Immersive experiences, as an innovative display method, have been widely applied in exhibition spaces such as museums and art galleries, aiming to enhance audience engagement and immersion. With the rapid development of artificial intelligence technology, narrative generation systems based on deep learning have emerged as an important approach to improving immersion. This study optimizes the narrative generation system using the TDBN algorithm, which combines Transformer and Deep Belief Network (DBN). The research results indicate that TDBN excels in narrative coherence, creativity, and user satisfaction. Specifically, TDBN scores 0.472, 0.765, and 2.43 on the BLEU-4, ROUGE-L, and CIDEr metrics, respectively, significantly outperforming other comparison models. In user immersion assessments, TDBN achieves an immersion score of 4.6, an information coherence score of 4.7, and scores above 4.0 for interactivity and user satisfaction, demonstrating its effectiveness in enhancing user experience. In summary, through the fusion of multimodal data and optimization of deep features, the TDBN algorithm can significantly improve the performance of immersive narrative systems, providing effective support for generating personalized, coherent, and engaging narrative content. It exhibits strong application potential.