Differences in modelled pavement temperature at German Road Weather Stations

The Canadian Meteorological Service's road surface model METRo is used by the DWD to simulate road surface temperatures at around 1,500 road weather stations (RWS). The modelled road temperature from METRo, together with other meteorological parameters from Model Output Statistics (MOS), is displayed on operational websites of winter services such as Autobahn GmbH.

The MOS provides a combined statistical interpretation of the IFS-HRES (ECMWF) and ICON (DWD) model forecasts at individual weather stations. The technique is based on multiple linear regressions by minimizing the Root Mean Square Error (RMSE). A wide range of model variables are used as predictors, including unobserved variables, as well as surface observations, precipitation radar and lightning detection for nowcasting.

The MOS is used to apply forecasted variables, such as air temperature, dew point, and precipitation parameters, to the METRo. The METRo then calculates the pavement temperature for the next seven days.

The RWS provide temperature data from federal roads, country roads, and international airports. The data undergo automated plausibility checks and are used to train the MOS and in METRo's data assimilation.

The MOS also calculates pavement temperatures. However, it has not been used operationally in the past due to a lack of quality testing. Thanks to years of automatically quality-assured measurement data at the RWS, the MOS can simulate the road temperature fairly well.

The RMSE was used to evaluate the accuracy of the METRo and MOS forecasts. It was based on 2.2 million quality-assured measured and modelled pavement temperatures at around 1,100 RWS stations between November 2023 and January 2024.

The RMSE was averaged over forecast times covering the next day and night, which is crucial for winter services. METRo's RMSE is 1.5°C, indicating high prediction accuracy given the pavement temperature measurement. MOS predictions are even better, with a calculated RMSE of 1.3°C. The largest differences between the models occur at noon and at night. Both models' forecasts are more accurate in regions of Germany where the data quality and quantity of observed pavement temperature at the RWS is higher.

Automated plausibility checks are crucial, as demonstrated by experiences with pavement forecasting in operational and pre-operational mode. If such checks are in place, a MOS can outperform a physical model, even with parameters that are difficult to observe. This is especially relevant for future developments in the field of AI.