Improving probabilistic weather forecasts in the Netherlands

Abstract

Forecasting wind speeds is important because of its large impact on society. The forecasts are issued by Numerical Weather Prediction (NWP) models. These NWP models often contain biases and have errors in dispersion, therefore they are subjugated to statistical post-processing. A commonly used method to perform post-processing is ensemble model output statistics (EMOS), where the goal is to fit the parameters of a probability distribution based on NWP output. In early versions of EMOS, linear regression was employed for this task. In recent approaches, more complex models such as neural networks have been introduced. While neural networks are able to significantly improve performance up to medium range wind speeds, they struggle with high wind speeds. The models are often trained using the continuous ranked probability score (CRPS), a proper scoring rule that equally weighs all possible forecast values. In this work, we propose using a weighted version of the CRPS (wCRPS) to address the challenges associated with extreme wind speeds. The wCRPS is a proper scoring rule that emphasizes particular regions of the forecast through a weight function. We also explore different parametric distributions, namely the truncated normal (TN), log-normal (LN), generalized extreme value (GEV) and mixture distributions. Our findings suggest that using the wCRPS with an appropriate weight function can enhance performance on extremes. However, for models using linear regression, we observed a body-tail trade-off, where increased performance on extremes came at the cost of worse predictions for average wind speeds. We developed an approach where the weight function is selected based on user preference by selecting hyperparameters using a multi-objective optimization algorithm. For the convolutional neural network-based models, we found that with an appropriate weight function the performance on extremes could be increased. Further investigation on the weight function of neural network-based models is advised, as the best choice of weight function may not have been included in our search space. Additionally, the best-performing weight function is shown to be model-specific. Regarding the choice of distribution, no significant effect was observed.