Constrained Operational Optimization of a Distillation Unit in Refineries With Varying Feedstock Properties

Abstract

This paper studies the challenging operational optimization problem of a distillation unit under varying feedstock properties (e.g., density and carbon content). This problem, in which changes in the feedstock properties are incorporated, aims to quickly obtain the operating variables that control the operating condition of the distillation unit. To solve this problem, we first model this operational optimization problem considering the ever-changing feedstock properties and practical technological constraints. Then, we propose an efficient soft-sensing strategy to rapidly measure the feedstock properties. Finally, motivated by the challenges caused by the varying feedstock properties, product yield and tray temperature constraints, we propose an optimization algorithm with global search and self-repair capabilities to optimize the operating variables of the distillation unit. The proposed algorithm integrates the optimization time and survival information of each individual into the proposed mutation strategy to improve its global search capability in the irregular feasible region of the operating variables. Based on the ranking and survival information of each individual, the adaptive strategies of the mutation factor and crossover probability are designed to balance the exploration and exploitation capabilities of the optimization algorithm. Subsequently, we propose an effective correction strategy to correct the infeasible operating variables and improve the optimization efficiency of the algorithm. Computational experiments on practical production data show the accuracy of the soft-sensing model and the superiority of the optimization algorithm for operational optimization of the distillation unit.

Research on Product Yield Prediction and Benefit of Tuning Diesel Hydrogenation Conversion Device Based on Data-Driven System

“…The first step in the proposed model is to collect plant data from the refinery, including feedstock characteristics, operating conditions, and product yields [8]. The data are then preprocessed, including outlier removal, data transformation, and normalization.…”

“…In multi-objective optimization, the problem is decomposed into multiple single-objective functions, which are then combined as a weighted sum. The weights assigned to individual objective functions can be adjusted based on their relative importance, as shown in Equation (8).…”

“…This helps to avoid strong coupling between process conditions, thereby improving prediction accuracy [7]. Chen and his colleagues adopted artificial neural networks (ANNs) to learn the sum of manipulative variables, such as pallet temperature, for optimizing ANN models [8]. Meanwhile, Song and his colleagues used polynomial models to model production and industrial processes [7].…”

Adaptive Weighted Hybrid Modeling of Hydrocracking Process and Its Operational Optimization

“…Recently, Long et al developed multiple FNNs to simulate the fluid catalytic cracking process . For the optimization of a distillation unit, Chen et al adopted an ANN to learn the relationship between operating variables and tray temperatures . Song et al developed a deep learning framework integrating the self-organizing map (SOM) and the convolutional neural network (CNN) for modeling the industrial HCUs to obtain better fitting and extrapolation performances .…”

Multisource Latent Feature Selective Ensemble Modeling Approach for Small-Sample High-Dimensional Process Data in Applications

“…Different input features also relate to various hyperparameters of the SEN model and its ensemble submodels. Multiobjective evolutionary can be employed to address this problem [32,33,34]. However, an effective joint optimization strategy can also solve the above problems.…”