Parameter sensitivity experiment results.

To enhance construction efficiency and economic performance in prefabricated building projects under limited resource conditions, this study proposes an integrated intelligent optimization model based on Building Information Modeling (BIM) semantic representation. The model is designed to generate optimal assembly plans under multi-objective trade-offs, achieving a balanced compromise between shortened construction periods, reduced costs, and minimized resource conflicts. The study begins by constructing an assembly semantic model using the publicly available BuildingNet dataset, extracting key components’ geometric structures and spatial topology to establish a data foundation suitable for multi-objective scheduling modeling. A multi-objective particle swarm optimization (MOPSO) algorithm enhanced with a dynamic objective weighting mechanism is then introduced. By allowing flexible prioritization of construction duration, budget cost, and resource usage, the model generates a diverse solution set and provides multiple candidate optimization schemes. Furthermore, a Deep Q-Network (DQN)-based reinforcement learning strategy is integrated to provide real-time feedback on each solution’s performance during simulated scheduling, enabling continuous policy updates and adaptive evolution. Experiments conducted on 100 standardized assembly tasks demonstrate the model’s effectiveness, producing feasible solution sets under varying objective weights. For a representative configuration, the model achieves an average construction period of 85.2 days, a budget cost of USD 1.486 million, and fewer than 1.7 resource conflict events. Compared with rule-based scheduling models, the Non-dominated Sorting Genetic Algorithm II (NSGA-II), and static MOPSO without feedback mechanisms, the proposed approach outperforms in terms of objective coverage, convergence speed, and solution diversity. It achieves superior results in key metrics, including hypervolume (HV = 0.683), solution spread (Spread = 0.227), and inverted generational distance (IGD = 0.017), validating its robustness and adaptability in complex trade-off scenarios. The findings indicate that integrating semantic modeling, evolutionary optimization, and learning-based feedback offers significant potential for dynamic multi-objective construction optimization, providing effective support for BIM practices oriented toward benefit–schedule–resource coordination.

为提升有限资源条件下装配式建筑工程的施工效率与经济效益，本研究提出一种基于建筑信息模型（Building Information Modeling，BIM）语义表征的集成智能优化模型。该模型旨在多目标权衡下生成最优装配方案，实现缩短施工工期、降低成本与减少资源冲突三者间的平衡折中。本研究首先利用公开数据集BuildingNet构建装配语义模型，提取关键构件的几何结构与空间拓扑关系，为多目标调度建模奠定数据基础。随后提出一种融合动态目标加权机制的多目标粒子群优化（Multi-Objective Particle Swarm Optimization，MOPSO）算法。通过灵活设定施工工期、预算成本与资源消耗的优先级，该模型可生成多样化的解集合并提供多套候选优化方案。此外，本研究还集成了基于深度Q网络（Deep Q-Network，DQN）的强化学习策略，在模拟调度过程中对各方案的性能进行实时反馈，实现策略的持续更新与自适应演化。在100项标准化装配任务上开展的实验验证了所提模型的有效性，其可在不同目标权重下生成可行解集合。针对一组典型配置，该模型可实现平均施工工期85.2天、预算成本148.6万美元，且资源冲突事件少于1.7起。与基于规则的调度模型、非支配排序遗传算法II（Non-dominated Sorting Genetic Algorithm II，NSGA-II）以及无反馈机制的静态MOPSO相比，所提方法在目标覆盖范围、收敛速度与解多样性方面均表现更优。其在超体积（Hypervolume，HV=0.683）、解散布度（Spread=0.227）与反向世代距离（Inverted Generational Distance，IGD=0.017）等关键指标上取得了更优结果，验证了其在复杂权衡场景下的鲁棒性与适应性。研究结果表明，将语义建模、进化优化与基于学习的反馈机制相结合，可为动态多目标施工优化提供巨大潜力，为面向效益-工期-资源协调的BIM实践提供有效支撑。