Deep reinforcement learning-based schedule optimization for parallel precast production

Purpose - With the increasing use of precast concrete elements in off-site construction, optimizing precast component production scheduling (PCPS) has become critical for improving construction efficiency. This study aims to develop a deep reinforcement learning (DRL)-based scheduling optimization method for parallel precast production to minimize earliness and tardiness penalties as well as the makespan. 

Design/methodology/approach - A parallel production process model is developed considering resource constraints, including crew quantities and fixed mold plates. A pre-trained DRL model is employed for rescheduling under varying precast orders with different quantities and due dates. The practicality of this approach is validated using real case data from field studies, comparing its performance with traditional dispatching rules (DPs) and the genetic algorithm (GA). 

Findings - The DRL-based method generates production schedules that are viable for practical applications. Compared to traditional DPs and GA, the proposed approach demonstrates superior stability, enhanced rescheduling capability and reduced computational time. 

Practical implications - The proposed DRL-based scheduling method offers a practical and efficient solution for optimizing precast production scheduling. It enhances decision-making in dynamic construction environments by reducing penalties and makespan while improving scheduling adaptability.

 Originality/value - This study expands the limited research on parallel PCPS by introducing a DRL-based approach, which integrates scheduling optimization with dynamic rescheduling adaptability under real-world conditions.