Due to its chaotic nature, the wind behavior is difficult to forecast. Predicting wind power is a real challenge for dispatchers who need to estimate renewable generation in advance to establish their strategies. To achieve an accurate wind power prediction, it is important to determine first which meteorological data need to be included in the predictor. For that purpose, this paper focuses on choosing the appropriate weather factors, namely spatially averaged wind speed and wind direction. These factors are selected according to correlation and importance measures. Then, the random forest method is proposed to build an hour-ahead wind power predictor. The random forest does not need to be tuned or optimized, contrary to most other learning machines. Both point and probabilistic forecasts are performed using the same inputs. The emphasis is put on the effect of wind speed and direction on the model performance, and the immunity of random forest to irrelevant inputs. The wind data used to test the proposed model are taken from Sidi Daoud wind farm in Tunisia. Results show an interesting improvement of forecast accuracy using the proposed model, as well as an important reduction of the different error criteria compared to classical neural network prediction. (C) 2017 Elsevier Ltd. All rights reserved.