The forecasting of short-term wind power is influenced by complex and changeable factors, such as climate environment, and meteorological characteristics. A single input of model features is difficult to meet the accurate forecasting requirements. However, to consider all the influencing factors, the input feature redundancy will increase the computational complexity of the forecasting model and have a negative impact on the forecasting results. Therefore, accurate and efficient feature selection is the key to improving the accuracy of data-driven wind power forecasting. In this paper, a short-term wind power forecasting method based on two-stage feature selection is proposed. Firstly, the meteorological factors and environmental factors affecting wind power forecasting are screened by using the Pearson Correlation Coefficient method. Secondly, a feature selection method of attention mechanism considering temporal correlation is proposed, which dynamically mined the potential correlation between wind power output and input characteristics from the perspective of time, and independently extracted the information of historical key moments. Finally, a bidirectional gated cyclic neural network (BIGRU) is selected as the forecasting model, taking into account the influence of historical data and future information. The results show that the proposed PCC-TA-BIGRU short-term wind power forecasting model can accurately select the factors affecting wind power forecasting and effectively improve the accuracy of short-term wind power forecasting.