Infiltration damage is one of the main causes of failure in earth-rockfill dams, and accurate prediction of seepage parameters for earth-rockfill dams is vital to ensure the safe operation of dams. This study, therefore, introduces a data-driven-based approach to predict the seepage flow in earth-rockfill dams. Taking the long short-term memory network (LSTM) as a base model, we introduced variational modal decomposition (VMD) to process original seepage data, and employed the sparrow search algorithm (SSA) to improve the LSTM model. The VMD-SSA-LSTM model was first carried out to predict the seepage flow of earth-rockfill dams. Approximately ten years of seepage data from a reservoir in Shenzhen were utilized to train the prediction model, and the performance of the VMD-SSA-LSTM model was evaluated. The LSTM and VMD-LSTM models were employed for comparison to highlight the superiority of the VMD-SSA-LSTM model. The predicted data closely aligns with the measured values demonstrating the effectiveness of the VMD-SSA-LSTM model in predicting seepage flow. This study introduces a new frame for predicting seepage flow in dams, which provides a reference for evaluating seepage flow in dams using machine learning models.