Reservoir numerical simulation technology is primarily used on the prediction of production in salt resistant polymer flooding wells for Class II reservoirs in Daqing Oilfield. However, for achieving the dynamic adjustment requirements in production, it does not fit well when meet the complex numerical modeling and computational processes. Limited by the current situation of lacking research on machine learning applications for production prediction in salt resistant polymer flooding wells of Class II reservoirs, this study take the polymer flooding dynamic data which with evident multivariate characteristics time series in different period into consideration, and use the dynamic data from plain water flooding period to water-cut decline period, explore the improvement method which used convolutional neural networks(CNN) to select characteristic parameters from various influence factors which affecting the production in water-cut decline period. On the premise of consisting with the oilfield development regulation, a CNN-LSTM-Attention integrated learning model for production prediction in salt resistant polymer flooding wells of class II reservoirs is established, which combining Attention Mechanism with Long Short-Term Memory networks (LSTM). The comparison results demonstrate that the integrated learning model established in this paper achieve superior prediction performance when with identical data, can explore the change regulation of dynamic production data in the DXD block effectively, and improve the accuracy of production prediction during the water-cut decline period of polymer flooding significantly. It also provides a scientific model for production prediction in other similar salt resistant polymer flooding blocks.