In order to solve the problem that the wind power is difficult to predict accurately due to the drastic change of wind speed, a short-term wind power prediction method based on the division of wind speed fluctuation characteristics is proposed. By setting the wind speed fluctuation threshold, the large fluctuation sequence in the historical data set is divided into three stages, including rising wind, fluctuating wind and descending wind. The dynamic time warping algorithm is used to mine the fluctuating wind similar data in the historical data, and a training sample data set is constructed combining with the corresponding historical wind power; a hunger game search algorithm is used to optimize the hyperparameters of the gated recurrent unit neural network, and a combined prediction model for three fluctuation stages is established. The wind power prediction values of different wind speed fluctuation processes are recombined in time series to obtain short-term wind power prediction results. The actual data of Xinjiang wind farm is used for simulation verification. The experimental results show that the proposed method can improve the prediction accuracy and generalization ability.