoscale rainstorm struck Guangzhou heavily from deep night of 6 May to early morning of 7 May 2017, and convection initiation was closely related to southerly wind enhancing in boundary layer near Pearl River Delta at night. Comparison between two subgroups in ECWMF ensemble forecasts shows members with strong rainfall have significant lowlevel convergence, updraft, moist air and convective instability in initial conditions. GRAPES mesoscale model performed well in forecasting the dynamic process of southerly wind strengthening in lower level, producing good forecasts for the rainfall process. Ensemble sensitivity analysis reveals that the high sensitivity depends on the relative position, strength of key synoptic systems for high pressure in Yangtze River Basin, high pressure on South China Sea and low pressure trough residing on South China. Precipitation forecast sensitivity to temperature is closely relative to convective instability in boundary layer, the boundary structure with warmer air in nearsurface and colder air near the top of boundary layer would increase convective instability obviously. Three sets of convectivescale simulations are applied to analyze sensitivity of convective precipitation forecast to initial thermal disturbance, lowlevel wind disturbance, and cloud microphysical scheme differences. Convective precipitation forecast shows more sensitivities to initial thermal disturbance compared with lowlevel wind disturbance, and forecast uncertainty could be estimated comprehensively through different cloud microphysical schemes. By carefully analyzing forecast sensitivity due to perturbations in initial conditions and physical process in the convectivescale ensemble forecasts, the predictability of mesoscale heavy rainfall events in warm seasons could be improved.