Incomplete time-series data forecasting based on clustering fill-in technique and ensembling neural network model

Abstract

This dissertation demonstrates the problem of incomplete time-series prediction by modelling the forecasting of several natural and social phenomena. The modeling consists of two main steps. The first step is to estimate the collected incomplete data, which are considered as missing data or missing values. The second step is to predict new data based on the nature of the data obtained from the first step. Our solution is to develop a new neural network model for forecasting incomplete time-series data and improving the accuracy of prediction. Two neural network models are proposed. First, various versions of EM-based algorithm and smoothing spline interpolation are used to preprocess the incomplete data sets. The individual networks are trained by supervised multilayer perceptron(MLP) with extended Kalman filtering. The ensemble construction is used for the combination of the individual networks. We name this type of network Fill-in - Generalized Ensemble Method (FI-GEM) networks. Second, each individual network uses a Finite Impulse Response model to perform the prediction. The outputs of all individual neural networks are combined by the genetic algorithm-based selective neural network ensemble method (GASEN). We denote this network as a reconstructed missing data-finite impulse response selective ensemble (RMD-FSE) network. Moreover, we proposed a new fill-in technique that is improved for estimating missing values based on clustering technique for characterizing the pattern of incomplete time-series data. The main idea is the time-series data are divided into separate subsequences of different sizes and, therefore, each subsequence can be viewed as a window. The imputation of missing samples is achieved by finding a complete subsequence similar to the missing sample subsequence and imputing the missing samples from this complete subsequence. The imputation accuracy of the proposed algorithm, namely varied window clustering (WDC) algorithm is comparable or better than the others traditional methods such as: the spline interpolation, the multiple imputation (MI), and the optimal completion strategy fuzzy c-means algorithm (OCSFCM) in case of the non-stationary time-series data.