Nonlinear analysis and prediction of river flow time series

In this paper we consider the problem of nonlinear modelling the discharge time series of a river in order to study the forecasting ability of a nonlinear approach. To this aim, we first check for some evidence of chaotic behaviour in the dynamics by considering a set of different procedures (phase portrait of the attractor, correlation dimension, the largest Lyapunov exponent, DVS diagram). Their joint application to our data allows us not to exclude the presence of a nonlinear deterministic dynamics of chaotic type. Secondly, we consider two kinds of nonlinear predictors: a univariate predictor, which is based only on the information of the discharges times series and a multivariate one, which also takes into account the information coming from rainfall data. By comparing these predictors with a linear predictor, we can conclude that nonlinear river flow modelling is an effective method to improve prediction in a statistically significant way.