Forecasting Machine Working Times in Off-Site Construction: A Machine Learning Benchmark for Supporting Production Planning

Abstract: Off-site construction relocates building processes to controlled manufacturing environments, enhancing productivity but increasing the risk of cost overruns due to the non-repetitive nature of module fabrication. This work presents a machine learning benchmark for a cyber-physical system that digitises manufacturing operations by capturing machine working times alongside design and production features. The benchmark spans tree-based and support vector regressors, with the best-performing model embedded in an industrial dashboard for data-driven scheduling. Validated in a UK steel frame manufacturer across eight months of production data, Random Forest achieved 10.84% median absolute percentage error on 420 frames, enhancing production planning and scheduling.