Heterogeneous Graph Reinforcement Learning for Dynamic Scheduling in a Multibranch Panel-Furniture Workshop

In panel-furniture manufacturing, dynamic scheduling is challenging because upstream motherboard release and parallel CUT-machine assignment are tightly coupled with downstream multibranch processing, finite-capacity buffers, and machine disturbances. This study investigates an event-driven scheduling problem in a multibranch panel-furniture workshop and aims to minimize the cumulative tardiness of all parts. To capture the structural dependencies of the shop floor, we construct a heterogeneous graph to represent the real-time production state. In this graph, candidate motherboards, parallel CUT machines, and downstream branch-level production units form different node types. The representation explicitly encodes motherboard–branch coupling features and machine–branch inflow relations, while aggregated branch-state features summarize detailed process-order information within each downstream branch. Based on this structured state representation, we develop a proximal policy optimization (PPO)-based actor–critic framework to learn adaptive motherboard release and CUT-assignment policies. We compare the proposed method with three heuristic rules and two PPO-based learning baselines under different machine-availability levels, buffer-capacity settings, and dataset scales. The results show that the proposed method achieves the best overall tardiness performance across the tested baselines and converges faster and more stably than the other PPO-based methods under different dataset scales and workshop conditions. These findings demonstrate the effectiveness of combining heterogeneous graph modeling with reinforcement learning for dynamic scheduling in multibranch panel-furniture production systems.