Application of periodic autoregressive process to the modeling of the Garonne river flows

Accurate forecasting of river flows is one of the most important applications in hydrology, especially for the management of reservoir systems. To capture the seasonal variations in river flow statistics, this paper develops a robust modeling approach to identify and estimate periodic autoregressive (PAR) model in the presence of additive outliers. Since the least squares estimators are not robust in the presence of outliers, we suggest a robust estimation based on residual autocovariances. A genetic algorithm with Bayes information criterion is used to identify the optimal PAR model. The method is applied to average monthly and quarter-monthly flow data (1959-2010) for the Garonne river in the southwest of France. Results show that forecasts are better off in the robust model rather than the unrobust model. The accuracy of the forecasts is also improved when the model is specified in quarter-monthly flows, especially for the dry seasons.