Real-time flood forecast using the coupling support vector machine and data assimilation method Xiao-Li Li, Haishen Lü, Robert Horton, Tianqing An and Zhongbo Yu ABSTRACT An accurate and real-time flood forecast is a crucial nonstructural step to flood mitigation. A support vector machine (SVM) is based on the principle of structural risk minimization and has a good generalization capability. The ensemble Kalman filter (EnKF) is a proven method with the capability of handling nonlinearity in a computationally efficient manner. In this paper, a type of SVM model is established to simulate the rainfall–runoff (RR) process. Then, a coupling model of SVM and EnKF (SVM þ EnKF) is used for RR simulation. The impact of the assimilation time scale on the SVM þ EnKF model is also studied. A total of four different combinations of the SVM and EnKF models are studied in the paper. The Xinanjiang RR model is employed to evaluate the SVM and the SVM þ EnKF models. The study area is located in the Luo River Basin, Guangdong Province, China, during a nine-year period from 1994 to 2002. Compared to SVM, the SVM þ EnKF model substantially improves the accuracy of flood prediction, and the Xinanjiang RR model also performs better than the SVM model. The simulated result for the assimilation time scale of 5 days is better than the results for the other cases. Xiao-Li Li Haishen Lü (corresponding author) Tianqing An Zhongbo Yu College of Science, State Key Lab of Hydrology-Water Resources & Hydraulic Engineering, Hohai University, Nanjing 210098, China E-mail: haishenlu@gmail.com Xiao-Li Li College of Electronics and Information, Nanjing University of Technology, Nanjing 210009, China Robert Horton Department of Agronomy, Iowa State University, Ames, IA 50011, USA Key words | ensemble Kalman filter, rainfall–runoff simulation, support vector machine, Xinanjiang rainfall–runoff model INTRODUCTION Accurate real-time flood forecasts are crucial for water resources planning and management, and reservoir and river regulation (Chang & Chen ; Rajurkara et al. ). Many approaches in artificial intelligence have been exploited for hydrological forecasting, such as artificial neural networks (Coulibaly et al. ; Taormina et al. ), genetic algorithm (Cheng et al. ), fuzzy theory (Nayak et al. ) and support vector machine (SVM) (Yu et al. ). Hydrological applications of the SVM have been investigated. Sivapragasam et al. () used the SVM model to perform one-lead-day rainfall–runoff fore- casting. Choy & Chan () determined objectively the framework of the radial basis function network using the SVM, and applied such a network to simulate the relation- ship between rainfall and runoff. Yu et al. () combined chaos theory and the SVM to forecast daily runoff. Bray & Han () developed a runoff prediction method based on the SVM by identifying an appropriate model structure and relevant parameters. In order to achieve an optimal training data set, Sivapragasam & Liong () divided the flow range into three zones, includ- ing high, medium, and low zones, and used different SVM models to forecast daily flows in different zones. Yu et al. () predicted hourly flood stages in real time using the SVM, and performed a sensitivity analysis on lagged input variables of the SVM. Although the SVM model has been broadly used in flood forecasting, the SVM model predictions usually substantially deviate from observations (Sivapragasam et al. ; Sivapra- gasam & Liong ). Sivapragasam & Liong () showed that the error is mainly due to a lack of training data in the model preparation. The data assimilation method can help to reduce prediction error and results in the best state esti- mates (Reichle et al. ). The ensemble Kalman filter 973 © IWA Publishing 2014 Journal of Hydroinformatics | 16.5 | 2014 doi: 10.2166/hydro.2013.075 Downloaded from https://iwaponline.com/jh/article-pdf/16/5/973/387428/973.pdf by guest on 19 December 2018