Multiscale characterization of Hydroclimatic Variables A Hybrid decomposition Based Artificial Intelligence approach for Rainfall prediction

Abstract

Prediction of non-linear systems like rainfall time series is of great concern among the meteorologists as it is a crucial step in efficient water resources planning and management. Rainfall time series data possess non-linear and non-stationary characteristics and therefore,for the prediction of rainfall the prime investigation is to be done on the characteristics of the climatic variable time series that influences rainfall. The traditional spectral analysis tools such newlineas Fourier Transform can handle only linear and stationary data and therefore is inappropriate for the characterization of rainfall. Later, Wavelet Transforms(WT) are found to be the better substitute with additional capability of capturing the non-stationarity of the signals that offers newlinebetter understanding of the spectral properties of the time series in time frequency domain. But the requirement of apriori basis in the selection of wavelet type and in the number of decomposition levels makes its implementation a difficult task. Empirical Mode Decomposition(EMD) is a data adaptive signal processing method that can capture the properties of non-linear and non-stationary signals and is therefore more appropriate for the characterization of rainfall time series. The method performs a data adaptive multiscale newlinedecomposition process which produces a set of oscillatory modes called Intrinsic Mode newlineFunctions (IMFs) with specific periodic scales for each IMFs and one residue component. The original time series can be obtained by summing up all the independent modes and residue. EMD make use of Hilbert Transforms that derives the instantaneous frequency and amplitudes of each IMFs for processing the spectral characterization. The primary objective of the present newlineresearch work is to understand the teleconnection between the climatic oscillations and rainfall. This is possible by spectral characterization on utilizing the decomposition technique and eventually use the information for improved prediction of the rainfall time series.