Data-Driven Forecasting Schemes: Evaluation and Applications

Abstract

A reliable multi-step predictor is very useful to a wide array of applications to forecast the behavior of dynamic systems. The objective of this paper is to develop a more robust data-driven predictor for time series forecasting. Based on simulation analysis, it is found that multi-step-ahead forecasting schemes based on step inputs perform better than those based on sequential inputs. It is also realized that recurrent neural fuzzy predictor is superior to both recurrent neural networks and feedforward networks. In order to enhance the forecasting convergence, a hybrid training technique is proposed base on the real-time recurrent training and weighted least squares estimate. The developed predictor is also implemented for real-time applications in material property testing. The investigation results show that the developed adaptive predictor is a reliable forecasting tool. It can capture the system’s dynamic behavior quickly and track the system’s characteristics accurately. Its performance is superior to other classical data-driven forecasting schemes.