A new pan-European dataset for gridded daily average wind speed based on in-situ observations

Objective

The purpose of the European Copernicus Climate Change Service (C3S) is to support society by providing information about the past, present and future climate. For the service related to in-situ observations, one of the objectives is to provide high-resolution (0.1x0.1 and 0.25x0.25 degrees) gridded wind speed fields. The gridded wind fields are based on ECA&D daily average station observations for the period 1970-2020.

Research question 

We address the following research questions: [1] How efficiently can we provide the gridded wind fields as a statistically reliable ensemble, in order to represent the uncertainty of the gridding? [2] How efficiently can we exploit high-resolution geographical auxiliary variables (e.g. digital elevation model, terrain roughness) to augment the station data from a sparse network, in order to provide gridded wind fields with high-resolution local features?

Approach

In our analysis, we apply greedy forward selection linear regression (FSLR) to include the high-resolution effects of the auxiliary variables on monthly-mean data. These data provide a ‘background’ for the daily estimates. We apply cross-validation to avoid FSLR over-fitting and use full-cycle bootstrapping to create FSLR ensemble members. Then, we apply Gaussian process regression (GPR) to regress the daily anomalies. We consider the effect of the spatial distribution of station locations on the GPR gridding uncertainty.

The goal of this work is to produce several decades of daily gridded wind fields, hence, computational efficiency is of utmost importance. We alleviate the computational cost of the FSLR and GPR analyses by incorporating greedy algorithms and sparse matrix algebra in the analyses.

Novelty   

The gridded wind fields are calculated as a statistical ensemble of realizations. In the present analysis, the ensemble spread is based on uncertainties arising from the auxiliary variables as well as from the spatial distribution of stations.

Cross-validation is used to tune the GPR hyper parameters. Where conventional GPR hyperparameter tuning aims at an optimal prediction of the gridded mean, instead, we tune the GPR hyperparameters for optimal prediction of the gridded ensemble spread.

Building on our experience with providing similar gridded climate data sets, this set of gridded wind fields is a novel addition to the E-OBS climate data sets.