Improvement of transportation cost estimation for prefabricated construction using geo-fence-based large-scale GPS data feature extraction and support vector regression

•GPS data of equipment can be abstracted and modeled to track transportation demand.•The support vector regression model is trained with project specification datasets.•The prediction model allows for the estimation of transportation demands and cost.•The conventional fixed-cost approach tends to over-estimate transportation costs. In panelized construction, transportation is an essential process linking a manufacturing facility to a project’s jobsite using hauling equipment (e.g., trucks and trailers). Accordingly, the cost associated with transportation operations is considerable compared to a traditional stick build. Nevertheless, transportation cost estimation has often relied on a fixed-cost approach, regarding the cost as part of the overhead cost, rather than conducting detailed estimation of actual transportation operations. This is because operation-level data might be challenging to collect and analyze in practice. In this regard, the prevalent use of GPS devices for construction equipment may provide an automated means of monitoring the operations of transportation equipment, and large and detailed spatial and temporal data can be generated from multiple pieces of equipment in multiple construction projects on a daily basis or even in real time. This study thus proposes a spatial and temporal data filtering and abstracting approach to transportation cost estimation using fleet GPS data which extracts equipment activities from the GPS data and accordingly predicts the transportation demands required for an individual project. From large-scale GPS data, key operation information, such as the number of trailers and durations required (i.e., transportation demands), is extracted using a geo-fence and a rule-based equipment operation analysis algorithm. Then, the extracted transportation demand information, along with related project specifications, is used to train support vector regression (SVR) models for the purpose of predicting the transportation demand in new projects, which is in turn utilized to estimate the transportation cost using the relevant transportation unit cost of the equipment. To evaluate the performance, GPS datasets collected from 221 panelized residential projects over a period of 8 months are used to train the prediction model and are compared with actual transportation costs estimated in practice. The results show that the SVR model has an accuracy of 86% and 88% in predicting the number of trailers and the duration, respect