Hybrid Deep Learning Model for Accurate Settlement Forecasting of Metro Tracks under Canal Diversion Engineering

The structural stability of metro systems is essential for safe and reliable urban rail operation. Large-scale underground construction may influence existing metro lines, making accurate settlement prediction necessary. Traditional empirical and numerical methods often fail to capture long-term settlement behavior. This study predicts track bed settlement of Hangzhou Metro Line 1 using monitoring data collected during the Grand Canal diversion construction. A hybrid model (CEEMDAN-BWO-BiLSTM-ATT model) integrating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, Beluga Whale Optimization, Bidirectional Long Short-Term Memory, and an attention mechanism is developed. Results from four monitoring points along the up line show good performance, with an average R 2 of 0.962, RMSE of 0.076 mm, MAE of 0.066 mm, and MAPE of 6.383%. Validation using a down-line monitoring point confirms accuracy and generalization. The results indicate that the model captures nonlinear settlement behavior and provides a reliable data-driven approach for metro deformation prediction. • A CEEMDAN–BWO–BiLSTM–ATT hybrid model is proposed for long-term metro track settlement forecasting. • Multi-scale decomposition, BWO-based tuning, and attention-enhanced BiLSTM are integrated to improve prediction. • Field monitoring at multiple points confirms reliable performance for practical settlement prediction and early warning.