Resumen
Product quality monitoring is an important element in industrial process control. In reality, most product qualities are difficult to be measured online due to technical or economic restrictions. Empirical inferential model is an effective solution to provide real-time analysis on product quality. Thus, this project analyzed different modelling techniques for C5 top composition of naphtha stabilizer column using historical plant data. Seven statistical and machine learning techniques were considered, which include multiple linear regression (MLR), stepwise linear regression (SLR), principle component regression (PCR), partial least squares regression (PLSR), regression tree (RT), gradient boosting regression tree (GBRT) and artificial neural networks (ANN). The aims of this project were to develop an inferential model of C5 top composition using the proposed modelling techniques, to compare the model prediction accuracy and to evaluate the limitations of each modelling technique for industrial application. The developed models were then assessed based on prediction accuracy and model development ease. The order of model performance obtained from this study was ANN (tangent sigmoid) > SLR > MLR > ANN (linear) > GBRT > PLSR > RT > PCR. Meanwhile, the overall assessment based on both model prediction accuracy and model development ease revealed that SLR was the most recommended modelling technique due to its reliable predictive performance and simplicity in developing an inferential model.