Accurate short-term forecasting of river water levels plays an important role in flood risk management and water resource planning. This study proposes a hybrid forecasting framework that integrates the Light Gradient Boosting Machine (LightGBM) and Long Short-Term Memory (LSTM) networks to improve prediction accuracy for hydrological time series. In the proposed approach, LightGBM is first employed to capture nonlinear relationships between lagged input variables and water levels, while the LSTM model learns the remaining temporal dependencies through residual modeling. To evaluate the effectiveness of the framework, a synthetic river water level dataset was generated to replicate key hydrological characteristics, such as seasonal patterns and stochastic fluctuations. Experimental results show that the proposed LightGBM-LSTM model outperforms the individual LightGBM and LSTM models across multiple evaluation metrics, including MAE, RMSE, NSE, and R². In particular, the hybrid model achieves the lowest prediction errors and the highest goodness-of-fit values, demonstrating improved forecasting acc and stability. These findings suggest that combining tree-based machine learning models with deep recurrent networks provides a promising approach for hydrological time series forecasting.