Abstract: Identifying the surrounding rock of TBM tunnels and predicting the excavation speed are crucial to ensuring the safe and efficient excavation of TBMs. In order to more accurately predict the TBM excavation speed, this study constructs a prediction framework based on PCA-BP and CNN-LSTM-Attention. First, Principal Component Analysis (PCA) is applied to process excavation parameters, such as cutter head torque, and surrounding rock parameters to explore the dynamic correlation patterns between surrounding rock parameters, excavation parameters, and excavation speed. Two principal component indicators are extracted, and a BP neural network is constructed to recognize surrounding rock grades. Then, the penetration, FPI, and TPI of each surrounding rock grade are used as input to predict excavation speed using the CNN-LSTM-Attention model. Finally, the effectiveness of the prediction framework is evaluated, the impact of surrounding rock grade classification on the prediction results is analyzed, and the model is compared with traditional regression models. The research results show that compared to BP neural networks and other traditional regression models, the CNN-LSTM-Attention model performs better. The R2 of the prediction results for each surrounding rockgrade is over 90%. Classifying surrounding rock grades significantly improves the prediction accuracy. The R2 for Grade Ⅱ, Ⅲ-a, Ⅲ-b, and Ⅳ surrounding rocks improves by 20%, 17%,17%, and 24%, respectively, compared to when no classification is performed. The mean square error (MSE) decreases by 88%, 86%, 81%, and 48%, respectively.