Effect of Normalization Techniques on Univariate Time Series Forecasting using Evolutionary Higher Order Neural Network
Sibarama Panigrahi1, H. S. Behera2
1Sibarama Panigrahi, Computer Science and Engineering, MIRC Lab, Majhighariani Institute of Technology and Science, Rayagad, India.
2H. S. Behera, Computer Science and Engineering, Veer Surendra Sai University of Technology, Burla Odisha, India.
Manuscript received on November 27, 2013. | Revised Manuscript received on December 13, 2013. | Manuscript published on December 30, 2013. | PP: 280-285 | Volume-3, Issue-2, December 2013. | Retrieval Number: B2493123213/2013©BEIESP
Open Access | Ethics and Policies | Cite
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Over the last few decades, application of higher order neural networks (HONNs) to time series forecasting have shown some promise compared to statistical approaches and traditional neural network (NN) models. However, due to several factors, to date, a consistent HONN performance over different studies has not been achieved. One such factor is preprocessing of time series before it is fed into HONN models. Normalization is one of the important pre-processing strategies which have a significant impact on forecast accuracy. Despite its great importance, there has been no general consensus on how to normalize the time series data for HONN models. This paper investigates how to better normalize the univariate time series for HONN models especially, the Pi-Sigma network (PSN). For this five different normalization technique (Min-Max, Decimal Scaling, Median, Vector and Z-Score) are used to normalize four univariate time series and corresponding forecast accuracy are measured using an evolutionary Pi-Sigma network. Results show that forecast accuracy using HONN models depends on the normalization technique being used. It is also noted that with PSNs, decimal scaling and vector normalization techniques provide statistically meritorious results compared to other normalization techniques.
Keywords: Normalization, Higher Order Neural Networks, Pi-Sigma Network, Differential Evolution, Time Series Forecasting.