Metamodeling has been used for design optimization of large-scale engineering problems. Kriging method is one of popular methods due to its accuracy and efficiency. There have been many attempts to improve the accuracy of Kriging method: blind Kriging, dynamic Kriging, etc. These attempts select adequate basis functions to describe the mean structure of the response. However, blind Kriging cannot describe a highly non-linear trend function accurately. In addition, dynamic Kriging takes significant computational time as the number of samples increases due to genetic algorithm and mean value of cross-validation error. In this paper, the new method that selects basis functions using process variance changes and cross-validation(CV) is proposed. Numerical study verified that the proposed method reduces computational time without sacrificing its accuracy.