An integrated multistep prediction system based on wavelet filter analysis and improved instance based learning (IIBL)

Abstract

In this paper we present a novel wavelet based forecast model integrating wavelet filters for denoising and Improved Instance based learning approach. The proposed model implements a novel technique that extends the nearest neighbor algorithm to include the concept of pattern matching so as to identify similar instances thus implementing a nonparametric regression approach. A hybrid distance measure combining correlation and euclidean distance to select similar instances has been proposed. To illustrate the performance and effectiveness of the proposed model simulations using Mackey-Glass benchmark series and a real time Nord pool time series used in day-ahead forecast of electricity prices have been carried out. We apply a comprehensive set of non redundant orthogonal wavelet transforms for individual wavelet subband to denoise the signal. The analysis of simulations demonstrate that the proposed wavelet based - IIBL model results in accurate predictions and encouraging results for both the series.