Comparative analysis of long short-term memory and storage function model for flood water level forecasting of Bokha stream in NamHan River, Korea

• Published in: Journal of hydrology (Amsterdam) 2022-03, Vol.606, p.127415, Article 127415
• Main Authors: Kim, Donghyun, Lee, Joonseok, Kim, Jongsung, Lee, Myungjin, Wang, Wonjoon, Kim, Hung Soo
• Format: Article
• Language: English
• Subjects: Flood water level forecasting; Gradient boosting (GB); Long short-term memory (LSTM); Storage function model; Support vector model (SVM)
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2021.127415

In this study, the applicability of machine learning models was investigated for real-time flood forecasting of a small river basin with a short time of concentration and the modes were compared with the storage function model (SFM), a rainfall-runoff model. Bokha stream basin located in the NamHan river, Korea was selected as the study area. Flood water level forecasting was performed for Heungcheon bridge station which is located in the downstream of Bokha stream using hydrological data observed at Bokha bridge station located in the upstream of the stream. For each of the upstream and downstream basins, Rainfall, water level, and discharge data from 2005 to 2020 were collected at two stations and especially the collected rainfall data were classified into 53 rainfall events using Interevent Time Definition (IETD) analysis of rainfall data. In addition, flood water level forecasting at downstream point was performed using machine learning models such as GB (Gradient Boosting), SVM (Support Vector Model), and LSTM (Long Short-Term Memory). Also the SFM which is a rainfall-runoff model was used for the forecasting. For the application of machine learning models, 33 rainfall events were used for learning and 23 rainfall events were used for evaluation. With the SFM, the flood discharge was forecasted first and then the flood water level was forecasted through the rating curve. The flood water level forecasted by each model was compared with the observed flood water level and predictive power for each model was evaluated by calculating NRMSE (Normalized Root Mean Squared Error). The NRMSEs for the models were ranged from 0.18 to 0.27, and the predictive power was good in the order of the LSTM model at 0.18 followed by the SFM at 0.21. Therefore, the LSTM model showed the best predictive power and was selected as the optimal model for real-time flood water level forecasting in this study. However, the SFM is currently employed in Korea for flood forecasting and warning, and the model well incorporates the basin characteristics, showing relatively good predictive power. Based on the models presented in this study, an optimal model suitable for real-time flood water level forecasting can be selected for flood forecasting and warning points of a small river basin, and it is expected that the forecasting results can be used as base data for decision making.