Aimed at the problems that PM2.5 concentration prediction model is unstable in efficiency and poor in generalization ability, a TwoDirection Attentionbased Recurrent Neural Network (TDA-RNN) is proposed based on the cyclic neural network and attention mechanism. Firstly, the temporal attention and the category attention in inputting data through attention mechanism are obtained in making them fused by TDA-RNN model, and then the fused data are encoded to obtain intermediate features through feature encoder. Finally, the intermediate features are fused with historical information of PM2.5 concentration, and the predicted values are obtained by feature decoder. The PM2.5 concentration in Beijing is predicted by several models. The results show that the prediction accuracy of TDA-RNN is higher than that of the Back Propagation Neural Network, the Long ShortTerm Memory, the Gate Recurrent Unit and the Moving Average model. In the antijamming test, while the input data having noise factors, the prediction accuracy of TDA-RNN decreases slightly, but still higher than that of other models. The TDA-RNN proposed in the paper is strong in feature extraction ability and high in prediction accuracy.And this can also be applied to multivariate time series prediction in other application scenarios.