Dynamic Real-Time Optimization of Modular Unit Allocation to Off-Site Facilities in Postdisaster Reconstruction Using Deep Reinforcement Learning

Abstract

Postdisaster housing reconstruction (PDHR) requires robust, efficient planning, and coordination among dispersed prefabrication facilities and jobsites to maximize prefabrication benefits. Manual construction methods and inherent risks often lead to unforeseen incidents and delays. Previous off-site construction studies focused on specific factors, neglecting uncertainties and the improvement of the Social Vulnerability Index (SoVI). Considering all critical parameters, this study proposes a real-time optimized allocation using deep reinforcement learning (DRL) for PDHR projects. The framework employs the Q-learning algorithm to generate the best real-time schedules for each prefabrication facility based on a current work status to complete a planned project. This is illustrated through a case study with ten project sites, four types of module layouts, and four prefabrication facilities. To demonstrate the superiority of the DRL-based method, the model was compared to the Monte Carlo Simulation and the Genetic Algorithm (GA) for the nine criteria related to time, cost, and social vulnerability. The DRL-based algorithm optimized the time parameter by reducing delay by 33.6% compared to the Monte Carlo simulation and by 46.4% compared to the GA. Similarly, it reduced missed deadlines by 35.1% compared to the Monte Carlo Simulation and by 13.3% compared to the GA. When comparing the cost parameter, the DRL model reduced overall cost by 3.4% compared to the Monte Carlo simulation and by 18.6% compared to the GA. In addition, it was able to prioritize the more vulnerable jobsites over less vulnerable ones to reduce delays and missed deadlines. In this regard, the proposed approach contributes to the body of knowledge by introducing a new automated model for PDHR project work distribution, considering productivity, cost, time, resources, uncertainties, and SoVI, thereby improving informed decision-making and overall project performance. This research presents an artificial intelligence (AI)-powered system tailored for managing large construction projects, especially crucial in postdisaster rebuilding efforts. By harnessing advanced AI techniques, this system streamlines operations in prefabrication factories, ensuring timely and cost-effective completion while considering critical social needs. The key factors that set this approach apart are adaptability and continuous learning. By constantly analyzing and learning from real-time project data, the system dynamically allocates resources, minimizing delays and cost overruns while also prioritizing areas with higher social vulnerability. However, it’s important to note that the system’s effectiveness relies on specific factory layouts and the availability of up-to-date information. In practical terms, implementing this system could revolutionize how construction projects are managed. It offers real-time insights, enabling informed decision-making, streamlined operations, and enhanced collaboration among all involved stakeholders. Beyond construction, the potential applications of this technology extend to thoughtful urban planning and expediting recovery in disaster-stricken areas. Despite existing challenges, this research paves the way for AI-driven solutions that not only optimize project efficiency but also integrate social responsibility into decision-making processes. It sets a promising path toward a future where AI plays a pivotal role in optimizing projects while concurrently addressing the critical needs of society.