In order to help system (transmission and/or distribution) operators to account for the stochastic nature of wind during the grid planning phase, univariate AutoRegressive Moving Average (ARMA) time series models for the long term modeling of wind speed are generally considered. Practically, different geographical correlation levels are observed in the bibliography when sampling wind generation in the framework of long-term analysis tools (e.g. sequential Monte Carlo algorithms). The traditional approach is to assess extreme correlation scenarios (entire independence, entire correlation). Some recent works try to better capture the correlation patterns of real data by using matrix methods, like the Cholesky decomposition. In this work, two methods for reproducing the actual correlation are compared on a statistical basis on the one hand, and in the framework of a Monte-Carlo reliability analysis on the other hand. It is shown that a good correlation model is mandatory for obtaining correct reliability indices.