Probabilistic forecasts of winter road maintenance operations based on machine learning

Planning the scale of necessary road maintenance actions based on outputs from numerical weather prediction (NWP) models can be challenging. By employing machine learning, these actions can be estimated more objectively. 

A machine learning model was developed to predict the hourly probability of salting and snow removal operations across different road maintenance categories. This model provides a three-day forecast of maintenance probabilities over a five-kilometer pixel grid covering mainland Finland. 

The model was developed collaboratively by the Finnish Meteorological Institute and Destia, Finland's largest infrastructure service company. This development was part of the JääVe project, funded by the National Emergency Supply Agency. The model uses a gradient boosting approach and is trained on reports of snow removal and salting operations provided by Destia. Inputs to the model include MEPS NWP model outputs, cyclical temporal variables, and static landscape features. 

The model has learned to account for varying regional and seasonal characteristics. For example, it predicts different probabilities for coastal versus inland areas and adjusts for variations between the sparse northern regions and the more populated southern settlements. The model's predictions are comprehensive, extending beyond the road station locations used for fitting to cover the whole Finland. 

 

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .