(91d) Application of Data Driven Techniques to Industrial Hydroprocessing Units

Hydoprocessing units in petroleum refineries play a critical role in removing impurities from the crude and crack the heavier crude to form lighter products for subsequent operations. These units have an intricate network of several unit operations and modeling such units play a pivotal role in estimating/predicting values of important variables, improving the control and optimization of the plant for efficient operation among several other applications. This presentation demonstrates the development and implementation of linear dynamic data-based models in estimating product qualities and other key monitoring variables in the hydroprocessing unit of an industrial refinery. Historical operational data from two different units was used and appropriate data-driven modeling strategies were formulated to address this problem. The usefulness of Dynamic-Partial Least Squares (DPLS) [1] and Subspace Identification [2] in the estimation of important variables of the units is demonstrated. In this study, the Subspace Identification methodology was suitably adapted and made more generalizable to handle missing data in both inputs and outputs. The Non-linear Partial Least Squares (NIPALS) algorithm, being robust to missing data, was used to replace some of the conventional steps in the Subspace Identification. The identified linear time invariant (LTI) model was then embedded inside a Kalman filtering state estimation technique to make use of the current process measurements to correct the model states as the process keeps marching along in time. The methods used in this study not only elegantly handle issues like missing data problem associated with real data sets but also conveniently address the problem of overfitting associated with data models, and the results demonstrate the suitability of using these linear dynamic models for specific problems in the context of hydroprocessing refinery. The results are a significant improvement over the current modeling techniques used in the industry where the units are operated and suggests practitioners the usefulness and potential application of these models.

[1] Wenfu Ku, Robert H. Storer, Christos Georgakis, Disturbance detection and isolation by dynamic principal component analysis, Chemometrics and Intelligent Laboratory Systems, Volume 30, Issue 1, 1995, Pages 179-196, ISSN 0169-7439

[2] Patel, N, Mhaskar, P, Corbett, B. Subspace based model identification for missing data. AIChE J. 2020; 66:e16538.