Deep reinforcement learning-based schedule optimisation method for off-site precast project

Abstract

Off-site construction (OSC) with precast components (PCs) has been widely adopted as a novel construction method in recent decades. It provides superior construction quality control, since the production of PCs is supervised in a factory environment, while the erection of PCs is conducted primarily at the job site. This means that off-site construction project scheduling (OSCPS) should consider both precast component production scheduling (PCPS) and construction project scheduling (CPS). Traditionally, production and construction scheduling methods have usually relied on human experiences. By utilising classical scheduling principles such as dispatching rules and the critical path method, schedulers can produce master plans with general milestones for the project. However, as the scale of an OSC project grows, the variety of PCs and construction activities also increases. The production of dozens of PCs across multiple production lines and hundreds of construction activities presents greater challenges for project scheduling. Schedulers need to consider optimising more complex combinations of parallel production and construction operation sequences.

The theoretical contribution of this study lies in three main aspects. First, the research extends the DRL application in large-scale PCPS, CPS and integrated scheduling. The proposed method can generate optimised schedules under various constraints. Through the modelling of the production and construction processes, as well as the design of the DRL framework, the proposed method further has the capacity for rescheduling. Unlike traditional optimisation algorithms that require repeated recomputation for each new scheduling instance, the learned DRL policies enable fast decision-making under varying project states. Second, this research introduces an enhanced DRL architecture featuring a VAS mechanism. The VAS mechanism enforces feasibility constraints by allowing the agent to sample only from actions that satisfy operational rules, thereby improving training stability and preventing unsafe or infeasible scheduling decisions.

This research can also be implemented practically. The methodology has been validated using real-world project scenarios. Customisable simulation environments are constructed based on these case studies, allowing for adaptation to various construction projects and PC facilities. Schedulers can use the proposed method to generate production and construction schedules that meet their project requirements. With the detailed schedules, project managers can pay more attention to controlling project duration, cost and quality.

Date of Thesis submission	2025
Original language	English
Awarding Institution	Western Sydney University
Sponsors	Australian Government Research Training Program Scholarship
Supervisor	Vivian Tam (Supervisor), Jun Wang (Supervisor), Khoa Le (Supervisor) & Anthony Butera (Supervisor)