Forecast-based AI Decisions for Winter Road Maintenance

Abstract: In countries with freezing temperatures, proper road maintenance is an important intervention to reduce traffic accidents and fatalities and improves road passability. While de-icing using road salt can be an effective intervention, it may also lead to harmful side effects such as damage to vehicles, groundwater contamination and plant stress. The timing of a de-icing activity is critical since an optimal application time of the road salt increases its efficiency thus reducing the amount of road salt necessary and alleviating its negative impacts. Timing the call-out of de-icing vehicles is a difficult task that is based on experience and available forecast data and is often associated with anxiety for the supervisor. A decision support system that can learn from previous and expert decisions based on similar forecast data and context could help alleviate some of the stress and perhaps help improve the timing of the call-outs. This thesis investigates to what extent machine learning (ML) methods can learn from previous supervisor call-out decisions as a basis for such a recommendation system. Results show that the ML methods AdaBoost regression, Support Vector Regression and Quadratic Discriminant Analysis classification outperform other relevant methods and display predictive power for call-outdecisions. Careful feature engineering choices and design of an output regression target value are important. Results and evaluation show that the top-performing ML models are able to emulate historical supervisor call-outs and could add value when integrated in a decision support system used to guide call-out decisions.