High-resolution (spatial and temporal) mesoscale models, such as the operational NCAR-Xcel 3.3-km WRF-RTFDDA, are able to simulate weather systems and their fine-scale features that cause wind power ramps. However, the model's ability in simulating the ramps is not obvious when the model performance is evaluated under traditional error statistics such as mean absolute error (MAE) and bias error (BE) because these metrics penalize their value when small spatio-temporal phase errors of the simulated weather features are present. Visual inspection of the 3.3-km model and observed wind speed time series often reveal a good one-to-one correspondence, especially for ramps caused by significant weather systems. In light of this fact, we verified the 3.3-km WRF ramp forecast by using a feature-based verification approach for three wind farms in the warm season of 2010 and the cold season of 2010-2011. The three wind farms are located in three disparate geographic regions: 1) lee side of the Colorado Rockies; 2) Plains region in Minnesota; 3) High Plains region in Texas where the low-level jet plays a major role. The feature-based approach computes a normalized metric (0 to 1) that is determined by comparison of four attributes between model and observed wind speed time series for a given ramp event: 1) the time window overlap between model and observed wind speed time series; 2) the ramp magnitude; 3) the maximum 1-h ramp rate; 4) the area under the curve of wind speed time series. The results reveal a very encouraging capability of the continuous 4-D data assimilation and forecasting cycles of the high-resolution WRF-RTFDDA system in forecasting wind ramps for 0 – 12h forecast ranges.