Water level prediction for the Three Gorges-Gezhouba cascade hydropower stations is crucial for their safe and stable operation and overall benefits. Nevertheless, due to the combined effects of multiple factors, such as the complex transformation between dynamic and static storage-capacity calculations and the unsteady flow downstream of the stations, traditional methods struggle to accurately predict short-term water levels. When the stations perform peak-shaving and frequency-regulation tasks under complex operating conditions, there is a risk of violating scheduling regulations and opening the gates, which may lead to engineering safety hazards and economic losses. In this study, we employed the Long Short-Term Memory (LSTM) deep-learning method to develop an ultra-short-term water-level prediction model for the Three Gorges-Gezhouba Hydropower Stations. We utilized water-level, inflow, and output data to forecast the ultra-short-term water-level processes of the stations. Subsequently, we analyzed the prediction accuracy of the model using data from peak-shaving scenarios. The results show that the model exhibits high overall accuracy, stability, and adaptability, and maintains stable prediction accuracy under different peak-shaving conditions. However, the prediction results tend to be homogenized at extreme water levels. The average error of 24-hour water-level prediction for the upstream of the Three Gorges and Gezhouba is less than 0.05 m. These findings can offer technical support for the refined scheduling of cascade hydropower stations.