A machine learning based predictive model for time-series modelling and analysis. Issue 3 (25th October 2021)

• Record Type: Journal Article
• Authors: Al-Haija, Qasem Abu
• Is Part Of: International journal of spatio-temporal data science. Volume 1:Issue 3(2021)
• Page Start: 270
• Page End: 283
• Publication Date: 2021-10-25
• Subjects: data science -- time-series -- regression -- forecasting -- neural network -- generalised delta rule -- NARX model -- autocorrelation function (AFC) -- cross-correlation function (CCF) -- prediction accuracy; Big data -- Periodicals; Data mining -- Periodicals; 005.7
• Languages: English
• ISSNs: 2399-1275

Abstract

Time series modelling and forecasting is an essential field of supervised machine learning because of its appreciated contributions into numerous research and real-life applications including the corporate, commercial, science and engineering applications. Therefore, substantial contributions have been devoted to developing competent predictive models. In this paper, we propose an inclusive time-series predictive model using two modelling techniques, namely; multi-layer feed forward neural networks (FFNN) based delta learning rule model and nonlinear auto-regression neural network (NARX) based external input model. The developed models have been trained with least possible prediction error for the 10th order one step ahead predictor for FFNN model and the 50th order two-step ahead predictor for NARX model. As a case study, we have employed the stationary time-series for yearly averaged sunspot activity during the period from 1719-2018. To evaluate the performance of the predictive models, the models have been trained for more than 1000 epochs and have scored the maximum prediction accuracy of more than 99% after 405 epochs recording a mean square error of (2.2~6.5)× 10 −2 for the training process. Eventually, the proposed models are considered comparative predictive model for any stationary time-series in several areas of study.