In the planning operation the currently used model in the long-term operation planning of the Brazilian Interconnected System, the uncertainty concerning streamflow is considered explicitly in the estimation of the expected cost-to-go function, as well as in the covering of the state space, by the use of multivariate hydrological scenarios. The set of all possible realizations of the streamflow stochastic process throughout the planning horizon forms a scenario tree. This tree represents the entire probabilistic universe on which are calculated the optimal operation strategies. As the scenario tree of the long-term operation planning problem has a high cardinality, makes it impossible to visit the complete tree due to computational effort. Therefore, only a portion of the tree (sub-tree) is covered. Currently the sub-tree is selected using the Monte-Carlo method with classical simple random sampling. The objective of this work is to propose a method for defining the sub-tree to be visited during the calculation of the optimal operating strategy for the Brazilian hydro-thermal power system by stochastic dual dynamic programming in order to obtain more robust results from this operation policy with regard to variations in the number of scenarios of forward and backward simulations, and variations in the sample hydrological scenario. There are two proposals for definition of the sub-tree: (i) change the simple random sampling to latin hypercube sampling or descriptive sampling in the multivariate streamflow scenario generation model, and (ii) apply multivariate statistical techniques to develop criteria that allow grouping similar objects in certain groups (clustering techniques). The proposals can be applied together or separately.