The invention discloses a short-term wind power prediction method based on an improved neural network. In the method, a neural network algorithm is selected to be combined with a particle swarm optimization algorithm, the short-term wind power prediction method based on the improved neural network is built, and the method comprises the steps of optimizing off-line parameters and predicting an algorithm subject. A historical data set of the off-line parameter optimization contains two months of historical data and is updated one time every week, and the output optimal parameters are transferred to the predicted algorithm subject; a wind powder prediction model of the predicted algorithm subject is combined with the predicted wind speed and direction and the off-line optimized parameters, then the short-term wind power prediction is completed, and a wind power prediction result is output. The method is high in practicability and has the advantages of being fast in convergence speed, high in precision and globally optimal.