comparative analysis of hybrid deep learning models for dam inflow prediction: lstm-gru, cnn-lstm, attention-lstm, and transformer approaches

This study offers the first comprehensive comparison among four hybrid deep learning architectures—lstm-gru, cnn-lstm, attention-lstm, and transformer—for multipurpose dam inflow forecasting under severe hydrological variability. the study employed a 14-year dataset (168 observations, 2010-2023) obtained from jiroft dam in iran and framed with hydrological and operational parameters including precipitation, reservoir capacity, agricultural discharge, and turbine functions. the lstm-gru architecture yielded the best performance by attaining 0.873 r² and 29.73 m³/s root mean square error (rmse) during the validation procedure and demonstrating the best balance among accuracy and generalizability. the model robustness was confirmed by advanced validation methods including taylor diagrams, violin diagrams, and statistical testing (kolmogorov-smirnov, ljung-box, and breusch-pagan tests). seasonal analysis revealed a seven times change in flow rates ranging across winter maxima of 391.5 m³/s and autumn minima of 56.2 m³/s. the models showed a widespread tendency to predict lower peak flows (percentage bias, pbias: -14.34% to -20.86%), suggesting the presence of operational safety buffers. precipitation–agricultural interactions were identified as the key forecasting variable (importance = 0.999). the model provides real-time support for decision-making on reservoir management, flood protection, and potable water supply under changing environmental circumstances and provides a validated model for ai-accelerated water resource management.