Abstract: Groundwater is the sole natural freshwater supply on coral islands. Accurately predicting its level dynamics is of great significance for island water resource management and ecological protection. Currently, long-term in situ monitoring data for coral island groundwater levels remain relatively limited, hydrogeological parameters across different islands are difficult to obtain, and dynamic prediction based on high-frequency monitoring data is particularly lacking. Based on continuous groundwater observation data from a typical coral island in the South China Sea between 1 November 2021 and 10 August 2024, this study analyzed the dynamic patterns of groundwater levels across different time scales. Subsequently, a hybrid prediction model combining wavelet analysis with a Long Short-Term Memory network (Wavelet–LSTM model) was constructed to achieve the prediction of groundwater levels in coral islands. The results indicate that the groundwater level ranged from -0.5 to 2.5 m, with an average level of 0.90 m. It presented significant multi-time-scale fluctuation characteristics, primarily including seasonal cycles, monthly cycles, daily cycles, and high-frequency pulse-like variations. Among these, seasonal and high-frequency variations were mainly controlled by precipitation, while monthly and daily cyclical changes were closely related to tidal dynamics. Model validation demonstrates that the proposed Wavelet–LSTM model achieved high predictive accuracy and showed excellent robustness. Moreover, the model showed strong transferability to independent coral-island settings with similar hydrogeological characteristics, highlighting its potential for regional application. The Wavelet–LSTM model constructed in this study can effectively characterize the multi-time-scale dynamic features of coral island groundwater systems. Its demonstrated generalization capability offers a practical and efficient approach for groundwater assessment and sustainable water-resource management on coral islands where hydrogeological observations are limited.