Aiming at the prediction of industrial process quality with noise interference and delay, in this paper,we propose a method of stacking noise reduction auto-encoder embedded with attention mechanism and bidirectional long short-term memory network. Firstly, the AE model is constructed in an unsupervised manner, and the industrial data is reconstructed with Gaussian noise processing to achieve denoising and de-redundancy. Secondly, embed the attention mechanism to reconstruct the weight allocation of process variables to achieve deep feature extraction. Finally, the BLSTM network is used to learn and reconstruct the time series trend characteristics of the data to overcome the delay between the data, and fully explore the potential relationship between the process variables and the quality variables, and finally achieve accurate prediction. Through the simulation experiments of single-mass variable prediction of debutane process and multivariate mass prediction of sulfur recovery process, it is verified that the proposed method has a more accurate prediction effect than other methods such as BP, LSTM, BLSTM and DAE-BLSTM.