ELMVDP: extreme learning based virtual data position exploration and incorporation method for escalation of time series forecasting accuracy

Abstract

Time series data are correlated in a nonlinear fashion which makes the future data prediction challenging. Particularly, the correlation among data at the fluctuation points is insignificant and it is hard to capture the underlaying nonlinearity at those points by conventional prediction systems. The accuracy of time series forecasting (TSF) is vastly influenced by the current and immediate past data rather by far away data points. This article proposes an extreme learning-based method for exploration of virtual data positions (ELMVDP) from the training data and incorporates them to the original time series to intensify the TSF accuracy of a single hidden layer neural network. Specifically, this method is useful for the time series having less volume of data which may not suffice to train a TSF model. The effectiveness of ELMVDP method is evaluated on time series available in the literature, compared with few similar deterministic and stochastic approaches, and observations from simulation studies show that ELMVDP method yields better predictions than others.