Real-time AGV cluster scheduling using attention-enhanced PPO and strategy transfer in large-scale solar cell production workshops with multiple concurrent tasks
针对太阳能电池生产车间中多任务并发下的AGV集群实时调度问题，提出一种结合注意力机制和策略迁移的深度强化学习方法，以最小化运输任务的平均等待时间和配送距离。
In large-scale solar cell production workshops, automated guided vehicles (AGVs) serve as pivotal equipment for material delivery. Achieving high-performance real-time AGV cluster scheduling under multiple concurrent tasks remains a significant challenge. Based on deep reinforcement learning and transfer learning, this study proposes a real-time AGV cluster scheduling method to minimise the average waiting time and delivery distance of the transportation tasks. First, considering the states of AGVs, buffers, and processing machines, a Markov decision process (MDP) model that adopts multi-rule-based action is established to describe the real-time AGV cluster scheduling problem and mitigate the curse of dimensionality. Second, an improved proximal policy optimisation (IAPPO) algorithm that integrates the multihead self-attention (MHSA) mechanism into both the policy and value function networks is proposed to optimise the scheduling strategy. Next, considering scale, resource proportion, and layout, a similarity assessment model for solar cell production workshops is constructed. Furthermore, an adaptive pretraining and fine-tuning scheduling strategy transfer method is presented to accelerate learning and reuse scheduling knowledge. Comprehensive simulation experiments demonstrate that the proposed IAPPO-based real-time scheduling method outperforms the baseline minimum transportation task rule by 12.32% and that the designed transfer strategy accelerates convergence by up to 4.5 times.