Post-Processing Surface Temperature Forecasts with XGBoost

Reliable road weather forecasts are crucial for ensuring road safety and optimizing road maintenance. They help to time road saltings effectively and to keep roads ice-free while avoiding unnecessary saltings. The Finnish Meteorological Institute produces road weather forecasts using a specified road weather model. This model is a one-dimensional heat balance model that utilizes atmospheric forecast from large-scale numerical weather prediction (NWP) model as input. One of the most important outputs of the road weather model is the road surface temperature, as it is essential to know whether it is below or above freezing.

Neither NWP forecasts nor the physics of the road weather model are perfect, which leads to errors in the forecasts. However, it is possible to reduce these errors by using machine learning model to learn from the past behavior of the forecast errors under different weather conditions. In this study, we develop a machine learning model using XGBoost (Extreme Gradient Boosting) to predict the difference between observed and forecasted road surface temperatures. This model can then be used to correct forecast errors. XGBoost is a type of gradient boosting algorithm that builds an ensemble of decision trees sequentially, where each new tree focuses on correcting the errors made by the previous ones. Unlike random forests, which aggregate the predictions of many independent trees, gradient boosting improves prediction accuracy by iteratively minimizing error.

The road weather model was run to road weather station points in Finland using atmospheric forecasts from the MEPS (MetCoOp Ensemble Prediction System) control member as forcing. Road weather station observations were used in the model initialization. The forecasts were done for each control member run, which means 8 forecasts for each day. The XGBoost model was trained with data from four winter periods (September-May) between 2020 and 2024. The winter period of 2024-2025 will be used for the final evaluation of the model's performance. The variables used as predictors contained forecasted atmospheric values from MEPS, forecasted road surface temperatures as well as time- and location-specific variables. Before training or evaluation, data quality control was performed on the data to remove outliers and missing values. Hyperparameter tuning is being performed to optimize the model’s performance. Although the tuning and training process is ongoing, initial results show that the machine learning model can reduce forecast error. The mean absolute error (MAE) of road surface temperature predictions across all lead times (1–62 hours) was 1.7 °C on the training set. After optimizing the number and depth of decision trees, the MAE decreased to 1.5 °C. This error estimate is based on four-fold cross-validation, where in each fold, one winter period served as the validation set while the remaining periods were used for training.