Forecasting Highway Maintenance Cost at the Early Stage Using Machine Learning

Accurate cost estimation has been a notable challenge for developing cost-effective highway maintenance plans. While multiple variables, including market indices, have been used to attain higher accuracy of cost estimation models, the difficulty escalates for long-term forecasting. Especially in the prediction of early-stage projects, the limited amount of information and the unpredictability of market conditions make it intricate to achieve the desired accuracy. This paper applies machine learning (ML) to develop a highway maintenance cost forecasting model using only the early-stage project details. Georgia Department of Transportation (GDOT) pavement data from 2017 to 2021, consisting of “bidding-related,” “asset-related,” and “project-related” variables, is input to ML methods including linear regression, random forest, extra trees, gradient descent, extreme gradient descent, and k-nearest neighbors. Extra trees regression showed the lowest mean absolute percentage error of 18.51% after fivefold cross-validation. This study will work as a reference for DOTs to implement ML in the long-term maintenance cost prediction and ultimately be utilized in the tradeoff analysis of optimal management strategies.