During deep excavation construction, deformation impacts on adjacent structures are inevitably induced, and field monitoring data are often contaminated by noise that degrades prediction accuracy. To address these issues, this study develops a joint denoising strategy combining CEEMDAN, sample entropy-based adaptive IMF screening, and an improved wavelet threshold (IWT) function, followed by a Bayesian optimization-based extreme gradient boosting (BO-XGBoost) model for surface settlement prediction. The developed method automatically identifies high-noise IMF components via sample entropy and processes them using an improved threshold function that overcomes the discontinuity of hard thresholding and the constant bias of soft thresholding, thereby preserving useful information while suppressing noise. Experimental results on a Wuhan metro deep excavation project demonstrate that the CEEMDAN-IWT method improves SNR by up to 4.09% and reduces RMSE by up to 8.00% compared with conventional CEEMDAN-wavelet threshold denoising. The BO-XGBoost model trained on denoised data achieves an RMSE of 0.09 mm and a MAPE of 3.54%, outperforming BP, LSTM, standard XGBoost, GRU, CNN-LSTM, TCN, and simple regression baselines. Feature importance analysis confirms that the denoised data retain physical interpretability consistent with soil deformation continuity. This framework provides a practical solution with promising accuracy for deformation monitoring and early warning in deep excavation engineering.