Application of Machine Learning for Temperature Prediction in a Test Road in Alberta

Pavement temperature prediction plays a key role in determining the structural capacity and deflection of asphalt pavement, owing to the viscoelastic behavior of asphalt. Thus, a high degree of accuracy is desirable for the prediction of asphalt temperature from available parameters such as air temperature and solar radiation. Asphalt temperature prediction models can be based on different approaches, including analytical, numerical, and statistical methods. Each of these models has its own strengths and weaknesses, and their accuracy varies based on site conditions. The goal of this research is to compare the accuracy of machine learning approach in general with existing models for prediction of temperature in asphalt pavements, based on 6 years of data collected from temperature sensors embedded in an instrumented test road in Alberta. A sensitivity analysis was performed to determine the most important parameters for prediction of temperature, and MATLAB regression learner was used for implementing machine learning-based algorithms based on parameters, which were determined to be air temperature, solar radiation, and day of the year. The machine learning methods were compared with existing literature models for prediction of average, minimum, and maximum daily pavement temperatures at different depths throughout the asphalt layer. The predicted results were validated by comparison with available field data. All machine learning algorithms used in this study resulted in the prediction of temperature values with higher accuracy compared to existing models, demonstrating the applicability of these machine learning models for improved pavement temperature prediction.