Sequential Recommendation focuses on using the user’s historical interaction sequence to predict their next item. Usually, the user’s global and local preferences will work together to varying degrees on the recommendation results. The existing sequential recommendation models tend to ignore the problem of attention imbalance when modeling users’ global preferences, and the influence of user behavior uncertainty is ignored when modeling user’s local preferences. To achieve this, we propose a novel parallel architecture that captures the global representation through attention weight normalization and dynamic local representation, allowing for adaptive fusion, termed as \(\textbf{PNSR}\) . Specifically, for the global representation, we capture global information using MLP and improved multi-head attention to normalize the weights between different item feature representation in the interaction sequence. For local representation, we integrate local convolution into Light-SA and employ MLP to combine candidate items with local preference representations, enabling the capture of real-time changing local semantic information. Additionally, we design a perceptual fusion unit to adaptively merge global and local representations. We perform experiments on three commonly used public datasets, showing that our approach surpasses most baseline models. The source code for the model implementation is available at https://github.com/ZhuYiZ-1205/PNSR .