Abstract: Convective-scale numerical weather prediction is sensitive to minor IPs (initial perturbations), and the evolution of these perturbations is model-, flow-, and scale-dependent. Constructing reasonable IPs for convection-permitting ensemble prediction systems has always been challenging. The authors use GRAPES (Global/Regional Assimilation and Prediction System) 3-km convective-scale model from the Center for Earth System Modeling and Prediction of the China Meteorological Administration. The authors employ a two-dimensional random function and background error of assimilation system in GRAPES 3-km to construct large-, meso-, and small-scale stochastic initial perturbation fields. Using these different-scale IPs, the authors conduct three convective-scale ensemble forecast experiments on a typical multiregional heavy precipitation weather process during summer in China. The authors analyze the spatiotemporal evolution and spectral decomposition characteristics of perturbation energy in these three experiments to elucidate the evolution characteristics of different-scale initial perturbations in a convective-scale model. This analysis provides a reference for constructing optimal initial perturbations in GRAPES convection-permitting ensemble prediction systems. The results reveal the following: (1) There are significant differences in the evolution of DTE (difference total energy) among the three IP experiments. The DTE of the large-scale IP increases with model integration, particularly in the middle and upper troposphere. However, the DTE evolution of the meso- and small-scale IP experiments exhibits an apparent diurnal cycle characteristic. In particularly, it exhibits a significant increase (decrease) during the period ranging from afternoon to evening (from night to morning) when convection is active (passive). The diurnal cycle is primarily caused by the small-scale component of DTE. The diurnal cycle of DTE may be due to the surface heating caused by solar short-wave radiation, which facilitates more active convection during the daytime than at night, and the convection directly affects the small-scale component of the DTE. In addition, the DTE of three IP experiments increases primarily due to the development of DKE (difference kinetic energy), whereas DPE (difference potential energy) cannot be neglected in the lower troposphere. (2) The DTE evolution of the large-, meso-, and small-scale IP experiments is flow dependent. In the mid-high latitudes, the increase in DTE for the large-scale IPs is dominant in regions with strong baroclinic instability (e.g., trough regions), whereas it does not develop in regions with relatively weak baroclinic instability (e.g., the northwest flow behind troughs). In the confluence region of north and south airflows, the DTE increases for the large-scale IPs are still dominant. However, the DTE of all three IP experiments hardly develops in the region affected by the South China Sea summer monsoon. A consistent relationship exists between the development of DPE and the ratio of large precipitation rates in this region. (3) The DTE spectrum reveals that the multiscale cascade characteristics of DTE change with integration periods. The downscaling cascade of DTE from the large-scale to the small-scale component is strong during the initial 3 h. However, for lead times after 6 h, the upscale growth of DTE from meso- and small-scale components becomes the main characteristic of the DTE spectrum. In conclusion, it is necessary to construct a scale-dependent and flow-dependent initial perturbation structure for different unstable weather regions, particularly when building convection-permitting ensemble predictions in regions with complex weather systems and nonuniform spatiotemporal distribution of dynamic instability, such as China.