An Integrated HRL-GAT approach for dynamic cooperative scheduling in two-stage distributed hybrid flow shops

Abstract With the global shift in manufacturing toward distributed production paradigms, multi-factory collaborative scheduling faces complex challenges arising from the dynamic coupling of part manufacturing and product assembly. This study addresses the two-stage distributed hybrid flow shop scheduling problem. It establishes a comprehensive collaborative optimization model that integrates parallel processing across distributed factories and centralized assembly stages. To overcome decision bottlenecks under dynamic disturbances, we propose a novel scheduling method that combines hierarchical reinforcement learning (HRL) with a graph attention network (GAT), referred to as HRL-GAT. The main innovations include (1) the construction of a tri-channel dynamic state matrix that fuses machine load, remaining processing time, and part type features, enabling adaptive state updates through rule-based transformations; (2) a dual GAT mechanism, where the local GAT optimizes operation-level resource allocation based on machine idleness, and the global GAT models inter-factory synchronization of similar operations based on differences in remaining time; (3) the establishment of a hierarchical decision-making framework, where the high-level policy generates scheduling guidance using global statistics, and the low-level policy selects feasible actions within a subset constrained by action masking. The combination of HRL and GAT significantly improves scheduling accuracy while reducing training time, as demonstrated in static experiments. Moreover, simplifying the policy structure through a “first-order hybrid + second-order fixed rules” approach enhances scheduling efficiency. When compared to traditional heuristic algorithms, HRL-GAT shows clear advantages in scheduling performance. In dynamic rescheduling, the policy reuse mechanism demonstrates optimal performance, achieving coverage rates of 81.1% and 76.3% for small and large datasets, respectively. Furthermore, the model achieved collaborative gains of 5.1% and 6.3% for small and large datasets, respectively. The proposed HRL-GAT framework provides a new, efficient collaborative scheduling method for distributed manufacturing systems, significantly enhancing the dynamic adaptability and collaborative efficiency of complex production systems.