This thesis presents an approach to modeling time series that have an autocorrelation structure that depends not only on the time interval between the observations, but the observed period. This approach consists of a combination between wavelet shrinkage and periodic autoregressive models – PAR(p). Wavelets have been used in the literature of time series as an auxiliary procedure for preprocessing the series under analysis, and the PAR(p) model has its importance recognized due to its use in hydrological series monthly. Specifically, in the Brazilian electricity sector, it is possible to observe the predominance of hydroelectric energy. One of the main characteristics of energy matrices with this composition is the strong dependence on rainfall patterns and future hydrological conditions, which turn the streamflow series very irregular and difficult to be modeled. Therefore, hydroelectric generation can be considered a stochastic variable and small advances in stochastic modeling of streamflows allow for better planning of the system´s operation. Thus, this thesis aimed to improve forecasting and scenario generation with the PAR(p) model and to contribute to stochastic modeling of flow. The main results indicated that the combination of methodologies achieves better results than only the PAR(p) modeling, both in simulated cases and in real time series.