Speaker
Description
Current state of the art numerical weather prediction models as COSMO-1 running at very high horizontal resolution (~1km) and using sophisticated assimilation techniques still show substantial biases in wind forecasts, above all in very complex topography such as the Alps.
The INCA-CH system, running at MeteoSwiss, tries to correct the COSMO-1 bias on a 1km grid using the available surface observations. The observed bias at the stations is then interpolated in space using inverse distance weighting (IDW). The observed limited spatial correlation of the COSMO-1 error and the very low spatial representativeness of weather stations suggest that the spatial interpolation using IDW could result in unwanted effects, and might even deteriorate the raw COSMO field.
Compared to post-processing of wind at individual weather stations, no established methods are available to produce bias corrected gridded fields. Here, we will explore machine learning algorithms to achieve this task. Specifically, we compare the currently operational INCA algorithm with a 2 step machine learning system. The first step consists in the removal of the systematic bias, correcting COSMO-1 wind estimations. An artificial neural network is trained on a complete year of data, considering COSMO-1 variables and several high resolution topographical parameters as predictors. Then, a second step is applied in order to correct the forecast error too. Another machine learning model is trained in real time and on single time steps, using the corrected wind of the first step as additional predictor.
The contribution provided by machine learning methods for post-processing of wind is evident, leading to a consistent reduction of the error with respect to station measurements. But the use of machine learning as spatial interpolation technique remains very challenging, in particular in regions characterized by complex topography.
