Flood bend flow prediction in intermittent river reach using a 2D hydraulic model and stacking-ensemble-based LSTM technique

Accurate prediction and analysis of flood bend flow in braided and intermittent river reaches are essential for providing effective early warnings of flood-related damage to embankments. Most previous studies have relied predominantly on hydrological influencing factors such as rainfall to train and construct predictive data-driven (DD) models. However, in the case of braided rivers, which exhibit complex hydrological characteristics, incorporating additional influential feature factors is essential. Therefore, this study proposed a new stacking-ensemble-based long short-term memory (LSTM) model capable of accurately predicting 2D bend flood flows. The novelty of this study is based on the integration of the different DD techniques and the geomorphological behavior in river bend reaches. This study focused on the river bend reach of the Shuiwei Embankment of the Da-An River in Taiwan as a case study. The HEC-RAS 2D model was utilized for the parameter calibration and model verification and further produced the simulated results under different flood return periods. By combining these simulated datasets with four distinct DD models, named LSTM, extreme gradient boosting regression (XGBR), light gradient boosting machine regression (LGBMR), categorical gradient boosting regression (CGBR) and a meta encoder-decoder LSTM model, a stacking-ensemble-based LSTM was trained and established. To evaluate and explore the performance and applicability of the proposed model, this study employed five quantitative indicators to compare its outcomes with those simulated via the HEC model and four individual DD models. The influence of the proposed models with different meta-learners on predictions was also conducted. The results of the assessment indicated that the proposed model yielded superior overall performance with Nash–Sutcliffe efficiency greater than 0.8 for the model training, test and validation. The proposed model achieved the root mean square error reductions of approximately 19.6%, 25.6%, 31.3%, and 32.9% compared with those of LSTM, XGBR, LGBMR, and CGBR, respectively. The findings of this study provide valuable insights and significant references for the planning of river management strategies, the deployment of river flood warnings and the safety of river embankments.