Application of metaheuristics in multi-product polymer production scheduling: A case study

Abstract

Chemical manufacturers produce a range of polymer product families on a large scale within complex and dispersed manufacturing plants. These plants are connected through pipelines and are highly dependent on each other. Such a group of connected plants is referred to as a value chain. The process involves the continuous flow of raw materials known as feedstock, from one plant to another, enabling the continuous production of polymers. When the flow of feedstock through the value chain is interrupted due to unexpected events or planned maintenance at a plant, it results in irreversible losses. Consequently, prompt decision support is required to manage these disruptions and ensure the continuity of the flow. This paper evaluates several metaheuristics for effectively scheduling the flow between the plants within the value chain of a chemical manufacturer. These metaheuristics aim to provide near-optimal solutions after disrupting events occur and for the scheduling of periodic production. Sixteen diverse algorithms were considered – including greedy search, tabu search, simulated annealing, and the genetic algorithm – for the profitability of new schedules in the shortest computational time after flow interruption. Moreover, subject-matter experts tested and evaluated several scenario disturbances in the value chain process. The genetic algorithm and variations like local search, tabu search, and greedy search produced the best results. The contribution of this study includes evidence to show that inter-plant scheduling in a multi-product polymer production chain can be done within a reasonable time to ensure continuous process flow. In addition, a novel encoding scheme of decision variables is presented, allowing for scheduling over short or longer time horizons. Finally, the study shows that the performance results of the metaheuristics can guide practitioners on which to select for future implementation; this study showed that the genetic algorithm with population sizes of 50 to 120 and more than 100 generations proved to be best, while results can be obtained in less than 10 minutes.