Adaptive neuro-fuzzy inference system-genetic algorithm vs. response surface methodology: A case of optimization of ferric sulfate-catalyzed esterification of palm kernel oil

Abstract

Abstract Ferric sulfate-catalyzed esterification process for palm kernel oil (PKO), which had an initial acid value (AV) of 22 ± 0.1 mg KOH/g oil, was modeled using response surface methodology (RSM) and adaptive neuro fuzzy inference system (ANFIS). The process parameters investigated in the AV reduction of the oil were methanol-to-oil ratio (2:1–3:1), catalyst loading (6–10 w/v) and reaction time (15–25 min) using Box Behnken design of RSM. The developed ANFIS and RSM models were both subjected to various statistical evaluation and they both showed high degree of accuracy based on the high values of coefficient of determination ( R 2 ) of 0.9662 and 0.9039 for ANFIS and RSM, respectively and low values of mean absolute error of prediction (MAE) 0.0506 and 0.1506, and average absolute deviation (AAD) of 2.3665 and 7.1179 for ANFIS and RSM, respectively. To minimize the AV for the PKO, the process parameters investigated were optimized using RSM and ANFIS coupled with genetic algorithm (GA). Optimum values of methanol-to-oil ratio of 2.96:1, catalyst amount of 6 w/v and reaction time of 15 min with a corresponding AV of 1.05 mg KOH/g oil (95.2% AV reduction) were established using ANFIS-GA, while the values obtained using RSM were methanol-to-oil ratio of 2:1, catalyst amount of 6 w/v and reaction time of 25 min with a corresponding AV of 1.54 mg KOH/g oil (93.0% AV reduction). Based on the statistical indicators employed for this work, ANFIS was a better prediction tool than RSM while GA outperformed RSM in the optimization of the esterification process. Ferric sulfate proved to be a good catalyst for PKO esterification.