Study of flood forecasting based on recurrent neural network for urban river in the piedmont plain

The floods in the piedmont plain city with complex underlying surface conditions exhibits characteristics of both mountain and urban floods, posing challenges for hydrological simulation and flood forecasting. In this study, we developed several flood forecasting models based on recurrent neural network variations in the Xiaoqing River Watershed above the Huangtaiqiao Hydrological Station in Jinan City, and assessed its predictive performance. The research findings demonstrate that the constructed flood forecasting model is suitable for forecasting both single flood events and providing continuous predictions for long series of processes. It has the capability to flexibly generate discharge and water level processes, while maintaining a high level of prediction accuracy within a specific forecast step. Among them, the model based on BiGRU(Bidirectional Gate Recurrent Unit) network exhibits the best prediction performance, with the weakest performance degradation as the length of the prediction step increases. Therefore, it can be regarded as a novel approach for riverine flood forecasting in piedmont plain cities.