Mechanism-informed prediction of complex karst water inflow: overcoming strong nonlinearity via a hybrid deep learning framework

Abstract To address the challenges of strong nonlinearity and non-stationarity in predicting water inflow in karst engineering, this study investigates an underground intake pumping station of the Shanxi Yellow River Diversion Project, systematically revealing the nonlinear causes of inflow dynamics. The research reveals that the low-permeability boundary formed by lining construction transforms karst seepage from matrix diffuse flow to conduit concentrated flow, significantly amplifying the system's pulse response to extreme meteorological forcing. Based on this mechanism, a hybrid predictive framework integrating CEEMDAN, LSTM, and a Self-Attention (SA) mechanism was developed, with a focus on comparing the predictive performance of two SA topological configurations: pre-positioned and post-positioned. Experiments confirm that the topological placement of SA significantly impacts nonlinear capture capability: when positioned subsequent to the LSTM layer as a "global feature recalibrator," it effectively compensates for the LSTM gating mechanism's "smoothing effect" on low-energy pulse signals, achieving high-fidelity capture of nonlinear abrupt change features through cross-temporal association reweighting. Results show that the post-positioned architecture achieves high accuracy with R² = 0.902, MAE = 0.002, and RMSE = 0.003. Compared to the pre-positioned architecture, the post-positioned configuration further reduces RMSE and MAE by approximately 50% and 67%, respectively, significantly enhancing hazard identification sensitivity in data-sparse karst regions. This provides a scientific basis for transitioning disaster prevention management from passive sealing to mechanism-driven proactive early warning.