Knowledge-Informed Semi-Supervised Robust Modeling for Mixed Oil Length Prediction in Multi-Product Pipelines Based on Student’s-T Mixture Model

Abstract Multi-product pipelines usually transport various petroleum products in certain batch, leading to inevitable mixed oil segment. Current soft sensor modeling methods for predicting mixed oil length face challenges characterized by nonlinearity, scarcity of labeled samples and interference from outliers. We propose a semi-supervised robust soft sensor called the KISSVBSMR. It employs the Semi-supervised Variational Bayesian Student’s-t mixture regression to empower the usage of unlabeled data, and it strategically selects the key variable based on physical significance to enhance mode identification. Additionally, by transforming the relationship between input and output variables derived from theoretical equation and incorporating a Bayesian approach, the KISSVBSMR can explore the prior knowledge regarding the regression coefficient and efficiently against outliers. Evaluation through numerical and real-life industrial datasets reveals the shortcomings of traditional methods and the superiority of KISSVBSMR in model performance. This novel approach improves the predictive accuracy of mixed oil lengths, optimizing multi-product pipeline operations and advancing knowledge-data hybrid modeling techniques.