In the operation of nuclear steam supply systems (NSSS), operators must thoroughly assess potential future conditions before performing any manual control. In recent years, rapid advances in time series algorithms have provided richer tools for predicting trends in nuclear power plants. NSSS data is characterized by high dimensionality, multiple periodicities, and low variance. While existing models perform well on various tasks, However, in scenarios characterized by varying degrees of coupling, multiple periodic patterns, and low-rank structures, achieving satisfactory prediction performance still faces numerous challenges. The main issues include the introduction of excessive noise due to complex coupling, difficulty in effectively extracting multi-periodic patterns, and the amplification of minor errors over long time series, which can lead to significant prediction deviations. For these issues, first, divide the feature data based on the degree of coupling and apply different embedding methods to different types of data. Then, for the high-coupling features, use multi-time-step embedding to extract information from different cycles, while applying variable-level embedding to low-coupling data. Additionally, fuse data of different coupling degrees through global tokens. Finally, use a non-autoregressive decoder to reduce error accumulation. Experimental results show that NuPreX performs well on the NSSS dataset, accurately predicting key parameter trends and contributing to the safer operation of nuclear power plants.