Interval predictor based on a Reversed Huber's error function

Abstract

In dynamical systems context, a predictor is a method that provides an estimation of the future system output using past information of the system. An interval predictor provides an outer estimation of the future output. The center of this interval can be used as central or nominal prediction. A method to formulate interval predictors is to assume an unknown but bounded error in the system measurements. The aim of this work is to study the benefits of using a Reversed Huber's function as error function in this kind of predictors. A Reversed Huber's function is a convex function, piecewise linear near zero but quadratic for large values. The paper provides a nonparametric formulation of the interval predictor and shows by a real world example that the proposed predictor can improve the performance of the central prediction.