Abstract: In order to improve the accuracy of wind power prediction, an ultra-short-term wind power prediction based on the overall average empirical mode decomposition (CEEMDAN), the Permutation Entropy (PE), the Wavelet Packet Decomposition (WPD) and the multi-objective optimization is proposed. First, the signal processing method consisting of the CEEMDAN, the PE and the WPD is used to reduce the randomness and volatility of the original wind power signals; then, the decomposed subcomponents are fed into the Long/Short-Term Memory (LSTM) network, and an improved Elite T-distribution Sparrow Optimization Algorithm (ETSSA) is used to optimize the number of hidden layer units of the LSTM to improve the prediction performance of the LSTM network; finally, the loss function is optimized with the three optimization objectives of accuracy, stability and pass rate added into it at the same time to improve the prediction performance of the model comprehensively. The experimental analysis of the measured data from a wind farm in a region in Inner Mongolia shows that, compared with other classical prediction methods, the proposed method has a more significant effect on improving the wind power prediction performance and a better prediction effect under different seasonal wind conditions.