Multi-objective risk optimization for sustainable modular infrastructure using machine learning and metaheuristics

This paper presents a unified predict–then–optimize framework that links calibrated machine-learning risk models with an evolutionary multi-objective optimizer to plan modular projects under uncertainty. A multi-source dataset, BIM/IFC exports, ERP costs, factory QC logs, transport manifests, and regional electricity factors- feeds tree-based and gradient-boosted models that return distributional estimates of schedule delay, cost overrun, and embodied CO₂e. Model reliability and interpretability are ensured via probability calibration and SHAP analyses, which consistently identify supplier on-time-in-full, defect rates, interface density, route length, load factor, and grid intensity as dominant drivers. These posterior predictive distributions parameterize a stochastic program solved with NSGA-II/III and MOEA/D, evaluated by hypervolume, IGD/GD, spread, and empirical attainment. Results show well-diversified Pareto sets with strong front quality (e.g., hypervolume proxy magnitude ≈ 0.90; IGD ≈ 0.30) and clear managerial “knees,” where small schedule concessions unlock disproportionate cost or carbon savings. Three exemplar plans illustrate actionable trade-offs: a time-lean option achieves 0 days tardiness at 3.69% overrun and 1123.7 kg CO₂e; a cost-lean option yields 0% overrun with 5 days tardiness and 292.0 kg CO₂e; and a low-carbon option reaches 69.5 kg CO₂e with 6 days tardiness and 0.90% overrun. The framework closes the gap between risk quantification and planning, transforming predictive insights into executable, auditable plans that enhance schedule reliability, budget discipline, and decarbonization in modular construction.

Text 123673.pdf - Published Version Restricted to Repository staff only Download (1MB) Official URL or Download Paper: https://link.springer.com/article/10.1007/s42107-0...

Actions (login required)

View Item