Comparing Kriging, Spline, and MLR in Product Properties Modelization: Application to Cloud Point Prediction

In most chemometrics applications concerning the prediction of physicochemical properties, regression models are preferred, because they are easy to implement and their posterior analysis provide simple interpretations. Interpolation methods are most currently used in geostatistics applications where the dimension of study area is generally limited. In this work, we proposed to develop kriging or splines models for predicting the properties of petroleum products. Kriging and splines have different foundations, because the former is based on stochastic assumptions and the latter is built on deterministic approach. A well-illustrated comparison of both methods was carried out through three suitable examples to highlight their similarities and their divergences. The advantages of using kriging or/and splines instead of classical regression models were also discussed. Results indicated the flexibility of interpolation methods, as they provide good accuracy for linear and nonlinear cases. They also confirmed previous studies which pointed out equivalence between kriging and spline models performances in some situations. However, the kriging approach has more valuable aspects, relative to other interpolation methods, since it provides a measure of prediction uncertainties. Kriging modeling were finally compared to multilinear regression for the prediction of diesel cloud point ranging from −39 °C to −12 °C. Models performances noted that kriging enables one to improve both accuracy and robustness.