Modelo híbrido EMD-RNA de previsão de cotas fluviais médias diárias em Manaus a partir de informações climáticas de larga escala

Abstract

This work compared the performance of Hybrid Empirical Mode Decomposition (EMD)-artificial neural network (ANN) models and conventional ANN models in forecasting annual extreme water stages that routinely penalizes Manaus, on the left bank of Negro River. Several studies reported significant performance gains using Hybrid EMD-ANN models. Nevertheless, the transitoriness of Intrinsic Mode Functions (IMFs) and the forecasting performance decline, comparing with training and hindcasting stages, where observed drawbacks, yielding the present investigation. First, an integrated oceanographic, meteorological and Amazon hydrological review is conducted, exploring their relations to define the predictors of the models. The subtleties of the tropical environment produce an atmospheric characterized by the duality between self-affinity and hierarchic scales. To deal with such peculiarities, the Detrended Fluctuation Analysis (DFA) was applied, revealing fractal relations between the SSTs, the Long Wave Radiation (OLR) on the Amazon and the levels in Manaus. In addition, Wavelet Coherence (WTC) was used to provide multi-scale insight via time-frequency spectra. Those analysis enabled the predictors selection and encouraged a discussion about the causes of intensification of events in Manaus in the last decades. Hybrid models gains where superior to 40% and 35% in high and low annual peaks forecasting, respectively, but revealed higher residuals variance, encouraging further investigations.