With the growing dependence on wind power generation, improving the accuracy of short-term forecasting has become increasingly important for ensuring continued economical and reliable system operations. In the wind power forecasting field, ensemble-based forecasting models have been studied extensively; however, few of them considered learning the features from both historical wind data and NWP data. In addition, the exploration of the multiple-input and multiple-output learning structures is lacking in the wind power forecasting literature. Therefore, this study exploits the NWP and historical wind data as input and proposes a two-stage forecasting framework on the shelf of moving window algorithm. Specifically, at the first stage, four forecasting models are constructed with deep neural networks considering the multiple-input and multiple-output structures; at the second stage, an ensemble model is developed using ridge regression method for reducing the extrapolation error. The experiments are conducted on three existing wind farms for examining the 2-h ahead forecasting point. The results demonstrate that 1) the single-input-multiple-output (SIMO) structure leads to a better forecasting accuracy than the other threes; 2) ridge regression method results in a better ensemble model that is able to further improve the forecasting accuracy, than the other machine learning methods; 3) the proposed two-stage forecasting framework is likely to generate more accurate and stable results than the other existing algorithms.