Abstract
With the goal of shaping a smart, economically-sustainable palm oil refining industry, this paper aims to improve the prediction modelling of Refined, Bleached and Deodorized Palm Oil (RBDPO) quality for early quality fault detection and remediation via the novel use of a moving windows form of prediction in addition to the conventional static window form. In this study, both Multiple Least Squares Regression (MLSR) and Partial Least Squares Regression (PLSR) model training techniques and prediction computations are carried out in the form of two moving data windows types, namely the expanding and rolling fixed-size data windows, together with the conventional static window form of prediction as control. Prediction improvement is observed consistently across both regression techniques when carried out in moving windows form, where both types of moving windows predictions, i.e. expanding and rolling windows, have fared significantly better than the static window form, with an average prediction error reduction of 20.6 % for Free Fatty Acid prediction, 55.9 % for Moisture Content prediction, 32.6 % for Iodine Value prediction and 34.2 % for Colour Value prediction in RBDPO. Among the moving windows themselves, the superiority of expanding windows over rolling windows and vice versa is insignificant. On the whole, this study paves the way for a revamped RBDPO quality prediction form which better reflects the dynamic, transient nature of the palm oil refining process, thus further consolidating its reliability for widespread practical use.
