Urban tunnel infrastructure, crucial for societal well-being, depends on reliable Tunnel Electromechanical Equipment (TEE), including ventilation, drainage, and lighting systems. A key challenge is these systems’ proactive and efficient maintenance, particularly under limited resources. This study introduces a novel deep learning-based multi-output prediction model developed to enhance the understanding and predictive accuracy Tunnel Boring Machine (TBM) performance, with a specific focus on machine wear and tear (y1) and adapting to ground conditions and geotechnical data (y2) in complex underground environments. The model employs an advanced deep learning approach, att-GCN, which innovatively integrates Graph Convolutional Networks (GCN) with a scaled dot-product attention mechanism. This combination notably improves model performance and interpretability. Experimental results indicate that att-GCN model achieves a Mean Absolute Percentage Error (MAPE) of 17.1% for y1 and 16.8% for y2, outperforming other established algorithms, including the Deep Neural Network (DNN)-Genetic algorithm hybrid. Furthermore, an online learning variant of att-GCN was developed that integrates real-time data during tunneling operations. This version demonstrated enhanced predictive accuracy, with a MAPE of 8.7% for y1 and 8.1% for y2. Applying att-GCN for real-time TBM performance estimation based on dynamic monitoring data offers significant insights for intelligent TBM control, improving construction efficiency and reliability.