基于循环神经网络的山前平原型城市河道洪水预报研究

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

山前平原型城市拥有复杂的下垫面条件,其洪水兼具了山洪和城市洪水的特征,为水文模拟和洪水预报增加了难度。以小清河济南市黄台桥水文站以上流域为研究对象,构建了基于循环神经网络变体的洪水预报模型,并评估了模型的预测性能。研究结果表明,所构建的洪水预报模型既适用于对场次洪水的预报,又适用于对长系列洪水过程的连续预测,且能够灵活地输出流量、水位过程,在一定预测步长内拥有较高的预测精度,其中基于双向门控循环单元(Bidirectional Gate Recurrent Unit,BiGRU)网络构建的模型预测性能最佳,且随预测步长的延长,其性能衰减最弱,能够作为山前平原型城市河道洪水预报的新方法新手段。

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.