Abstract: The rolling mill vibration seriously affects the quality of thin-gauge high-strength steel products in the high-speed tandem cold rolling process. Regarding the problem, a vibration energy prediction method that integrated mechanism and data-driven approaches was proposed in this paper. It aims to break through the key bottlenecks that restrict the stability of production and the improvement of efficiency. Traditional analytical methods based on steady-state assumptions struggle to accurately describe dynamic roll gap characteristics under vibration conditions, while purely data-driven models lack physical mechanism interpretation. Therefore, a parabolic dynamic velocity field considering work roll vertical vibration was constructed firstly. An analytical model encompassing internal plastic deformation, friction, shear, and tension powers was derived based on the upper-bound approach. It can reveal the vibration mechanism from an energy perspective. Subsequently, 12 critical parameters, such as roll gap change rate, rolling force and back tension stress, were selected using the maximal information coefficient. Based on this, a bidirectional long short-term memory （BiLSTM） network integrating the attention mechanism was established to extract deep temporal features. Learning rate decay and the Dropout strategy had also been employed to optimize the training process. It demonstrates that the proposed Att-BiLSTM model performs excellently in both prediction accuracy （correlation coefficient R=0.968, determination coefficient R²=0.925） and computational efficiency （prediction time of 3.5 ms）. The model has excellent engineering practicability and can provide early warning for self-excited vibrations. Meanwhile, the feature importance analysis further clarifies that the roll gap change rate and the rolling force are the key influencing factors, and reveals that the recent historical information has a stronger contribution to vibration identification. Through the effective integration of mechanism and data, this work not only achieves high-precision and high-efficiency prediction of the rolling mill vibration energy, but also provides a new approach for vibration mechanism explanation and process optimization. It has significant value for promoting the high-quality production of strips.