Integrating numerical models with deep learning techniques for flood risk assessment

Abstract

Floods are among the natural disasters that pose significant threats to both urban and rural infrastructure, as well as the lives and properties of individuals. Streamflow prediction is essential for obtaining hydrological information and is critical for a variety of water resource projects. While precise daily streamflow predictions are indispensable, forecasting streamflow according to the limited data can help reduce computational time and enhance the efficacy of flood early warning systems. The purpose of this research is streamflow forecasting with the Long Short-Term Memory (LSTM) approach for the next 20 years. The peak streamflow extracted from the LSTM model was entered into HEC-RAS software and obtained flood zone maps and hazard maps. Furthermore, the effectiveness of the proposed method was assessed through statistical analysis, including the coefficient of determination (R²), Mean absolute error (MAE), Root mean square error (RMSE), Nash-Sutcliffe efficiency (NSE), Kling-Gupta efficiency (KGE) and Mean bias error (MBE). In addition to the numerical comparison, the models were evaluated. Their performances were evaluated based on graphical plotting, including scatter plot, violin plot, box plot and Taylor diagram. In the chosen model (MD-8), the values RMSE (m³/s), R², MAE, NSE, KGE and MBE are 4.57, 0.98, 2.56, 0.98, 0.94 and 0.17 during the training period, respectively, and 6.40, 0.92, 3.81, 0.89, 0.87 and 0.09 during the testing period, respectively. The simulation was tailored to the daily streamflow series of the Nesa river in Iran, which spans over 40 years. It is evaluated the results of generating flood zone maps using both the 2D HEC-RAS and LSTM models. The water inflow volume into the reservoir was found to be 76.3 million cubic meters, based on the peak streamflow predicted by the LSTM approach. The present model results demonstrate that the volume of water inflow into the reservoir for return periods of 25, 100 and 500 years were calculated as 76.26, 148.73 and 149.22 million cubic meters, respectively. Additionally, the Difference Flood Hazard (DFH) maps are obtained, illustrating the difference in flood hazard under various conditions.