Utilizing artificial neural networks to predict the asphalt pavement profile temperature in western Europe

Abstract

The temperature profile significantly influences the structural performance of asphalt pavement since it influences various mechanical parameters such as stiffness, strength, and fatigue life. The present study aims to achieve two main objectives. Firstly, to explore the potential of Machine Learning (ML) approaches in predicting asphalt pavement profile temperature in the western Europe climate. Secondly, to determine the impact of different weather parameters on the effectiveness of the prediction models. Therefore, three ML algorithms are used to develop asphalt temperature prediction models: autoencoder network, Feedforward Neural Network (FFNN), and Long Short-Term Memory (LSTM). The performance of different ML algorithms is assessed by comparing the accuracy of the prediction models. The analysis results indicated that the autoencoder network is the most accurate algorithm in predicting asphalt pavement temperatures at various depths. Regarding the weather parameters' input dimension, a direct relationship was observed between the number of input dimensions and the performance of the prediction models, where the model using all weather data performs best. The accuracy of the developed ML and a validated FE model were compared with the collected experimental data in terms of Mean Absolute Error (MAE). The average MAEs were calculated between 0.25 and 0.31 degrees C and 1.05-1.19 degrees C for ML and FEM approaches, respectively. The results showed the ability of ML algorithms to predict the temperature of asphalt pavement with a higher degree of accuracy compared to numerical models with no extra information about material properties or boundary conditions.