Abstract:In view of the large randomness and non-stationarity of wind power, it is often difficult to obtain higher accuracy by directly inputting the prediction model, a short term wind power prediction based on feature selection and improved sparrow search algorithm (ISSA) optimizing convolutional neural network-bidirectional gated recurrent unit (CNN-BiGRU) is proposed. Firstly, variational mode decomposition (VMD) is used to decompose the original power into a set of subcomponents containing different information to reduce the non-smoothness of the original power series and improve the predictability, while the number of model decomposition is determined by observing the central frequency method. Secondly, the feature selection is carried out by random forest (RF) feature importance method for each component, and the influencing factors with higher contribution to the prediction of each component are selected from the meteorological feature factors such as wind speed, wind direction, temperature and air density to form the input feature vector. Then, the CNN-BiGRU prediction model of each component is established, and for the problem that the parameters of the neural network algorithm are difficult to adjust manually configured parameters with large randomness, ISSA is used to hyperparameterize the model to find the optimal combination of parameters adaptively. Finally, the prediction values of each component are superimposed to obtain the final prediction results. Simulation experiments are conducted with actual data from a wind farm in Inner Mongolia, China, and compared with various single and combined prediction methods. The results show that the proposed method has higher prediction accuracy compared with other methods, in which the average absolute percentage error value reaches 2.6440%, and the model accuracy and generalization are verified on other data, and the average absolute percentage error values are 4.3853%, The average absolute percentage errors were 4.3853%, 3.1749%, 1.5761% and 1.3588%, which were all within 5%, proving that the proposed method has better prediction accuracy and generalization ability.