Enhancing ERA5 wind speed time series variability through ramp events detection and correction

Wind variability poses challenges for planning and designing reliable power system with high share of renewables. Their representation in reanalysis dataset such as ERA5 remains an active research area, as it can provide reliable long-term mean wind speed estimates but is missing high frequency variability. The aim of this study is to build a machine-learning model that can correct the representation of variability in the ERA5 wind speed time series. The objective is to reduce not only the averaged error (e.g. the root mean squared error or the mean bias), but also the time series variability characteristics (e.g. the autocorrelation error). Our study focuses on five sites in the North and Baltic Seas. Wind speed measurement were collected using floating lidar systems and a meteorological mast at 100-m height, and these were selected at hourly timestamps to be used as the ground truth. From this observed wind speed dataset, our first objective is to find the periods of high variability that must have been presented in the ERA5 wind speed time series, which we refer to as ramp events. The predictors are constructed from physically motivated variables in the ERA5 dataset, and their relevance and redundancy are evaluated through a process of feature selection. This approach provides insight into the atmospheric mechanisms that drive variability in ERA5 wind speed and offers an explainable basis for improving variability representation and, ultimately, long-term mean estimates derived from ERA5 reanalysis data.