With the further promotion of the "Double Carbon" goal, China's wind power industry has developed rapidly in recent years. How to accurately and effectively predict wind power is crucial to realize the safe grid connection of wind turbines and maintain the stable operation of the system. Aiming at the problems such as input feature redundancy, insufficient generalization ability, and insufficient capture of inherent wind power output characteristics in existing wind power forecasting methods, a short-term wind power prediction model based on feature selection and similarity based fusion using long short term memory network-attention mechanism (LSTM-AM) is proposed. Firstly, least absolute shrinkage and selection operator (Lasso) regression is used to optimize input features to reduce redundancy. Then, the long short term memory network-attention mechanism（LSTM-AM） fusion network model is established using long short term memory and attetion mechanism. Finally, similar historical samples are extracted by Euclidean distance calculation and weighted with the model output as the final predicted value. The experimental results show that compared with traditional methods, the prediction accuracy of the proposed method is significantly improved, and it shows higher accuracy in wind power prediction, which can provide support for power system planning and operation and the in-depth application of renewable energy.