Very High Resolution Coupled Weather and Wind Power Modeling
Given the potential role of specialized, high-resolution NWP to improve power forecasts, we will discuss such an approach for the prediction of wind energy for the five main islands in the Canary Islands archipelago, namely (from east to west), Lanzarote, Fuerteventura, Gran Canaria, Tenerife and La Palma. We will present how we are using fine-resolution NWP forecasts with machine learning (ML) methods such as linear SVRs, Elastic Net and Lasso to model wind energy from the NWP output and compare the resulting model forecasts with those derived from the coarser resolution ECMWF model.
Given the geography of the archipelago, we have applied NWP techniques at a turbulence scale for the Canary Islands. To resolve turbulent eddies, which contribute to ramp-up or ramp-down events, we have utilized the Large Eddy Simulation component of the WRF-ARW (version 3.3.1) community NWP model at 668m horizontal resolution. Fifty vertical levels are incorporated with at least ten in the planetary boundary layer to capture the conditions above, below and through the blade extent of the various turbines that are deployed. Several model configurations were evaluated to generate numerical experiments for retrospective analysis of significant ramping events in 2010 and 2011 as identified via hourly power data from the wind farm operators. The effort is focused on recent events to minimize the impacts of any changes in the wind farms and their operational conditions. Given an eventual goal of enabling operational forecasting, the experiments were done as hindcasts. Therefore, the model has been initialized with 0.5-degree data from NCEP's Global Forecasting System. In addition, scaling experiments were done to maximize the efficiency of the model configuration on a small Power7-based HPC cluster. The first phase was to build a hindcast-based climatology of six months to one year, with one 24-hour forecast per day and output available every 5 minutes in order to capture the transient nature of wind events. Output variables include those that drive the turbine energy extraction process, namely, turbulent kinetic energy, volumetric vorticity, horizontal and vertical wind velocities as well as surface wind gusts and a clear-air turbulence index.
The complex topography of the Canary Islands warrants the utilization of 90-m terrain data from NASA's Shuttle Radar Topography Mission. The oceanic influence on coastal winds and convection requires the NASA 1-km sea surface temperature analyses for model initialization. Since data from automated weather sensors are limited, both validation and variational assimilation are somewhat problematic. As a result of the large domain size, high-resolution and aspect ratio, additional challenges include avoidance of numerical instability and the computational cost associated with the hindcast generation, given the eventual transition from research to operations. These issues can be addressed through domain splitting/parallelization, one-way nesting, and adaptive time-stepping techniques.
The very fine resolution data from the turbulent-scale NWP input can be extremely large for the ML models. This clearly precludes the application with non-linear ML models, as their time complexity would be prohibitive. However, it has been observed in other fields that simple linear models can yield good results for problems with large dimensional inputs. We will apply two such linear methods here. Our first option will be linear Support Vector Regression (SVR), which uses the so-called hinge-loss that penalizes only forecast errors above a certain tolerance. The second option will be the Elastic Net and Lasso methods. Both combine a square error function with an L1 regularization penalty term. To this Elastic Net also adds a quadratic penalty, as done in ridge regression. Both approaches present two important properties. First the models are built solving a convex optimization problem and, thus, have a unique minimum value. Moreover, the hinge loss of SVR and the L1 regularization of Lasso and Elastic Net result in sparse final models with many zero coefficients that enable a fast application to new data and also can be exploited for ranking the predictive NWP variables.
The forecasts produced by the combination of the turbulent-scale NWP forecasts and the linear models will be compared with those obtained using ECMWF forecasts at a 0.25 degree resolution. The linear models using the former have so far been built using a comparatively short two-month training period. In contrast, the ECMWF forecasts were used with the non-linear models built using a 12-month training period. Nevertheless, and as we shall illustrate numerically, the performance of the models using the turbulent-scale data are comparable with that of the ECMWF-based forecasts. This clearly points to a significant potential for the application of fine-resolution NWP models for forecasting wind energy over islands or, more generally, isolated renewable energy systems.
In addition to the scientific and computational results to date, we will present its current status and our plans for future work.