Speaker
Description
Weather forecasts have been steadily improving in quality over the last decades. At the same time, users demand more and more localized weather information for any locations and the large amount of weather forecast data and the pressure on resources demand the automation of the forecasting process. The ongoing improvements in forecast quality are due to advances in numerical weather prediction (NWP) and the advent of ever more powerful supercomputers that allow to simulate future weather and its uncertainty with increasing resolution. Such physics-based computer models, however, are not free of systematic errors. Further, even operational models with a spatial resolution of as high as 1km are not able to capture the full complexity of the topography such as the Alps. Statistical postprocessing can therefore be used to calibrate NWP forecasts to further improve forecast quality and adapt it to local conditions and increase their value consequently. Here we explore deep learning approaches, artificial neural networks (ANN) in particular, to calibrate deterministic and probabilistic forecasts of several parameters like temperature and wind in Switzerland. The machine learning approaches allow us to accounte for local effects in complex terrain such as valley direction and exposure. Further, the methods are generalized to calibrate the forecasts at unobserved sites allowing the provision of improved predictions at any location in Switzerland. In all cases, ANN lead to significant improvements over the direct NWP output. The improvement through ML is comparable in magnitude with improvements achieved with classical postprocessing approaches. Cross-validation results show further, that the ANN-based postprocessing provide a substantial added value at unobserved sites compared to direct model output.
