Ultra-short-term wind speed prediction is of great significance to the dispatch and control of wind farms and the safe operation of smart grids. Existing prediction methods based on neural network algorithms suffer from the shortcoming of lack of guidance in hyperparameter optimization. This paper proposes an ultra-short-term wind speed prediction method based on a genetic algorithm-long short-term memory (GA-LSTM) network hybrid model. The number of neurons in the hidden layer of LSTM and the dropout rate of the dropout layer are optimized by GA to improve prediction accuracy. A preservation strategy is proposed to record the excellent LSTM structures that may appear during the training process and reduce the computational effort by half. The proposed method is validated using actual data from a wind farm. The results show that GA can effectively optimize the hyperparameters of LSTM, and the GA-LSTM hybrid model has good accuracy in wind speed prediction tasks from one to four steps ahead.