Wavelet based nonlinear autoregressive neural network to predict daily reservoir inflow

![CDATA[In spite of the ability of Artificial Neural Network (ANN) to handle nonlinear relationships in data, ANNs fail to predict with high accuracy in the presence of non- stationarity. Hydrological processes in nature exhibits non stationarity due to many interrelated physical and other interrelated factors such as chaotic weather conditions. This paper presents the modelling of one such hydrological process, inflow, to the top most reservoir in the major cascaded reservoir system in Sri Lanka. This daily inflow series has been investigated to be nonlinear and nonstationary. Thus, the difficulties encountered in modelling the inflow series is addressed through a pre-processing strategy based on wavelet transform. Among the methods available in dealing with the nonstationary nature, the wavelet transform was used due to its ability to determine the frequency content of the signal and to assess and determine the temporal variation of this frequency content. The inflow series is decomposed in to several sub series using discrete wavelet transform (DWT). Consequently the appropriate sub series resulted through the wavelet transform are used to model the original inflow using Nonlinear Autoregressive Artificial Neural Network with Exogenous Inputs (NAR-ANN). The results of the NARANN with modified inputs are compared with the results of the base model i.e. NAR-ANN with raw inputs as well as a previously fitted cluster based modular NAR-ANN. The results confirms the superiority of the wavelet based approach over the other approaches, as it has the ability of capturing useful information on various resolution levels.]]