Abstract
Palm oil is a highly efficient feedstock for large-scale biodiesel production, as it yields significantly more oil per agricultural area than other common oil crops, such as soybeans or sunflowers. However, crude palm oil often faces a high level of free fatty acid (FFA) problems, hindering biodiesel production. Pretreatment, such as esterification, is thus employed to convert FFA to fatty acid methyl ester (FAME) and avoid undesired byproducts. Response surface method (RSM) has been widely and effectively used to optimize the pretreatment conditions. To address this challenge, raw palm oil (a mixture of palm stearin and palm fatty acid distillate) with the initial FFA of 3-90% was used in the pretreatment process. A four-factor-three-level Box-Behnken experimental design was deployed to estimate the final FFA as a function of reaction time (0.5-4 h, X(1)), molar ratio of methanol to FFA (3:1-24:1, X(2)), catalyst (0.5-8 wt % based on FFA, X(3)), and initial FFA (3-90%, X(4)). Three different mathematical models were obtained and validated over different ranges of FFA in palm oil. The optimum conditions were 2.73 (X(1)), 22.02:1 (X(2)), 3.90 (X(3)), and 21.88 (X(4)) for 3-30% FFA; 2.34 (X(1)), 16.57:1 (X(2)), 3.12 (X(3)), and 60.00 (X(4)) for 30-60% FFA; and 2.40 (X(1)), 16.05:1 (X(2)), 3.12 (X(3)), and 90.00 (X(4)) for 60-90% FFA, respectively. After validation, the results showed that palm oil at 1-30 and 60-90% FFA gave fewer errors of 0.60 and 0.58% respectively, than the other ranges of 30-60% at 1.25%. Therefore, a machine learning approach was used to improve the optimum conditions, comparing decision tree, random forest, and gradient boosting. It was found that the decision tree gave the highest R (2) of 0.9762, RMSE of 1.2130, and MAE of 0.4070. The new optimum conditions from the predictive model of 3-90% FFA were 2.25 (X(1)), 15:1 (X(2)), 11.5 (X(3)), and 46.5 (X(4)) via a gradient boosting model with the least percentage error to obtain %final FFA less than 1%.