Multi-time Step Deterioration Prediction of Freeways Using Linear Regression and Machine Learning Approaches: A Case Study

Abstract

Abstract Multi-time step deterioration prediction of road pavements can provide more options for effective maintenance and rehabilitation decision under limited resources as compared to single time step prediction. This paper presents the 1–4 time step ahead prediction of the pavement cracking, rutting and roughness using their past values together with climate and traffic data as model inputs. Three prediction models were adopted, including the simple multiple linear regression (MLR) model and two sophisticated machine learning models, namely, support vector regression (SVR) and genetic programming (GP) models. An industry dataset of spray seal freeways is used to demonstrate the application of the methodology developed in this study. After calibration with observed data, the three prediction models are tested with unseen datasets using three performance indicators, namely, mean squared error (MSE), mean absolute percentage error (MAPE) and coefficient of determination (R squared). Among many results, all three models are unable to predict cracking with acceptable prediction accuracy. On the other hand, the rutting and roughness can be predicted with relatively good accuracy up to 4 time steps ahead. The sensitivity analysis shows that roughness and rutting prediction depends significantly on their previous or lagged values and not on remaining inputs such as annual average daily traffic (AADT) and rainfall. The methodology developed in this study is also applied to another dataset of asphalt freeways, which have similar model inputs. Similar findings are found with this dataset of asphalt freeways to that of spray seal freeways. The simple MLR model can produce similar prediction performance to the sophisticated SVR and GP models for rutting and roughness, suggesting the use of the MLR model as a benchmark for any development of prediction models for pavement deterioration.