The accuracy of dam deformation prediction is crucial for the structural stability and overall safety of dams. In recent years, optimization algorithms have been widely used to enhance the prediction accuracy of dam models. However, traditional optimization algorithms often get trapped in local optima, which limits the performance of the models. To address this issue, a stochastic search mechanism was introduced into the grey wolf optimizer (GWO), and the GWO was further improved through the Metropolis acceptance criterion to optimize its performance. Subsequently, an advanced dam settlement prediction model was innovatively constructed by integrating empirical mode decomposition (EMD), modified grey wolf optimizer (MGWO), and long short-term memory network (LSTM). The model was validated using actual data from the Wuyi reservoir in Xinjiang, where EMD was applied to process the data, and a detailed analysis of the different variation characteristics of each component was conducted. Then, the hyperparameters of the LSTM were finely tuned using the MGWO to achieve accurate prediction of dam settlement. A comparative analysis of the model before and after EMD decomposition was conducted. The results showed that the proposed EMD-MGWO-LSTM model for dam settlement prediction exhibited significant advantages across four error performance indicators, demonstrating higher fitting accuracy and superior predictive performance. The findings enhance its adaptability, maintaining fast response and accurate predictions in the complex and variable dynamic operating environment of dams. This research provides technical support for dam safety monitoring and early warning, vigorously promoted the development of water conservancy modernization of flood control and disaster mitigation technology and upgrade.