Preliminary Cost Estimation of Pavement Maintenance Projects through Machine Learning: Emphasis on Trees Algorithms

Abstract

Accurate cost estimation is crucial for developing cost-effective pavement maintenance plans, but it remains challenging, particularly during the preliminary stage due to limited project information. The objective of this research is to create a cost forecasting model specifically for pavement maintenance projects at the initial stage. The study employed data preprocessing and feature selection methods to apply trees-based algorithms, including extreme gradient boosting random forest, and extra trees. Data from the Georgia DOT (GDOT) from 2017 to 2021, encompassing variables related to bidding, assets, and projects, were processed and fed into the chosen machine learning algorithms. Model performances were assessed and compared using key metrics to identify the optimal combination of predictive model and feature selection method that achieves a reliable preliminary cost estimation at the initial stages of the plan development process. The extra trees regression with SelectFromModel feature selection demonstrated superior performance, achieving a mean absolute percentage error of 13.60% and a coefficient of determination of 91.68%. The results underscore the applicability of the extra trees algorithm in cost forecasting through the following aspects: (1) the model is suitable for the preliminary stage when only limited features are available; (2) the model provides consistent prediction results despite the lack of extensive or high-quality data; and (3) the model is tailored for maintenance projects by integrating the asset characteristics into the model. The findings contribute to the body of knowledge through (1) identifying the significant features for determining preliminary cost estimates of pavement maintenance projects; and (2) developing a reasonably accurate model for cost estimating in the initial stages of plan development process. It is anticipated that the outcome of this research will provide transportation agencies with a practical approach of integrating machine learning into their current practices of preliminary cost estimation, ultimately utilizing the public fund more efficiently.