Postprocessing wind forecasts with deep learning in complex terrain

To improve and automate the quality of weather forecasts to the public, MeteoSwiss is redesigning its statistical postprocessing suite. The effort aims at producing calibrated probabilistic predictions to any arbitrary point in space and up to a 15-day lead time, by seamlessly integrating multiple numerical weather prediction models into a unique consensus forecast.

For hourly wind forecasts (mean, gust, and direction), the task is formulated as a regression problem in a supervised machine learning framework, where station measurements are used as labels, and co-located NWP forecasts as features. To improve the estimates at ungauged locations, additional static topographical features are derived from a 50m digital elevation model. The probabilistic component is included by training the neural network not to produce a deterministic prediction, but the parameters of a conditional probability function. To this end, the Continuous Ranked Probability Score (CRPS) is used as a loss function.

The dataset includes a range of surface parameters at hourly resolution produced by the operational forecasts from three NWP models (the deterministic COSMO-1 model, at 1 km horizontal resolution; the 21-member COSMO-E, 2 km; and the 51-member ECMWF IFS ENS at about 18 km). The data cover the whole of Switzerland over a period spanning more than four years (mid 2016 to end of 2020). Wind measurements from over 500 surface weather stations are included as reference dataset. The study uses a train-validation-test split in both space and time to assess the ability of the postprocessing model to generalize to unseen locations and times.

The results indicate that, despite the challenging nature of the problem, the postprocessing model can improve over the baseline NWP forecasts in terms of CRPS on the test set. In particular, the model is effectively correcting for biases relating to altitude error and other misrepresentations in the NWP topography. The results show that it is feasible to downscale numerical predictions to a substantially higher spatial resolution. Moreover, the conditional probabilities shows consistent improvements in terms of calibration, although it remains a significant portions of undetected peak events (positive outliers), possibly to be related to unpredictable phenomena (e.g., thunderstorm gusts). Finally, first results seem to suggest that the gain in prediction skill is mainly driven by a better statistical reliability rather than higher statistical resolution.