Quantifying the Predictability of Winter River Flow in Iberia. Part I: Interannual Predictability

Abstract

The interannual variability and predictability of the winter streamflow of the main Iberian Peninsula international rivers (Douro, Tejo, and Guadiana) are examined for the period 1923?2004. In the first part of this paper, a singular spectral analysis was carried out to isolate the main oscillatory components of the streamflow series. Results showed a similar model structure for the three rivers, including the following components: (i) a nonlinear trend that contains variability at periods of 20?30 yr, (ii) modulated amplitude oscillations with associated periods in the bands 2?3, 4?5, and 6?8 yr, and (iii) a red noise process. These models accounts for the bulk of winter river flow variance, ranging between 64% (Guadiana) and 96% (Douro). In general, the amount of variability associated with the low-frequency component is similar to that associated with the interannual variability. The analysis of the association between the North Atlantic Oscillation (NAO) and the streamflow variability proved this relationship to be complex and nonstationary. In particular, it is found that only when the NAO presents high amplitude oscillations is this mode capable of dominating the streamflow variability. In Part II, autoregressive-moving-average (ARMA) models were fitted to the filtered streamflow series and an interannual forecasting experiment was conducted. Results were tested against the raw streamflow series. The percentage of variance explained by the models ranged from 25% to 62%. Additionally, the ARMA models presented useful one-year-ahead forecasting skills. Particularly during the validation period (1986?2004) the models performed between 51% and 53% better than climatology. The skill against persistence proved to be much greater, indicating that the climatology is a better benchmark than persistence for streamflow forecasting in Iberia. Finally, the developed models were, in most cases, able to accurately predict the phase of the streamflow, with a percentage of agreement that ranged from 54% to 90% throughout the validation period.