Abstract
Seepage significantly impacts the stability of earth and rockfill dams, making effective monitoring essential. Traditional Partial Least Squares (PLS) methods handle multicollinearity well but often lack predictive accuracy. Integrating neural networks, particularly Bidirectional Long Short-Term Memory (BiLSTM) networks, enhances accuracy by improving nonlinear data processing and memory of long-term dependencies. This research presents a novel PLS-BO-BiLSTM seepage model for rockfill dams, combining PLS with BiLSTM and Bayesian Optimization (BO). The model employs normal and Rayleigh distribution functions to account for lags in water depth and precipitation, optimized using the Grey Wolf Optimization (GWO) algorithm. Engineering case studies demonstrate the model's high predictive accuracy and generalizability, especially during sudden seepage increases caused by heavy rainfall.