Comparative study of data-driven intelligent flood forecasting methods for small- and medium-sized rivers

In recent years, data driven flood forecasting methods have been widely used in flood forecasting, and good results have been achieved. But most data-driven models are applied to large basins, seldom in small basins. Flash floods in small- and medium-sized rivers, which are mostly located in data-poor mountainous areas in China, are featured by abruptness, rapid concentration and short forecasting time. The support-vector-machine (SVM) model, the BP neural network model, the RBF neural network model and extreme learning machine (ELM) model respectively are established and the used to forecast flash floods in Changhua basin. The results show that the SVM model and RBF network model have accurate prediction in the low flow section with simple parameters while BP network has better performance in the high flow section with less stable forecast results for the low flow section, and that the ELM model is not stable with large deviations. As a result, the SVM model or RBF model was adopted for the low flow section, and BP network for the high flow section. This final combination model shows better performance in experiments.