A Hybrid Model Combining the Cama-Flood Model and Deep Learning Methods for Streamflow Prediction

Climate warming will accelerate the global hydrological cycle and intensify the risk of extreme precipitation and floods. Accurate and reliable streamflow forecasting is fundamental to flood risk mitigation. In this study, we develop a streamflow prediction model by coupling physics-based models, namely, the variable infiltration capacity (VIC) and catchment-based macroscale floodplain (CaMa-Flood) models, with deep learning methods, i.e., the recurrent neural network (RNN) and long short-term memory (LSTM), which complement physics-based models. Two hybrid models, namely, the VIC-CaMa-Flood-RNN (VCR) and VIC-CaMa-Flood-LSTM (VCL) models, are established that provide the advantages of both physics-based and data-driven models. The results show that (1) the VCL model achieves the best performance among the proposed models in streamflow and flood prediction. It outperforms the VCR model, with a potential increase of up to 4.94% in Nash Sutcliffe efficiency coefficient (NSE) and 1.18% in correlation coefficient (R), as well as an improvement of 15.8% in the maximum flood volumes (MAX). (2) in this study, we investigate the actual contribution of various input features (precipitation, maximum temperature, minimum temperature, and wind speed) to the hybrid model-simulated streamflow. The results show that the minimum temperature is the most significant feature, followed by precipitation, maximum temperature, and wind speed. When the maximum and minimum temperatures are considered as temperature features, temperature and precipitation are the most important features affecting the hybrid model-simulated streamflow, with the actual contribution exceeding 80%. (3) during the 2040 and 2090 s, considering the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios, the monthly average streamflow will increase with increasing temperature, and flood seasons will be prolonged. This study is a novel attempt to couple physics-based and data-driven models, which can further improve the streamflow and flood prediction accuracy and provide reliable support for future flood risk assessments.