A downscaled ensemble prediction system for offshore weather
Transcription
A downscaled ensemble prediction system for offshore weather
A downscaled ensemble prediction system for offshore weather forecasts PO.003 Torge Lorenz Uni Research Abstract Results from the ensemble forecast Sea-surface temperature perturbations A short-term weather forecast for the North Sea is being developed to reduce costs and improve safety of marine operations during installation and maintenance of offshore wind turbines. We apply the mesoscale Weather Research and Forecasting (WRF) Model to downscale the global ensemble prediction system from the European Centre for Medium-Range Weather Forecasts (ECMWF) to a horizontal grid spacing of 3 km in the North Sea. Near-surface wind speed is one of the key variables in determining safety of marine operations. To increase the ensemble spread and improve the probabilistic forecast of near-surface wind speed, we introduce perturbations of the sea-surface temperature (SST) into our downscaled ensemble. The SST perturbations are designed to represent smallscale SST features which are not resolved by the largescale ECMWF model but should be present on the smaller scale of our WRF ensemble. Their amplitude is based on statistical analysis of the Multi-scale Ultrahigh Resolution (MUR) Sea Surface Temperature Analysis on different spatial scales. Both the dynamical downscaling and the additional small-scale SST perturbations are shown to increase the ensemble spread in near-surface winds. The large-scale ensemble forecast at the ECMWF is initialized twice every day, at 00 and 12 UTC, with 50+1 ensemble members. For a test period of 5 continuous weeks in Autumn 2014, we dynamically downscale each of these global ensemble forecasts to a 3-km grid in the North Sea for a short-term forecast of 27 hours. The sensitivity of the ensemble spread and the ensemble mean to the downscaling and introduction of additional small-scale perturbations of sea-surface temperature will be investigated. The figures below show an example of the actual smallscale SST features which are missing when the WRF ensemble is initialized with the large-scale ECMWF ensemble data (Figure 4) and the small-scale SST perturbations which we introduce to represent these missing features (Figure 5). The aim of the perturbations is not to exactly recreate the missing features, but to create SST features with the same properties. These perturbations can then be used to create different initial states of SST, which on average all have the same spatial variability. Sea-surface temperature analysis When the regional WRF ensemble is initialized with large-scale SST data from the ECMWF ensemble, the SST in the WRF model will be too smooth and lacking genuine features on the 3-km scale of the WRF model. To quantify the loss in spatial variability in the SST, we compute the spatial variance between high-resolution MUR SST data (0.011 grid spacing) smoothed to 3km grid spacing and to ECMWF resolution. From Figure 1 below we can see that most of the additional variability in SST on the 3-km scale of the WRF model can be accounted for by a proportional relationship of 1.8 times the spatial variability present in the large-scale SST data inherited from the ECMWF ensemble. Our small-scale SST perturbations are thus designed to increase, on average, the spatial variability in SST by 80%. Figure 3. Original ECMWF EPS forecast for 10-m wind speed at the location of research platform FINO3. Shown are only the 21 ensemble members which are downscaled in the regional WRF ensemble. The example above illustrates how the spread in the original ECMWF ensemble forecast of 10-m wind speed (Figure 3) is increased by the dynamical downscaling with additional small-scale SST perturbations (Figure 4). An increase in ensemble spread is visible already in the first hours of forecast time. This indicates that the downscaling with additional SST perturbations may improve the common issue of under-dispersive shortterm ensemble forecasts. The mean value (blue line), here shown for the medians of the first 98 percentiles of the large-scale SST data set, is followed well by the according leastsquares linear fit (red line). The proportionality factor from the least-squares linear fit is 1.84 ( 0.01), with a near-zero constant of 2*10-4 K2. Figure 5 (below). Example of small-scale SST perturbations applied to the regional WRF ensemble forecast. Summary and Conclusions Figure 4. Forecast of 10-m wind speed from the regional WRF ensemble forecast with additional small-scale SST perturbations. Figure 1. Spatial variance of SST on the 3-km scale of the WRF model as a function of spatial variance of SST on the larger scale of the ECMWF ensemble forecasts. Figure 4 (above). Example of SST features missing when WRF is initialized with large-scale ECMWF EPS data. The improvements in terms of accuracy (RMSE) and other common ensemble verification metrics, such as continuous rank probability score, reliability and rank histograms, will be quantified using the in-house verification software from the Norwegian Meteorologisk Institutt, HARP (Hirlam-Aladin R Package). With HARP, the ensemble forecasts can be compared to a vast array of meteorological observations, both before and after downscaling and with and without additional SST perturbations. Most of the SST variance on smaller spatial scales can be accounted for by proportionality with SST variance on larger spatial scales (Figure 1). SST perturbations on the spatial scale of the downscaled ensemble forecast are developed exploiting this proportional relationship. These perturbations qualitatively represent the actual missing features on that scale well (Figures 4 and 5). The dynamical downscaling, together with the additional small-scale SST perturbations, improves the ensemble spread already in the first hours of the forecast (Figures 2 and 3). Both the dynamical downscaling and the SST perturbations help to accommodate additional uncertainty in the short-term weather forecast and add value to the risk assessment for marine operations related to offshore wind energy. Acknowledgments The MUR SST analyses are freely available at the Physical Oceanography Distributed Active Archive Center. The ECMWF EPS data were provided via an exclusive operational data stream from the European Centre for MediumRange Weather Forecasts.