Graph reinforcement learning-driven source-load cooperative scheduling optimization for textile production

• Proposes graph reinforcement learning-driven source-load co-scheduling for textile production. • Constructs knowledge graph to characterize heterogeneous correlations of production-energy systems. • Establishes multi-objective MDP model to quantify dynamic trade-offs among makespan, carbon emission, and energy cost. • Utilizes heterogeneous attention GNN with adaptive greedy strategy, trained via PPO algorithm for real-time scheduling. • Case study demonstrates 2.9 % energy cost reduction, 10.1 % carbon emission decrease, and 1.1 % makespan increase. There are engineering issues caused by the traditional energy scheduling model where the energy generation (source) adjusts according to the energy consumption (load) in textile industrial parks, leading to a disconnect between the energy system and the production system, low energy efficiency, and high carbon emissions. This research proposes a graph reinforcement learning-driven source-load collaborative optimization method. Based on a domain knowledge graph fusing textile production processes, equipment status, and energy consumption relationships, the method explicitly characterizes the heterogeneous correlation of the source-load system. It further establishes a multi-objective Markov Decision Process model to quantify the dynamic game relationship among makespan, carbon emission index, and energy cost. To capture the complex, long-range dependencies in production processes, heterogeneous attention graph neural networks are used to capture the long-range dependencies in production links, combined with an adaptive greedy sampling strategy for optimizing scheduling decisions. Additionally, the Proximal Policy Optimization algorithm is adopted to train a hybrid scheduler, facilitating deep collaborative decision-making for both production processes and energy scheduling. Validation based on actual data from a Shanghai textile enterprise shows that the proposed method outperforms the benchmark, achieving a 10.1 % decrease in carbon emissions, a 2.9 % reduction in energy cost and a minimal 1.1 % increase in makespan. Furthermore, compared to the OR-Tools solver, the performance differences are limited to <4.0 %, 2.8 %, and 0.1 %, respectively, for energy costs, carbon emissions, and makespan.