The integration of wind energy into the electrical grid is complex due to the high variability of wind fields when electricity should be available all time. In this context, accurate wind power forecasts have to be given 48 h before. However, the major difficulty is that wind power is highly nonlinear and non-stationary. This paper proposes a methodology to cope with these two issues by a two folded model. First, the time-series are decomposed into a low and high frequency components to deal with non stationarity. Second, the nonlinearities are accounted by regimes defined by wind direction. The model called D-TARX is compared with other models with only regimes, only decomposition, and none. Results show that our model outperforms other models according to statistical criteria. The methodology is straightforward while more work could be performed to continue towards accurate wind power forecasts.