<p>Pipes conveying fluid are important components in modern engineering systems, such as aerospace, marine and nuclear power equipment. Under external excitations, these pipes may exhibit complex nonlinear vibration phenomena. Predicting the long-term nonlinear vibration behavior of such pipes based on limited data has become a key scientific challenge in both academia and industry. To address this issue, a novel hybrid neural network model is proposed, by integrating three core architectures: the convolutional neural network (CNN), long short-term memory network (LSTM), and Transformer model. It is enhanced with attention mechanisms and optimized via a Bayesian approach for hyperparameter tuning. Specifically, the model leverages CNN for extracting spatial feature, LSTM for capturing temporal dependencies, and Transformer for modeling global dynamic characteristics. Two base architectures, i.e. the CNN-Attention-LSTM-Attention and the CNN-Attention-Transformer, were constructed integrated through a multi-expert feature fusion strategy, resulting in the Bayesian-optimized multi-expert feature fusion (BO-MEFF) model. The computational validation and autoregressive predictions were conducted using limited vibration displacement response data from the nonlinear forced vibration of a fluid-conveying pipe. The results demonstrate that the predictions of the proposed model are in close agreement with those from classical numerical simulations, confirming its stability and effectiveness in capturing nonlinear dynamic characteristics under varying excitation conditions. Thus, this study establishes a novel neural-network-based forecasting framework for predicting nonlinear vibrations in the dynamical system of pipes conveying fluid, and is expected to offering computationally efficient and accurate alternative for addressing complex fluid–structure interaction problems in slender structures.</p>