Knowledge-Driven reinforcement learning for Prefabricated construction supply Chain Scheduling: A Two-Phase large language model integration framework

Prefabricated construction supply chains require coordinated scheduling across production, transportation, and assembly under transport uncertainty, storage constraints, and assembly precedence requirements. However, existing approaches rely on stage-isolated optimization, static heuristics, or black-box AI models that cannot ensure engineering feasibility, interpretability, and adaptive decision-making under dynamic operational conditions. To address this problem, this study proposes a knowledge-driven AI framework that formulates the three-stage scheduling problem as a Markov Decision Process and solves it using a Maskable Proximal Policy Optimization agent under a two-phase large language model integration architecture. In the offline phase, engineering constraints are extracted from technical standards via document-grounded prompt engineering, reducing manual constraint specification effort while providing traceable provenance, and embedded into reinforcement learning as action-masking predicates. In the online phase, an entropy-triggered LLM consultation mechanism provides interpretable scheduling reasoning at high-uncertainty decision points, while a Random Forest classifier enriches the reward function with delay-risk predictions to guide dispatching under weather variability. The two-phase architecture ensures that symbolic engineering knowledge constrains the action space while adaptive learning optimizes within the feasible region. Evaluated across eight scenarios spanning capacity constraints, deadline pressures, and uncertainties, the proposed framework consistently outperforms rule-based and myopic baseline policies under most scenarios. Sensitivity analysis identifies site storage capacity and dispatch batch size as dominant feasibility determinants. This study contributes a knowledge-driven and constraint-aware AI framework that bridges symbolic engineering knowledge, interpretable reasoning, and adaptive reinforcement learning for industrialized construction logistics, advancing trustworthy and uncertainty-resilient AI decision-making paradigms for complex construction supply chains.