Author(s): Song-Yue Yang; Y. S. Gan; Kuo-Wei Wu; Tsung-Tang Tsai
Linked Author(s): Song-Yue Yang
Keywords: SRH-2D LSTM BiLSTM GRU BiGRU Transformer ViT
Abstract: To effectively achieve real-time flood forecasting, many researchers have begun to train artificial intelligence (AI) models using the results of two-dimensional (2D) hydraulic simulations. This approach significantly reduces the computational resources required for forecasting and provides predictions more rapidly than traditional 2D hydraulic models. While previous studies primarily focused on simulating the water depth and flow velocity of 2D fixed-bed hydraulic models, this study explores AI models' application in simulating 2D mobile-bed hydraulic models, particularly forecasting water depth, flow velocity, bed shear stress, and bed erosion. This paper presents a case study in the lower reaches of Taiwan's Jhuoshuei River, proposing a real-time flood forecasting method that integrates the SRH-2D (Sedimentation and River Hydraulics – Two-Dimension) model with AI. During the research, observed flow data were first input into the SRH-2D model to generate the required AI training datasets. These datasets were then used to train, validate, and test various AI models, including LSTM (Long Short-Term Memory), BiLSTM (Bidirectional LSTM), GRU (Gated Recurrent Unit), BiGRU (Bidirectional GRU), Transformer, and Vision Transformer (ViT), to assess their forecasting performance. ViT accurately predicted water depth, flow velocity, and bed shear stress thanks to its self-attention mechanism, effectively capturing global dependencies within spatially continuous hydrological model data. The study found that all AI models achieved the highest accuracy in predicting flow velocity, followed by water depth and then bed shear stress, while the accuracy of predicting erosion depth was the lowest. When the ViT model was used to predict fewer variables, its overall prediction accuracy significantly improved, demonstrating its potential to focus on learning key features. Overall, the AI models showed a significant improvement in prediction efficiency compared to the SRH-2D model, highlighting their tremendous potential in real-time flood forecasting.
Year: 2025