Plug and schedule: scalable and modular architecture for integrated and dynamic scheduling based on graph neural networks and reinforcement learning

Manufacturing companies increasingly require scheduling approaches that integrate practical constraints such as workforce availability, maintenance, setup dependencies, energy objectives, and uncertainty, yet existing research offers limited support for transferring them into heterogeneous industrial data landscapes. This paper proposes a scalable and modular architecture for integrated and dynamic scheduling that combines an Asset Administration Shell-based middleware for standardised data integration, a heterogeneous graph representation of scheduling instances, a graph neural network-based State Modeler, and a Proximal Policy Optimisation-based RL-Scheduler coupled with event-based simulation. The architecture is validated on the production system of a German machinery manufacturer with 59 manufacturing resources and 39 employees and on public benchmark instances. On the industrial instances, the learned policy reduces the makespan against the mean of seven dispatching baselines by 2.93% to 6.76% across six evaluation cells covering small-part, large-part, and combined production under deterministic and stochastic machine failures, with statistical significance in every scenario. ... mehrThe energy-aware configuration preserves this makespan performance while significantly reducing the relative energy cost in most scenarios. The findings indicate that the proposed architecture provides a practically transferable pathway for configurable RL-based scheduling in heterogeneous production environments.