Optimal schedule generation for single-channel crude transfer using a multi-model approach. (April 2022)

• Record Type: Journal Article
• Authors: Paranjape, Aditya A.; Baranwal, Mayank; Wagle, Satyavrat; Lotti, Rushi; Majumder, Sushanta; Bullière, Anne-Laure
• Is Part Of: Computers & chemical engineering. Volume 160(2022)
• Publication Date: 2022-04
• Subjects: Crude transfer scheduling -- Reinforcement learning -- Optimal control -- Mixed-integer linear programming; Chemical engineering -- Data processing -- Periodicals; 660.0285
• DOI: 10.1016/j.compchemeng.2022.107732
• Languages: English
• ISSNs: 0098-1354

Abstract

Highlights: The paper builds a multi-model approach for scheduling the operations of a refinery under tight operating constraints. The multi-model approach involves different degrees of temporal and spatial aggregation. We present multiple scheduling techniques in the paper: meta-heuristics or business rules, mixed-integer linear programming (MILP) and reinforcement learning (RL). The specific refinery model that we consider in the paper is amenable to an ILP-based solution, but the solvers based on business rules and RL offer the possibility of extension to refinery models where the extent of nonlinearity might prevent us from employing ILP reliably. In the present paper, ILP offers a baseline for assessing the two other solvers. Our innovation, which promises to be useful in practice, involves combining two or more approaches (e.g., RL + business rules or RL + MILP) in a hierarchical manner using the multi-model abstraction. The objective is not only to improve the quality of the solution, but also to enable the scheduling algorithm to deal with uncertainties and retain some degree of explainability. The problem addressed in the paper is based on the requirements of a real refinery. The model considered in the paper is sanitized (due to non-disclosure agreement), but the challenges identified and tackled in the paper are realistic. Abstract: This paper presents three techniques for scheduling for crude transfer between a port and a refinery on a single pipeline in the presence of stringent flow constraints. The three techniques are based on metaheuristics (business rules), mixed integer linear programming and reinforcement learning. In addition to comparing the three techniques, we also demonstrate how knowledge gleaned from one technique (in our case, the metaheuristics) can be used to design an algorithm based on another technique (in our case, reinforcement learning). A novel feature of our approach to reinforcement learning, in particular, is the use of low-fidelity, reduced-order simulators for training the scheduler and supporting it with a post-processor based either on business rules or on integer programming for ensuring compatibility with the constraints. The set of constraints considered here includes temporal restrictions on the use of the pipeline, flow constraints in the tanks that feed the column distillation units in the refinery, and the need to ensure a certain minimum residence time for crude in a given tank for dewatering.