Abstract: Obtaining short-term probabilistic information of future wind power will improve the comprehensive energy dispatching of power grid. Therefore,a method of short-term wind power probability density prediction based on multi-level semantic attention mechanism is proposed. In order to obtain as much coding information as possible at the semantic level,Recurrent Highway Network（RHN）is introduced at the encoding stage. The underlying association information of input features is extracted to the maximum extent through the deep RHN network structure. A multi-level semantic attention mechanism is designed to integrate local attention vector under different semantic layers,which further strengthen the expression ability of coding feature vector. Combining the output of network with quantile regression and kernel density estimation methods,the prediction results and continuous probability density distribution of future short-term wind power are obtained. The experimental results show that the proposed method of short-term wind probability density prediction is superior to other comparison models,in terms of the accuracy of prediction value and the distribution of prediction results with uncertainty.