A Novel Hybrid Model for Hourly Streamflow and Water Level Prediction from Radar Reflectivity Using Deep Learning Approaches

Urban flooding has become a significant challenge for metropolitan areas; therefore, reliable water level and streamflow prediction models are crucial for effective flood control and planning. In this study, we develop a hybrid model, namely SGGP, for hourly water level and streamflow predictions in the Jungrang urban basin along the Han River in South Korea. This model includes two sub-models, the first of which produces three-hour mean areal precipitation (MAP) from quantitative precipitation forecasts (QPFs) based on the Spatial-scale Decomposition Method (SCDM) using Gate Recurrent Units (GRU), namely SCDM-GRU. The second sub-model is utilized to predict hourly-ahead water level and streamflow by integrating a GRU with a particle swarm optimization (PSO) algorithm. Radar data, rainfall, water level, and streamflow data were collected from 2008 to 2022 and are used to establish and evaluate the performance of the model. The SGGP model is evaluated using root mean square error (RMSE), correlation coefficient (CC), mean absolute error (MAE), Nash-Sutcliffe efficiency (NSE), and mean absolute percentage error (MAPE) in comparison with four other deep learning models. The results indicate that the SCDM-GRU achieves a better MAP value, with CC > 0.77 in the first 60 min and CC > 0.6 at the 180-minute time step, compared to the SCDM alone. In streamflow and water level forecasting, SGGP demonstrates superior performance, achieving the highest CC and NSE (> 0.97) and the lowest MAPE (0.66) for 3-hour-ahead predictions. The results show that the proposed SGGP model achieves accurate results in multistep-ahead water level and streamflow predictions.