The safe operation of nuclear power plants relies on accurate prediction of key parameters, especially under complex conditions and unexpected accidents. Traditional time-series prediction models, such as LSTM and GRU, face challenges in capturing long-term dependencies and maintaining computational efficiency. The Informer model, based on Transformer architecture, addresses these limitations with its sparse self-attention mechanism and multi-scale learning capabilities, improving both prediction accuracy and computational efficiency. This study applies the Informer model to forecast critical nuclear power plant parameters, effectively modeling long-sequence dependencies in dynamic scenarios. Experimental results demonstrate that compared to LSTM, the Informer model significantly enhances prediction accuracy and reduces computational costs. Additionally, the impact of prediction step size on model performance is analyzed, revealing an optimal step size that balances dynamic response requirements with real-time performance. By identifying trends and risk inflection points in critical parameters, this method provides scientific support for operator decision-making, particularly in reactor shutdown scenarios. The approach enhances situational awareness and emergency response, contributing to safer and more efficient plant operation.