Resumen
Hydrological series are largely characterized by a strong random component in their behavior. More noticeable changes in the behavior patterns of rainfall/runoff temporal series are recently being observed. These modifications are not a trivial issue, especially in regards to peculiarities, non-linearities, diffused influences or higher time orders of dependence. This study mainly aimed to analyze the temporal dependence of an annual runoff series dynamically. This approach comprises a coupling between classic techniques (Autoregressive Moving Average Model, ARMA) and novel ones, based on Artificial Intelligent for hydrological research (Bayesian Networks, BNs). An ARMA model was built to provide reliable data to populate BNs. Then, causal reasoning, through Bayes?s theorem, allows the identification of the logic structure of temporal dependences within time series. Furthermore, the resultant conditional probability permits the quantification of the relative percentage of annual runoff change, and provides the right time order of dependence. This research introduces an original methodology able to build a logic structure for a stochastic analysis of temporal behavior. This approach also aimed to provide a powerful and graphic modeling method for improving the understanding of the dynamic runoff series temporal behavior. This was successfully demonstrated in two unregulated river basin stretches, belonging to the Duero river basin which is the largest basin in Spain.