Abstract: Aiming at the problems of insufficient utilization of historical information and the fixed multi-dimensional input weight ignoring the importance of features in different time dimensions in current wind power prediction process,a wind power prediction model based on feature variable weight is proposed. Random forest（RF）is used to analyze the degree of influence of wind speed,wind direction,temperature and other meteorological characteristics at different heights on the wind power and cumulative contribution rate is used to complete the extraction of meteorological features. Singular spectrum analysis（SSA）is used to denoise the extracted features and historical power information,and the denoised data is used as input to establish a cascaded FA-CNN-LSTM multivariate prediction model to predict ultra-short-term wind power. By adding feature attention mechanism（FA）to CNN-LSTM network to adaptively mine feature relationships at different time,the weights of input features at different time dimensions can be dynamically adjusted to enhance the attention of key features at prediction moment,and the prediction performance can be improved. The case study shows that the proposed method can effectively improve the accuracy of ultra-short-term wind power prediction.