A proactive digital twin framework for dynamic pull scheduling in modular construction factories

Achieving continuous flow in modular construction is challenging due to high-mix variability, where traditional static methods often fail, leading to excess Work-in-Process (WIP) and flow stagnation. This study proposes a Digital Twin framework utilizing a novel "Offline Learning, Online Control" architecture. The methodology employs Simulation-Based Optimization to derive optimal control policies offline, which are synthesized into a lightweight surrogate model. This model drives an autonomous Scheduler Agent that acts as a virtual pacemaker, regulating real-time flow to shift production from a 'push' to a dynamic 'pull' strategy. Specifically, the agent monitors the cumulative count of finished units at critical downstream constraints, throttling upstream releases until the cumulative count of finished units from the current project surpasses a dynamically calculated threshold. To validate this approach, a case study was conducted on a wall panel assembly line in Edmonton, Canada. For this specific implementation, a Genetic Algorithm was used to train a linear regression surrogate policy against stochastic demand scenarios. Experimental results demonstrate a 38.5% reduction in average Cycle Time (Lead Time) compared to the facility’s baseline practice. These findings confirm that integrating simulation-based learning with real-time surrogate control successfully stabilizes flow efficiency, minimizes WIP accumulation, and prevents gridlock in constrained manufacturing environments.