Forecasting the Duration of Three Connected Wings in a Generalized Lorenz Model

Abstract

This work considers the problem of predicting wing changes, and their duration, in systems able to support three interconnected wings (spirals) in the space of the variables. This is done by exploring the alignment of covariant Lyapunov vectors (CLVs) known to precede the occurrence of peaks and regime changes in some chaotic systems. Here, the alignment of the CLVs is combined with classification procedures, machine learning techniques and nearest-neighbors analysis to predict the duration inside coupled wings. The classification procedure has distinct classes defined as the number of maxima of one variable inside one wing, proportional to the time spent inside each wing. In general, we observe that predictions of significant duration times (around [Formula: see text] oscillations) inside a wing can be efficiently predicted until four subsequently visited wings. For five to ten subsequently visited wings, the prediction efficiency decreases significantly for more than ten oscillations inside each wing. The numerical comparison shows the superiority of the nearest-neighbors methods over the multilayer perceptrons procedure for the predictions. Remarkably, accuracies larger than [Formula: see text] are obtained for predictions of classes for [Formula: see text] subsequently visited wings. Accuracies higher than [Formula: see text] are obtained for predictions of classes in the [Formula: see text] subsequently visited wings.