<p>Real-time intelligent perception of complex surrounding rock conditions is essential for achieving automated and intelligent control of Tunnel Boring Machines (TBMs). Leveraging the extensive big data obtained from the TBM excavation of a water resources allocation project, this study applied deep learning algorithms to enable real-time characterization of surrounding rock features and to predict key tunneling parameters. The raw data were first processed, followed by the establishment of pre-training models for surrounding rock feature characterization based on three deep temporal algorithms. These models enabled real-time acquisition of surrounding rock feature vectors. The results of Multi-Layer Perceptron (MLP) surrounding rock classification based on the derived feature vectors indicated that the Long Short-Term Memory (LSTM) pre-training model exhibited the highest prediction accuracy among the three models. The effectiveness of the surrounding rock feature vectors in characterizing rock conditions was further validated through similarity analysis and dimensionality reduction visualization. Subsequently, a baseline model for real-time prediction of tunneling performance parameters was developed, along with two improved models that incorporated either the rock grade or the surrounding rock feature vectors as additional input. Experimental results demonstrated that both improved models effectively enhanced prediction accuracy and stability. Notably, the model utilizing surrounding rock feature vectors as input enabled real-time performance prediction, thereby overcoming the inherent delay associated with traditional rock grade acquisition.</p>