Hybrid-input ensemble neural network for improving flood-forecasting performance

Abstract

Owing to the increasing severity of flood disasters in recent years caused by climate change, flood forecasting technologies are receiving increasing attention. For real-time river-stage forecasting, hybrid methods that combine the rainfall-runoff-inundation (RRI) model and machine-learning techniques have been actively researched. This study proposes a method for appropriately selecting calculated flow-rate input points and an effective deep neural network (DNN) structure for a hybrid method that employs a DNN to forecast river water levels using flow rates calculated by an RRI model and the observed river water levels. The results show that upstream flow rates, including that of each tributary, significantly contribute toward more accurate forecasts. Moreover, a hybrid-input ensemble neural network, which combines rainfall- and flow-rate-based DNN forecasts, improves the forecast accuracy by approximately 20% for a forecast lead time of up to 24 h compared with a simple DNN model. The proposed hybrid model demonstrated better forecasting accuracy than the simple DNN, even after training on a small amount of flood data, indicating its potential for application in cases with limited past observational data.