(546g) Recent Advances and New Directions for Optimization of Production Scheduling in Crude-Oil Refineries

There is an increasing innovation and improvement in production scheduling of crude-oil refineries. Commercial solvers for optimization of industrial problems are progressively robust and efficient; the processing capabilities of computers has evolved tremendously with concurrent processing; techniques and problem-solving strategies are constantly being created and improved. These factors contribute to achieve more efficient production schedule (in operational and economic point of view) and to better integrate wider scale and scope of refining processes. In this way, several studies in the literature addressed techniques: a) to reduce the complexity of the problems by fixing, excluding or relaxing of part of the problems; b) to facilitate the implementation of the problems (programming) by the schedulers or final users in the plant site; c) to reduce the differences between the model and the plant (better prediction) by using closed-loop parameter feedback; d) to access broader and higher quality data, etc. The intention of this work is to cover the recent advances in modeling and solving aspects related to the crude-oil refinery scheduling and give an overview of the new directions in such technology.

Within the modeling aspects, in the last decades, most of works found in the literature for production scheduling optimization in crude-oil refineries have chosen not to use discrete-time modeling due to the combinatorial complexity of the problem. However, considering their selection of continuous-time, the main disadvantages of this approach are the difficulties of execution of tasks by the operators in the plant and the needs of defining previously the points to divide the time horizon. Today, with advances in computer processing, commercials solvers for optimization (e.g., GUROBI), and problem-solving techniques, automated decision using the discrete-time approach has become feasible within reasonable processing time (CPU) and might be used in large-scale problems involving crude-oil processing [1].

In addition to the technical advances mentioned previously, there are new directions in the development of heuristics to be employed in intuitive aspects of the crude-oil refinery scheduling. These methods include a fixation strategy by considering the feedstock storage assignment before a decomposed formulation that integrates the scheduling in a mixed-integer linear (MILP) problem by testing good local and the global optima in a pool of nonlinear (NLP) problems considering, among other techniques, randomized initial point of the nonlinear variables. The MILP construct the time-varying solutions of setups of unit-operations and their connections for the NLPs. The pre-scheduling feed assignment strategy proposed by Kelly et al. [2], which consists of pre-defining the selection of tanks for storage of crude-oils in the refinery, clustering those with similar characteristics by minimizing the bulk crude-oil qualities when different oils are sharing storage tanks. The MILP-NLP decomposition strategy named Phenomenological Decomposition Heuristic (PDH), proposed by Menezes et al. [3] for strategic planning, reduced the complexity of the problem and enable optimization within minutes (< half hour). In this strategy are optimized, iteratively, an MILP model (or logistics problem) considering only logic and quantity variables and an NLP model (or quality problem) considering only quantity and quality variables, setting the optimized binary variables in the MILP solution.

Kelly et al. [1] solved the MILP-NLP scheduling optimization for a time horizon of seven days considering time-periods of two hours. The authors pre-defined feedstock storage assignment [2] to the fixed in the further MILP-NLP scheduling. The modeling of the industrial-scale crude-oil refinery scheduling problem uses the UOPSS (Unit-Operation-Port-State Superstructure) [4] [5]. The problem includes 5 atmospheric distillation units in 9 operational modes and 35 tanks among storage and feeding/charging tanks. The logistic problem (MILP), with approximately 30 thousand continuous and 30 thousand binary variables, 6.5 thousand equations and 80 thousand inequalities (53,802 degrees of freedom) was solved in 128.8 seconds using 8 multi-tasks (threads) in the CPLEX 12.6 solver. The quality problem (NLP), with 102,539 continuous variables, 58,019 equality and 768 inequality constraints (totaling 44,520 degrees of freedom) was solved in 10.3 minutes in the IMPL 'SLP tool using CPLEX 12.6 solver. The MILP-NLP gap between the two solutions was 3.5% after two PDH iterations. However, the authors did not consider the complete network of process-shops and blend-shops beyond the distillation unit.

Despite numerous studies of crude-oil refining scheduling during the last two decades of literature, only one covered the complete refinery network including process units downstream to the distillation unit. Xu et al. [6] included refining operation units such as reforming, cracking, hydrotreating, as well as blending (mixing of products and additives) and fuel storage. However, there is a relatively small number of variables in the continuous-time model for a time horizon of two weeks (<100 binary variables), making it difficult to use for real cases. Franzoi et al. [7] included the oil storage and feeding steps, the distillation unit modeled as cascade towers and the refinery set of process-shops. The authors optimized the problem with about 3,000 binary variables and 4,000 continuous variables for a time horizon of 5 days with periods of 4 hours. In addition, they used a strategy of online parameter feedback to cope with uncertainties and to reduce the offsets or inaccuracies over the life-time of the problem as an effective way to close the gap between predictions and productions.

