Two-stage Intelligent Dynamic Scheduling Framework for Re-entrant Blocking Flow Shops with Central Buffer

Abstract

Re-entrant blocking flow shops are commonly encountered in modern discrete manufacturing processes, where jobs revisit stations and blocking occurs without intermediate buffers. To improve efficiency and minimize makespan, this study proposes a two-stage intelligent dynamic scheduling framework for re-entrant blocking flow shops with a central buffer. A system-level disjunctive graph is constructed to model interactions among operations, machines, and the buffer, and the scheduling problem is reformulated as a two-stage Markov Decision Process. A hybrid Graph Neural Network – Long Short Term Memory Network is developed to encode spatial–temporal features, while a hierarchical policy alternates between operation–resource selection and job assignment. Experimental results across multiple cases show that the proposed method clearly outperforms representative heuristic rules, achieving an average makespan improvement of 8%. Moreover, it attains solution quality comparable to genetic algorithm optimization with extended computation time while maintaining millisecond-level inference latency. These results demonstrate the framework’s effectiveness, scalability, and real-time applicability for intelligent scheduling in complex manufacturing environments.