The production of long artificial series of multiple streamflows given for say monthly or weekly values is considered, taking into account the usually very nonnormal distributions involved, serial dependence, seasonal variation, and the correlation between different streams. It is proposed that the distribution of flows for a given stream and month be fitted by gamma type distributions, and that these distributions then be transformed to obtain normally distributed variates. A separate canonical correlation analysis for each successive pair of months enables one to set up a nonstationary seasonally varying Markovian process that is then transformed back into the original nonnormal variates. The uses of such simulated series for calculating the probability behavior of a system of dams and for the approximate optimization of their operation is described, together with the estimation of standard errors.