Using, as example, 4-hour as time-step, it pushes forward to a faster and more efficient scheduling, considering this approach in the direction to an online scheduling engine. According to Gupta et al. (2016a), new information may become available over time and should be used as soon as possible, as may occur changes that make the current schedule suboptimal or even infeasible, such as process interruptions, task delays, oscillations in product demand. In this way, the execution of an online scheduling aims to improve the adaptation to uncertainties as well as computing new information within the scheduling re-optimization, which implies in a potential reduction in costs and profit increase for the refinery production. Besides, according to Gupta et al. [8] [9], rescheduling should be performed not only when necessary, but regularly, since it considers new information and contribute to further cost reduction. Recent studies involving re-scheduling strategies were developed by Zhang et al. [10] for the front-end problem in a refinery, where production is re-scheduled whenever there is a malfunction in a tank (due to mechanical problems); and by Pattison et al. [11] who created a mechanism to re-schedule periodically an air separation unit to compute new information regarding prices and demands; or when there are process or market-related disruptions.

In addition, according to Kelly and Zyngier [12], a continuous cycle of improvement is required to provide process feedback and reduce the deviation between model prediction and actual plant values. Thus, using data and information measured with the use of automated systems significantly improves the reliability of the model, especially using techniques of simultaneous parameter estimation and data reconciliation to improve the quality of the information to be updated.

[1] KELLY, J. D.; MENEZES, B. C.; GROSSMANN, I. E.; ENGINEER, F. Crude-oil Blend Scheduling Optimization of an Industrial-Sized Refinery: a Discrete-Time Benchmark. In Foundations of Computer Aided Process Operations, Tucson, United States, 2017a.

[2] KELLY, J. D.; MENEZES, B. C.; GROSSMANN, I. E.; ENGINEER, F. Feedstock Storage Assignment in Process Industry Quality Problems. In Foundations of Computer Aided Process Operations, Tucson, United States, 2017b.

[3] MENEZES, B. C.; KELLY, J. D.; GROSSMANN, I. E. Phenomenological decomposition heuristic for process design synthesis of oil-refinery Units. Computer Aided Chemical Engineering, v. 37, p. 1877-1882, 2015b.

[4] KELLY, J. D. The Unit-Operation-Stock Superstructure (UOSS) and the Quantity-Logic-Quality Paradigm (QLQP) for Production Scheduling in the Process Industries. In Multidisciplinary International Scheduling Conference Proceedings, New York, United States, v. 327, 2005.

[5] ZYNGIER, D.; KELLY, J. D. UOPSS: A New Paradigm for Modeling Production Planning and Scheduling Systems. In European Symposium in Computer Aided Process Engineering, London, United Kingdom, 2012.

[6] XU, J.; ZHANG, S.; ZHANG, J.; WANG, S.; XU, Q. Simultaneous scheduling of front-end crude transfer and refinery processing. Computers & Chemical Engineering, 2017.

[7] FRANZOI, R.E.; MENEZES, B.C.; KELLY, J.D.; GUT, J.W. Effective scheduling of complex process-shops using on-line parameter feedback in crude-oil refineries, Process Systems Engineering, 2018.

[8] GUPTA, D.; MARAVELIAS, C. T.; WASSICK, J. M. From rescheduling to online scheduling. Chemical Engineering Research and Design, v. 116, p. 83-97, 2016a.

[9] GUPTA, D.; MARAVELIAS, C. T. On deterministic online scheduling: major considerations, paradoxes and remedies. Computers & Chemical Engineering, v. 94, p. 312-330, 2016b.

[10] ZHANG, S.; WANG, S.; XU, Q. A New Reactive Scheduling Approach for Short-Term Crude Oil Operations under Tank Malfunction. Industrial & Engineering Chemistry Research, v. 54, n. 49, p. 12438-12454, 2015.

[11] PATTISON, R. C.; TOURETZKY, C. R.; HARJUNKOSKI, I.; BALDEA, M. Moving horizon closed‐loop production scheduling using dynamic process models. AIChE Journal, 2016.

[12] KELLY, J. D.; ZYNGIER, D. Unit-operation nonlinear modeling for planning and scheduling applications. Optimization and Engineering, v. 18, p. 133–154, 2017.