Hierarchical collaborative scheduling in cloud manufacturing based on dynamic contract net protocol and deep reinforcement learning

Cloud manufacturing facilitates collaborative production by integrating distributed manufacturing resources via service-oriented platforms. However, its scheduling process is constrained by resource heterogeneity, complex task dependencies, and inefficient platform-enterprise coordination. To address these limitations, this paper proposes a deep reinforcement learning (DRL) approach for multi-task scheduling that integrates a contract net protocol (CNP) with graph neural networks (GNNs). First, a platform-enterprise collaborative scheduling framework is proposed. Then, two CNP-based schemes are designed and extended into a Dynamic CNP (DCNP) which updates evaluation weights to improved service selection. Moreover, an RGCN-LDDQN algorithm is proposed, which integrates relational graph convolutional networks (RGCN), long short-term memory (LSTM), and double deep Q-network (DDQN) to model inter-task dependencies and support global scheduling decisions. A bi-level priority constraint is integrated to mitigate conflicts, and a priority-aware P-RGCN-LDDQN is obtained, which enhances feasibility and stability. Subtasks are autonomously scheduled using a genetic algorithm at the enterprise layer based on platform-issued objectives and constraints. Simulations with respect to automotive parts assembly tasks are conducted. Results indicate that, under DCNP, RGCN-LDDQN reduces makespan by 21.88% and comprehensive cost by 15.64% compared to conventional DDQN, while maintaining service quality. Additionally, P-RGCN-LDDQN achieves further reductions of 9.15% and 6.69% over RGCN-LDDQN.