Wind Energy Forecasting in Marine Tropical Environments: Can Lidar Data and Numerical Weather Prediction Help?

Wind forecasting in support of renewable energy production can be a challenging task, particularly in the context of an isolated island located in areas with few weather observations, such as over the tropical oceans. One such scenario is the island of Aruba located in the Caribbean Sea, where a wind project provides energy to the island population. To help fill the relative data void, a WindCube Scan lidar has been deployed to provide observations of the evolving trade winds out over the ocean. The instrument can observe winds in the marine boundary layer out to about 15 km away from the wind project and might provide an early warning for horizontally propagating disturbances to the flow upstream of the wind project. Efforts have been undertaken to determine how lidar observations can help improve wind forecasts at the wind project. Considered approaches include (1) adding a small portion of the upstream lidar observations as additional features in a hybrid machine learning forecast system, (2) taking a purely data-driven approach with deep-learning (DL) methods on the lidar data, and (3) numerical weather prediction (NWP).

First, we added the lidar features with the highest lag-correlation to the predictor set for the site-specific wind forecast. Second, we trained a deep neural network to predict the wind lidar data as if they represented image sequences and then added these forecasts as new inputs to the site-specific wind forecast. From these first two approaches, we observe only about 8-10 minutes of skill and see an overall reduction in the mean absolute error of the resulting wind power forecasts of 8-10%. For a different and complementary approach, the potential offered by NWP forecasts is also investigated, first by looking at the simple downscaling of a global weather model using Weather Research Forecast (WRF) kilometer-scale forecasts. The initial focus is on case studies corresponding to days with significant wind ramp events observed by anemometers at hub height. An analysis of wind speed forecast errors reveals that forecasts are generally unbiased, with a variance slightly larger against lidar data (error standard deviation around 2.0 m s^-1) compared to nacelle anemometers at the wind project (error standard deviation around 1.5 m s^-1). Forecast error statistics are similar whether forecasts are generated at a horizontal resolution of 3 km or 1 km. Also, the absence of smaller forecast errors early into forecasts indicate that initial conditions (downscaled from global model analyses) do not capture the detailed flow structure in the vicinity of the wind project. Finally, a closer inspection of lidar observations suggests that occurrences of the significant ramp events are related to rapidly evolving small-scale perturbations embedded in the trade winds. These results suggest that sub-kilometer scale forecasts are likely needed to obtain a more realistic representation of small-scale flow structures, combined with the assimilation of lidar data for more accurate initial conditions. Our early results show lidar observations have a modest positive impact on wind power forecasts over very short horizons, while requirements toward positive contributions from NWP have come into greater focus.