Short-term ensemble prediction of surface winds in the North Sea at Uni Research Computing
by Torge Lorenz, Uni Research Computing - [torge.lorenz AT uni.no]
Torge Lorenz is a researcher at Uni Research Computing, financed by NORCOWE. He will be defencing his PhD thesis at the University of Bergen this year.
Near-surface wind speed is a crucial parameter for the safety of marine operations related to offshore wind energy, such as turbine installation and maintenance. At Uni Research Computing, we are designing a high-resolution ensemble prediction system for 10-meter wind speed in the North Sea. We apply the mesoscale atmospheric Weather Research and Forecasting (WRF) Model with a horizontal grid spacing of 3 km to downscale dynamically global ensemble predictions from the European Centre for Medium-Range Weather Forecasts (ECMWF) for short-term forecasts of 27 hours ahead.
The ECMWF ensemble consists of several numerical weather predictions that are computed over the same period of time. The different ensemble members are subject to small differences, or perturbations, to the model's initial state and the physical processes that are parameterized within the model. These perturbations are designed to account for uncertainties in the forecasts, and by employing such an ensemble of weather predictions, one obtains probabilistic forecasts of the future state of the atmosphere.
To obtain high-resolution ensemble predictions, we set up a regional WRF domain to cover the North Sea and its surroundings with a spatial resolution of 3 km. The extent of the North-Sea domain is shown in Figure 1. The limited-area WRF model receives its initial and boundary conditions from the different members of the global ECMWF ensemble prediction system, and an ensemble of dynamically downscaled weather predictions is computed.
Figure 1. Extent of the limited-area WRF domain. The model's landmask is shown in gray. The horizontal grid spacing within the domain is 3 km. While the North Sea is the area of main interest, the domain has been extended to the west, the predominant wind direction, to allow the model to resolve incoming cyclones from the Atlantic and topographically-induced features around the British Isles.
In our downscaled ensemble, we also investigate the outcomes of time-lagged initialization. ECMWF ensemble forecasts are initialized twice a day, at 0000 UTC and 1200 UTC. This means that successively initialized and downscaled forecasts have a certain overlap in simulation period. For as long as there is overlap between two successive ensemble forecasts, these forecasts may be combined to an even larger ensemble, thereby obtaining a so-called time-lagged ensemble. Our downscaled ensemble forecasts are initialized every twelve hours, have a simulation period of 27 hours and will thus have an overlap of 15 hours simulation time with the ensemble that is initialized next. The practice of time-lagged initialization is a way to address uncertainty in the initial-state analysis, as forecasts for the same point in time are based on different initial states.
The ensemble spread is defined as the square root of the mean variance over the ensemble. It provides information about the differences between the different forecasts within the ensemble and thus about the uncertainty in the forecasts. A large ensemble spread means that the different ensemble members diverge in their predictions and that there is a high uncertainty associated with the forecast. A low ensemble spread translates to rather high confidence in the forecast, as the forecasts from individual ensemble members are close to each other even though the members' initial states and physical processes have been perturbed.
A commonly used metric for the accuracy, or skill, of an ensemble prediction system is the root-mean-square error (RMSE) of the ensemble mean. Although the ensemble mean does not necessarily represent a physically consistent state of the atmosphere, it is usually more skillful than any single ensemble member. The overall reliability of an ensemble prediction system may then be assessed by the so-called spread-skill ratio, i.e. ensemble spread divided by RMSE of the ensemble mean. In a reliable ensemble prediction system, the ensemble spread and the RMSE of the ensemble mean are thought of to equal one another, and the spread-skill ratio should be 1. This is because the RMSE for an unbiased estimator, such as a well-calibrated ensemble, is equal to the square root of the variance, which is the definition of the ensemble spread.
The spread-skill ratio for our downscaled ensemble predictions of 10-m wind speed is shown in Figure 2, as a function of forecast lead time.
Figure 2. Spread-skill ratio for predictions of 10-m wind speed from the ECMWF ensemble (blue), WRF 3-km ensemble (black) and WRF time-lagged ensemble (yellow). Modelled wind speeds are validated against observations at 248 measurement stations, from ensemble predictions initialized between 17 October and 10 November 2014.
The dynamical downscaling improves the spread-skill ratio considerably: the ratio is almost doubled at forecast lead times greater than zero. Changes in the spread-skill ratio are due to both an increased ensemble spread and a decreased RMSE of the ensemble mean. The ensemble spread is improved by the downscaling, as the decreased grid spacing allows for a more detailed representation of mesoscale flow features, and wind fields in the 3-km WRF ensemble are less smooth than those in the larger-scale ECMWF ensemble. The RMSE is improved due to the enhanced representation of the coast line, coastal terrain and topographically-induced flow features.
The time-lagged initialization improves the spread-skill ratio further. The ensemble spread in the time-lagged WRF ensemble is increased because of two reasons. First, half of the ensemble members had more simulation time available to develop divergent model states. Second, the time-lagged ensemble members are initialized with different initial-state analyses, which may result in more divergent predictions, depending on the forecast uncertainty associated with the model's initial state.