Scheduling of steelmaking-continuous casting process by integrating deep neural networks with mixed integer programming

This study addresses the scheduling problem in the steelmaking-continuous casting (SCC) process. The SCC process is a hybrid flow shop with three stages, and we focus on job dispatching in the second stage, the refining stage. Our primary aim is to develop an algorithm applicable to real-world scenarios, mirroring field engineers' decision-making and handling the process's complex features. We propose a deep neural network (DNN)-based approach, trained on engineers' past decisions, achieving up to 97% accuracy. However, DNN alone falls short of outperforming engineers in scheduling objectives, specifically minimizing the total completion time in the refining stage. Hence, we introduce a novel approach combining DNN with mixed integer programming (MIP). In the integrated approach, the DNN initially makes decisions, but when confidence in the accuracy of a DNN-based decision is lacking, as determined by a developed reliability measure, it is supplemented with a decision derived using MIP. Experiments demonstrate that this integration improves scheduling objectives, surpassing engineers' performance. Furthermore, filtering inaccurate decisions enhances the accuracy of the DNN-based decisions. The proposed approach has been successfully implemented in one of South Korea's largest steelmaking companies.