As long as the wind speed can be forecasted， the output of a wind power unit can be calculated according to the unit's power curve. A short-term combination forecasting model of wind speed is developed based on the seasonal periodicity and time-continuity of wind energy. With the pattern recognition technique， two kinds of samples to seasonal periodicity and time-continuity are selected separately. The two kinds of samples are used respectively to train two back propagation（BP） neural network models to obtain the lateral and vertical wind speed forecasting values. By importing the two values to a BP neural network again， the wind speed value is finally predicted. Forecasting results of a domestic wind farm indicate that the proposed combination forecasting method is valid and effective in short-term wind speed forecasting.