Abstract: In ensemble forecast, by summing up ensemble members, filtering the uncertainty, and remaining the common component, the ensemble mean with better results can be obtained. However, filtering works only when the perturbation develops nonlinearly, if perturbations propagate in a linear space approximately, the positive and negative couple members will counteract each other, lead to little difference between ensemble mean and control forecast, thus making the ensemble insignificant finally. In 1-2 days ensemble forecast based on singular vector calculation, to avoid this mutual cancellation, the couple members originated from the same singular vector should not be all both into the ensemble system, the only candidate should be the one with better forecast. Based on the ingredient analysis of initial perturbation development, one method for selecting ensemble member is presented, which can satisfy the above requirement. The regional model MM5V1 of NCAR/PSU (the National Center of Atmosphere Research/the University of Pennsylvania) and its corresponding tangent and adjoint models are used. The ensemble spread and forecast errors are calculated with dry energy norm. Two mesoscale vortex cases on the meiyu front in the Yangtze River basin are used. The initial perturbation putting onto the control forecast develops with nonlinear mode integral, the practical calculating indicates that the perturbation is mainly linear, and most of those linear components could be interpreted by the tangent model. This ingredient analysis shows the necessariness for choosing the optimized ensemble member in short forecast. According to the theory analysis of perturbation ingredient, among a couple of members from different singular vectors, those members performing better always have smaller or greater spreads compared with other members. Following this principle, one optimized ensemble and one inferior ensemble are formed. Results of practical calculation show that the ensemble mean of optimized ensemble is more accurate than that of inferior ensemble, and also performs better than the traditional ensemble with positive and negative member simultaneously. As for generating the initial perturbations, the initial perturbations derived from multiple orthogonal singular vectors grow much quicker than those from the single singular vector, the former can effectively enlarge the range of spread of ensemble numbers, and thus raise the forecast accuracies of ensemble members.