Probabilisitic road temperature forecasting using machine learning techniques – comparison of a multi-model approach in the Austrian Alps

Road weather conditions, especially road temperature, have a major impact on road safety even more so in case of unusual early/late snow events leading to damages to logistic infrastructure, road infrastructure, and fatalities. Road maintenance services use meteorological forecasts as well as targeted road temperature and precipitation forecast to estimate when, where, and how often they have to treat roads before and during such events happen and rely, thus, on accurate and targeted predictions.

Numerical weather prediction models are able to provide a good guesstimate but still lack the detail, temporal and spatial resolution, which is needed especially in regions with rugged terrain. Here, we take a two-fold approach in (i) implementing different post-processing methods, deterministic and probabilistic, and (ii) evaluate the skills of the separate models and a combined multi-model ensemble approach. Furthermore, simplistic transfer learning approaches are implemented to test the models’ skills in unobserved areas. For step (i) the following methods are used: (a) multilinear regression, (b) Kalman filtering, (c) random forest, (d) a feed forward neural network, (e) a transformer neural network, (f) the EMOS model, and (g) the Metro model.

Two Austrian regions are considered here, namely Tyrol and the state of Salzburg. Results for these regions show that the simplistic transfer learning approaches are only in a few cases better than a temperature height correction approach. The Kalman filter results show that a careful parameter selection is needed to achieve good results. They rely, too, on onsite measurements and cannot be applied to regions with no measurements. All other methods are able to improve the raw NWP forecasts indicating that overall a mix of method is better suitable than relying on one single method (deterministic/probabilistic).