RWIS based road condition prediction using machine learning algorithms

Abstract

The need for a forecasting model of road conditions is becoming evermore critical, given the effects of ever-increasing severity in weather. Drastic changes in weather, especially cold fronts, often lead to dangerous roads. Consequently, traffic efficiencies are diminished and, even worse, accidents resulting in loss of life and property could increase. Across the nation, states are responsible for anticipating inclement weather and treating roads accordingly. Treatment costs can be reduced with more precise road condition predictions. The development of machine learning capabilities has enhanced the utilization of Big Data Systems throughout various sectors, including road climatology, making weather forecasting much more efficient and reliable. The study reported in this thesis analyzed various road climatology data, including sub-surface temperature at two- and six-inches from Road and Weather Information Systems (RWIS) deployed by Oklahoma Department of Transportation (ODOT) along the I-35 corridor at various road-bridge intersections aimed at producing a reliable and robust forecast model for predicting road surface temperature in the near and distant future. The predicting importance of each factor is analyzed statistically, and then manually, to determine its requirements for the forecast model. The study also determined the best forecast model after comparing a newly developed neural network with common regression techniques previously available through Machine Learning. Results showed that the novel neural network model offered a reliable 12-hour prediction for road surface temperature at a frequency of five minutes, depending on available historical data from RWIS. Two additional classification models provided Road Conditions Classes. The first was based on time series historical data from RWIS, and the second was based on historical and future data from a GFS (Global Forecast System). Together, these models accurately forecast local road surface temperatures 12-hour in advance of inclement weather in five-minutes frequencies at RMSE of ±1.67. They also accurately classified road conditions at a rate of more than 87.984%.