Using the observation data of a certain wind farm in Gansu, an ultra short-term fast rolling wind speed forecast method is proposed based on the Long Short-Term Memory (LSTM) neural network model by evaluating the forecast accuracy under different input variables and model time window lengths. The results show that the change in wind speed itself plays a leading role in the ultra short-term wind speed forecast. Better simulation results can be obtained when input variables only include the wind speed data at different altitudes. By evaluating the impact of time window lengthL of LSTM on simulation capability, it is found that when L≤24 h, the model works well, which means that the change of ultra short-term wind speed is mainly related to the change of its own near time. When L>24 h, the simulation effect of all schemes decreases, which means overly long L reduces the simulation accuracy. By evaluating the wind speed forecast capability of LSTM in the next 4 hours, it is found that the simulation accuracy decreases gradually while the prediction time increases. The forecast ability is ideal in the next 2 hours, and the RMSE is less than 2 m·s^-1. LSTM proves economical and practical with low requirements for computing resources and has high application potential in operational wind speed forecast practice.