Proactive energy-carbon aware production scheduling in SMT lines: A novel graph meta-reinforcement learning approach

The transition toward sustainable manufacturing requires Surface Mount Technology (SMT) lines to adopt energy-carbon aware production scheduling. Traditional paradigms treat machines as discrete states, failing to capture the continuous thermodynamic inertia of thermal equipment like reflow ovens. Coupled with dynamic electricity prices and fluctuating carbon factors, classical optimization and standard multi-agent reinforcement learning become inadequate. To address the deep coupling of discrete manufacturing logic and continuous thermodynamics, we reformulate the scheduling problem as a semi-Markov decision process and propose a novel graph meta-reinforcement learning architecture termed GMeta-MATD3. Specifically, graph neural networks are utilized to extract spatial–temporal topologies across production and microgrid networks. To navigate environmental non-stationarity, a meta-learning module is embedded within each agent to infer latent physical contexts and predict peer behaviors, enabling temporal energy arbitrage via deliberate “wait” actions. Additionally, an attention-based centralized critic dynamically decomposes joint values to resolve the multi-agent credit assignment problem. Extensive experiments on real SMT data coupled with IEEE distribution systems demonstrate significant performance improvements over state-of-the-art baselines. GMeta-MATD3 achieves the best Pareto front, yielding a 13.1h makespan, 1020.5 RMB energy cost, 810.2 kgCO₂ emissions, a 1.5% tardiness rate, and zero grid violations. Crucially, its thermodynamic stepping mechanism reduces transition energy from 12.5 to 3.2 kWh/batch, and proactive peak avoidance cuts the load peak-to-valley difference by 34.5%, maintaining daily carbon intensity below 0.55 kgCO₂/kWh. Scalability tests confirm that executing 500-batch schedules in under 500 ms.