Recurrent Neural Networks for Pavement Performance Forecasting: Review and Model Performance Comparison

Abstract

Accurate pavement performance forecasting is critical in supporting transportation agencies’ predictive maintenance strategies: programs that prolong pavement service life while using fewer resources. However, because of the complex nature of pavement deterioration, high accuracy for long-term and project-level pavement performance forecasting is challenging to traditional models. Therefore, researchers have taken advantage of machine learning (ML) technology to create more sophisticated models in recent years. However, there are no extant studies that compare different ML models on a singular, real-world, large-scale, and comprehensive pavement data set to evaluate their capability for pavement performance forecasting. Thus, the goal of this study is to critically evaluate ML models, such as multiple linear regression (MLR), fully connected neural network (FCNN), recurrent neural network (RNN), gated recurrent unit (GRU), long short-term memory (LSTM), and a hybrid LSTM-FCNN model, on Florida’s statewide, 31 year historical pavement data set. The results demonstrate that the RNN, GRU, LSTM, and LSTM-FCNN models perform significantly better than MLR and FCNN for predicting time-series pavement condition, with the LSTM-FCNN model performing the best. This result provides a valuable demonstration and recommendation to transportation agencies and researchers that RNN-based ML models are a promising direction to improve the accuracy of pavement performance forecasting.