Precast components on-site construction planning and scheduling method based on a novel deep learning integrated multi-agent system

The performance of precast construction relies on efficient on-site planning and scheduling of precast components. However, existing research primarily addresses either assembly sequence planning or construction process scheduling. The interactions between these elements and the dynamic nature of construction environments are often overlooked. This paper presents a novel two-stage planning and scheduling methodology that leverages a multi-agent system (MAS) and deep learning to account for construction environment, crews, materials, equipment, and assembly sequencing. First, precast components within the construction area are categorized into groups based on their coupling states. A knowledge-enhanced graph transformer is then introduced to determine the assembly sequence of intra-group components, considering their physical attributes and topological relationships. Subsequently, the MAS is developed to orchestrate the sequence of inter-group construction activities, addressing factors such as site congestion, worker movements, and material consumption. The knowledge-enhanced graph transformer is integrated into the MAS as a smart agent, to facilitate coordination between intra-group and inter-group planning. Finally, an engineering case is conducted to validate the effectiveness of the proposed method. The results demonstrate that, compared to conventional MAS methods, the groupbased plan leads to a 3.1 % reduction in construction time, a 6-10 % decrease in crew walking distance, and effectively mitigates assembly difficulties. These outcomes provide construction managers with innovative strategies and perspectives for optimizing construction workflows.