Forecasting Short-Term Road Surface Temperatures – A Neural Network-based Approach

Abstract

The ability to forecast road surface temperatures (RST) in advance is an important asset for winter maintenance (WRM) operations as it effectively reduces WRM costs through more efficient use of their maintenance resources, which in turn would improve road safety and mobility of traveling public during winter months. However, considering the number of road networks that must be monitored and the extent to which RST varies over time, WRM agencies have constantly been looking for better ways to generate accurate road weather forecasts as they strive to optimize their WRM services and provide safe travel conditions. Despite its significance, there exist significant gaps in knowledge and methods in current literature for generating reliable short-term RST forecasts. In addition, forecasting itself is a challenging task that often involves many unknown variables creating uncertainties through their interactions. As a result, forecasting accuracies may vary significantly across different regions – an important aspect that has been often neglected in the existing literature. To tackle these issues, this thesis aims to develop highly reliable RST forecasting models using one of the most sophisticated and successful methods; namely, neural networks. In particular, this thesis attempts to assess the performances of the two uniquely developed neural network-based short-term forecasting models and investigate the hypothesis that geographical (e.g., latitude and altitude) attributes affect forecasting accuracy for improved generalization potentials. RST measurements collected by six selected stationary road weather information systems (RWIS) stations in Alberta, Canada, were utilized to validate the feasibility and applicability of the proposed method developed herein. The findings indicate that the two developed models were found to generate highly accurate results. The first method using conventional artificial neural network (ANN) models yielded forecasts with mean absolute error (MAE) values of 0.74, 1.34, 1.82, 2.40, 2.84, and 3.30°C for 1-h, 2-h, 3-h, 4-h, 5-h, and 6-h ahead forecasts, respectively. The second method using long short-term memory (LSTM) models, on the other hand, generated even more promising results, with average MAE values of 0.63, 1.17, 1.79, 2.28, 2.74, 3.14°C for the 1-h, 2-h, 3-h, 4-h, 5-h, and 6-h forecasting horizons, respectively. The novelty of this study lies in investigating the probable effect of several external factors on model performance, where it was revealed that forecasting horizon and geographical attributes influenced forecasting accuracies. Upon investigating the hypothesis that locational attributes affect forecasting accuracies, the results showed that accuracy improved with increasing latitude and decreasing elevations, which are worthwhile findings that can potentially lead to developing more refined models for generating highly accurate location-specific RST forecasts.