Novel automatic model construction method for the rapid characterization of petroleum properties from near-infrared spectroscopy

Petroleum fuels play an important role in economic society, and near-infrared analysis has been widely used in the characterization of petroleum fuels due to its effectiveness and efficiency. However, near-infrared spectra are high dimensional data that require high computational cost, thereby complicating the practical application. Previous model construction methods are also restricted to pre-designed model structure, thus limiting the model construction ability. To solve these problems, this paper proposes an automatic model construction algorithm that increases the diversity of model structure with linear-in-parameter representation and optimizes the model structure with genetic programming. The proposed automatic construction algorithm is verified on datasets from literature and experiments with real petroleum products compared with Partial Least Square (PLS) and Support Vector Machine Regression (SVR). The proposed method outperforms PLS and SVR in overall predictive accuracy and reduces the involved variable number by at least 70% during the model construction process. Good performance in prediction accuracy, model complexity, and interpretability on the rapid characterization of different petroleum products can assist refinery plant in product control and management during production.

Automated summary

This paper proposes an automatic model construction algorithm that increases the diversity of model structure with linear-in-parameter representation and optimizes the model structure with genetic programming. The proposed method outperforms traditional methods in predictive accuracy and variable reduction, and can assist refinery plants in product control and management during production.