How level of detail of activity duration data influences look-ahead schedule performance in prefabricated construction

"How many parts will we install tomorrow?" Field managers on prefabricated construction jobsites need to answer this question when updating the look-ahead schedule (LAS) each day. If too few parts are scheduled, crews finish early and both crew and crane time are wasted. If too many parts are scheduled, crews work overtime, and additional costs are incurred for laydown and trucking changes. Field managers rely on activity duration data to inform LAS updates. Project teams collect these data at various levels of detail (LoD) from low (logging the number of parts installed per day, regardless of type) to high (logging the activity duration and type of each lift). My dissertation studies how LoD of activity duration data influences LAS performance for the "how many parts tomorrow" problem. More specifically, how does LoD of activity duration data influence (1) magnitude of LAS duration errors (2) accuracy of LAS duration uncertainty quantification, and (3) ability to minimize cost of waste from LAS duration errors? To answer these questions, I developed the Uncertainty-based Look-ahead Optimizer (ULO) which optimizes the number parts in LASs based on estimated LAS duration distributions obtained through randomly sampling input distributions fitted to activity duration data. I first used the ULO to organize crane sensor data from 2,693 crane lifts over 2 precast parking lot construction projects into activity duration data at three LoDs: aggregate LoD, part LoD and lift LoD. For each LoD, I used the ULO on each day of the project schedule to generate a next-day LAS that only uses data from all previous days. Finally, I quantified the performance of the LASs generated by each LoD of activity duration data and compared the results. The comparison shows a 40-59% (50-89 minute) decrease in next-day look-ahead mean absolute error (NL-MAE), a 71-74% increase in next-day look-ahead mean likelihood (NL-ML), and a 26-35% ($2,250-$3,770) decrease in next-day look-ahead mean cost of errors (NL-MCE) between the aggregate-level LASs and the lift-level LASs across the two case studies. My dissertation contributes two new insights. Firstly, increasing LoD of activity duration data from aggregate to part-level significantly improves LAS performance (-43% NL-MAE, +36% NL-ML, -25% NL-MCE on average), and less so from part-level to lift-level (-11% NL-MAE, +26% NL-ML, -8% NL-MCE on average) at a project level. Secondly, increasing LoD of activity duration data on specific days with major outlier activities taking more than 1. 5 hours than expected reduces LAS performance in terms of NL-MAE (60-67% of days), NL-ML (33-70% of days), and NL-MCE (67-70% of days). My dissertation also contributes the ULO model, which can help field managers automatically create buffered LASs in less than 1 minute per day instead of spending 2 hours per day doing so manually based on experience. Finally, my dissertation contributes three metrics, NL-MAE, NL-ML and NL-MCE, for benchmarking how well LAS generation methods address the "how many parts tomorrow" problem across a project schedule. These contributions should facilitate future research on finding more precise answers to the "how many parts tomorrow" problem using data-driven construction management methods.