Abstract: For the uncertainty of wind power, an ultra short-term wind power prediction method based on improved MFO(Moth-flame optimization) attention LSTM(Long short-term memory) is proposed. Firstly, the original power data is decomposed into multiple IMF(Intrinsic mode functions) components by using the adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) to eliminate the influence between different components, calculate the sample entropy of each component, and combine the approximate values of sample entropy to reduce the operation scale. Secondly, by introducing Chebyshev chaotic mapping, Cauchy mutation, Sigmoid function-based inertia weights to improve the traditional moth fire fighting algorithm, and the improved moth fire fighting algorithm is compared with the traditional MFO, PSO(Particle swarm optimization) algorithm. It is proved that its optimization ability has been greatly improved. Finally, the Attention mechanism is used to calculate the different weights of the hidden layer state of the LSTM neural network, and the improved moth-fighting optimization algorithm is used to optimize the hyperparameters of the Attention-LSTM. The combined IMF components are modeled respectively, and each component model is superimposed to obtain the final power prediction curve. Through the simulation analysis of the measured data of a wind farm in Jinzhou, the results show that the proposed model has high prediction accuracy and has certain reference significance for actual engineering.