As an important dynamics parameter， stiffness has a key role in turbopump vibration reduction. Therefore， a prediction model that incorporates attention mechanism and bi-directional long short-term memory （BiLSTM） neural network is proposed. Vibration information is input and the time-related historical features are effectively extracted by the LSTM network. Subsequently， the BiLSTM network is built by the inverse superposition of the two-layer LSTM network. This is to accommodate the complex and lengthy sequences of dynamics information， and thus the nonlinear features between parameters are extracted. The weights of the features are obtained by introducing the Attention layer， which will enhance the key information. Finally， the prediction model is trained with turbopump dynamics data. The results show that for turbopump stiffness characteristics， the Attention-BiLSTM model has a significant advantage in sequence data processing， with a mean absolute percentage error （MAPE） of 2.194 5%. In contrast， the MAPEs of RNN， LSTM， and BiLSTM models are 10.497 7%， 5.497 3%， and 2.798 6%， respectively. It can be seen that the method effectively avoids the complex dynamical inverse problem solving and achieves the dynamic identification of nonlinear parameters.