Invited seminar presentation at ABB Corporate Research
Abstract: Monitoring of industrial processes yields large amount of measurement and operation data, which have the potential of improving the operation of the processes. However, in real-life applications, the potential of data is rarely fully exploited. The main reason is that, in order to be useful for the operational decisions, the raw data often requires processing by sophisticated data analytics or machine learning methods. At the planning and scheduling level, the operational decisions are traditionally generated by mathematical optimization methods. In this seminar, we present work that aims at improving planning and scheduling optimization via selected machine learning methods. The main strategies are to 1) improve the accuracy of the optimization parameters by prediction methods and 2) reduce the size of the decision space by data-driven elimination of non-sensible decisions. In addition, we propose a framework of where decisions of rescheduling procedures are generated using reinforcement learning.