Urban pavement performance modeling methodology using artificial intelligence

Abstract

Pavement performance models are an extremely useful tool for road management agencies, allowing them to estimate pavement performance and designate the optimal corrective measure for road maintenance. Recent research has presented favorable results when using Machine and Deep Learning techniques in pavement performance modeling. This research proposes a methodology for developing first-phase performance models for urban pavements, managed at the network level, for short-term condition prediction using Machine and Deep Learning techniques in their preparation. Specifically, random forest regression (RFR), support vector regression (SVR), gradient boosting regression (GBR), artificial neural networks (ANN) and recurrent neural networks (ANN) are used. With these algorithms, models are built to predict the Chilean Urban Pavement Condition Index (ICPU) using a synthetic database of our own elaboration for the simulations. Regarding the results obtained, the iterations offer favorable results for short-term prediction, obtaining low average prediction errors (0.4% for the GBR algorithm) and RMSE results close to zero (0.063 for the GBR algorithm), which is of interest for each of the algorithm alternatives mentioned above, since they place them as recommendable tools to make predictions on the performance of urban pavements.