Accurate prediction of the water level can guide urban flood control and calamity reduction, as well as water conservancy construction to improve the speed of urban flood emergency response. Data-driven water level prediction models, especially the long short-term memory (LSTM) models, have shown advantages in simulating the strong nonlinear relationships of hydrological elements in nature and thus are widely used. However, the collection of hydrological data in nature is often accompanied by noise and human interference factors, which affect the prediction performance of the models. To address this problem, this study develops a new prediction model combining singular spectrum analysis (SSA) and LSTM, i.e., the SSA-LSTM model. Specifically, SSA first decomposes the observed time series into periodic, trend, and noise components, and then LSTM is used to train the model on the denoised time series to obtain the final prediction results. In this study, the water levels of Guoyang Sluice in the Guohe River Basin from May 1971 to December 2020 are selected as the data set for experiments: 1) The original time series data of water levels are decomposed into multiple trend and noise components (RC₁–RC₁₂) by SSA, and the components (RC₁–RC₁₀) are selected as the trend term and reconstructed into a new water-level time-series signal. 2) The reconstructed signal is trained and verified by the LSTM model, and the predicted results are compared with those of the LSTM model. 3) To obtain the optimal SSA-LSTM model, this study conducts single-step prediction performance evaluation experiments for different time steps (7, 14, 21, 28, and 35 d). The experimental results reveal that the coefficient of determination R2, root mean square error (RMSE), and mean absolute percentage error (MAPE) of the SSA-LSTM water-level prediction model are better than those of the LSTM model at different time steps. The pre-processing of the water level at the Guoyang Sluice by SSA can effectively improve the prediction effect of LSTM. Compared with the traditional LSTM models, the SSA-LSTM model has the characteristic of high reliability and low errors and is more adaptable in water-level prediction applications, which can provide a better decision basis for the rational scheduling of urban flood control, irrigation, water supply, and other water conservation measures.