The invention belongs to the technical field of electric power system forecasting and controlling, in particular relating to a method for forecasting short-term power in a wind power station. The method comprises the following steps: firstly, acquiring actual measured wind speed and power data, and processing the acquired data by adopting a differential autoregressive moving average model and a power forecasting model based on the combination of wavelet transformation and a neural network, thus obtaining a super short-term forecasting power; then acquiring the wind speed, wind direction, air temperature, humidity and atmosphere pressure data of different heights, processing the acquired data by adopting a BP (back propagation) neural network, a self-adaptive fuzzy neural inference system model and a least square support vector machine model through a power combination forecasting method based on a maximal information entropy principle, thus obtaining short-term forecast power; and finally processing the super short-term forecast power, the average error of the super short-term forecast power, the short-term forecast power and the average error of the short-term forecast power which are utilized as input by a combination forecasting method based on the BP neural network, thus obtaining the final short-term forecast power. According to the invention, the accuracy of forecasting power in 72 hours in the future in the wind power station is improved, and a reference is provided for the resonable dispatch of a power grid.