Abstract
Production scheduling in the process industry is an area of significant activity in research and of great practical importance for the performance of industrial companies. In the vast majority of research papers, the scheduling problem is formulated as an off-line problem where a number of jobs is scheduled on a number of resources and the efficiency of the formulation and the solution algorithms is discussed. In reality, however, scheduling is a continuous activity that has to react to the arrival of new orders, to variations in processing times, breakdowns, lack of resources etc. This is termed real-time (or online) scheduling. Available commercial solutions usually provide solutions with relatively long update intervals due to the necessary computation times and a delayed flow of information from the manufacturing execution systems where the data on the current state of the production is collected. Thus the computed schedules are outdated quickly, if not already at the point in time when they become available, and new orders and all kinds of variations and disturbances are handled by the operators and plant supervisors. There has been little concern for systematic solutions for real-time scheduling that continuously react to the changing situation in the plant. In this contribution we present a simulation-optimization approach to real-time scheduling where a detailed discrete-event simulation of the plant is coupled with an evolutionary algorithm. The simulation model is continuously updated with the latest plant data, and the optimization algorithm continuously improves the schedule based on the currently available information. We validate our approach using an example of a multiproduct, multistage batch plant in the pharmaceutical industry from the literature, demonstrating that the proposed approach can generate high-quality solutions quickly. We compare the results with those of an ideal (clairvoyant) scheduler, demonstrating that our method effectively manages unforeseen disturbances.