This paper addresses the problem of forecasting for monthly mean streamflow series, in which we call the forecast horizon (h), the interval of time between the last observation used in fitting the model prediction and the future value being predicted. The analysis of the forecast error is made on the basis of the forecast horizon. These series have a periodic behavior on average, ariance and autocorrelation function. Therefore, we consider the widely used approach to modeling these series that initially consists of removing the periodicity in mean and variance of the streamflow series and then calculating a standardized series for which stochastic models are adjusted. In this study we consider the series to the standard periodic autoregressive models PAR (p m). Orders p m of the adjusted models for each month are determined from the analysis of periodic partial autocorrelation function (PePACF), using the Bayesian Information Criterion (BIC) applied to PAR models, proposed in (MecLeod, 1994) and analysis of PePACF proposed in (Stedinger, 2001). The forecast errors are calculated on the basis of parameters adjusted and evaluated for forecasting horizons h, ranging from 1 to 12 months on the original scale of the flow series. These errors are compared with estimates of the variances of the flows for the month being predicted. As a result we have an evaluation of the predictive performance in months of the adjusted models for each month.
PAR model; streamflow forecasting; forecast error; hydroelectric systems
