Portfolio optimization problems involving Conditional Value-at-Risk (CVaR) are often computationally intractable and require complete information about the distribution of returns, which is rarely available in practice. These difficulties are compounded when the portfolio contains a lot of assets. In this paper, we consider the worst-case CVaR (WCVaR) under mixture of vine copulas distribution uncertainty, which can capture complex and hidden dependence patterns in multivariate data.We compare the out-of-sample performance of the robust strategies based on the mixture of R-vine copulas, mixture of C-vine copulas, mixture of D-vine copulas and nominal CVaR method. The experimental study shows that the robust models based on mixture of R-Vine copulas and mixture of C-Vine copulas perform the best in terms of average returns, Sharpe ratio and cumulative returns. The performance of robust mixture of D-Vine copulas model might be advantageous when the correlation between assets is low.