Abstract: In order to fully explore the effective information in dam deformation monitoring data and improve monitoring model prediction accuracy, a dam deformation prediction model based on IKOA-optimized SAGRU (IKOA-SAGRU model) was proposed.First, a self-attention mechanism was introduced into the gated recurrent unit (GRU).By calculating the contribution rate of time-dimension features, the model effectively captured key time-series features in the measured data, enhancing its sensitivity to key information.Next, the Kepler optimization algorithm (KOA) was improved through three strategies: chaotic map initialization, Runge-Kutta-based position update, and ESQ reinforcement exploration, to automatically optimize the hyperparameters of the self-attention gated recurrent unit (SAGRU).Case studies showed that the improved Kepler optimization algorithm (IKOA) outperformed the sparrow search algorithm, the gray wolf optimizer, the northern goshawk optimization algorithm, and the original KOA in optimization speed and accuracy.The RMSE of the IKOA-SAGRU model′s prediction accuracy was 48.45%, 54.56%, and 58.14% lower than that of the GRU, LSTM, and XGBoost models, respectively.Notably, at key inflection points and peaks in the measured displacement changes, the optimized model achieved better fitting performance.The results demonstrate that the model can effectively extract time-series features from dam deformation sequences, addressing the problems of limited GRU memory capacity, the slow convergence speed of traditional optimization algorithms and their tendency to fall into local optima, thereby significantly improving prediction accuracy.