Refined bleached deodorized palm oil quality prediction using multivariate statistical process control tools

Abstract

Multivariate statistical process control (MSPC) has been widely used for quality prediction and monitoring in palm oil refinery processes. Currently, the refined, bleached deodorized palm oil (RBDPO) quality is determined based on the relationship between crude palm oil quality and process parameters, with the assumption that the process is static and not affected by the time-varying characteristic of the palm oil refinery process. However, the prediction is less accurate since the generated regression coefficients from static prediction models do not reflect the current process status and remain constant over time. Therefore, this study was conducted to introduce a new framework for regression coefficients improvement via dynamic prediction models. The dynamic prediction models were developed by integrating the MSPC prediction tool with time-series expansion methods where the prediction models were adapted to new process dynamics. Data collected from an industrial palm oil refining plant were used as the case study in this research. Four MSPC models, namely linear principal component regression (PCR), linear partial least squares (PLS), nonlinear principal component regression based on nonlinear iterative partial least squares algorithm (NIPALS-PCR) and nonlinear partial least squares based on nonlinear iterative partial least square algorithm (NIPALS-PLS) were used to determine the relationship between the quality and process variables. Time-series expansion methods were used to trace the dynamic behaviour based on five approaches, namely static, moving window (MW), recursive window (RW), exponentially weighted moving window (EWMW) and exponentially weighted recursive window (EWRW). The findings show that the combination of the linear prediction model with the time-series expansion method showed a more reliable prediction performance than the nonlinear prediction model. The performance of the PCR EWMW model in predicting the RBDPO quality is improved by 12.02 % (11.96 % for free fatty acid, 6.92 % for moisture content, 16.13 % for iodine value and 13.01 % for colour) compared to other prediction models. The sensitivity of the regression coefficients was also improved where the regression coefficients fluctuated very smoothly and showed high convergence to zero value when using the PCR EWMW model. This shows that the implementation of the linear dynamic prediction model was better than the static prediction model. Therefore, the linear dynamic prediction model for quality prediction was the best for it has the greatest prediction improvement and showed a better trend of the regression coefficient.