Author(s): Chongyu Sun; Guanghui Yan; Jiawei Lin; Saiyu Yuan
Linked Author(s): Saiyu Yuan
Keywords: Flood forecasting; Xin’anjiang model; LSTM; Coupled model; Runoff Process Vectorization
Abstract: In coastal areas, the peak height of typhoon-related rainstorms and floods presents significant challenges for traditional hydrological models, particularly due to the short confluence time and the complex dynamics of rising and retreating water. To tackle this issue, this paper proposes a new coupling model, XAJ-LSTM-RPV, which integrates the Xin’anjiang model (XAJ) with Long Short-Term Memory Network (LSTM) and employs Runoff Process Vectorization (RPV). This approach enhances the physical mechanisms of hydrological modeling while numerizes the rise and fall characteristics of the flow process, , reducing training gradient errors in machine learning input-output data and improving flood forecasting accuracy. The XAJ-LSTM-RPV model was tested in the Baixi River Basin of Ningbo City in the Yangtze River Delta, with comparisons made against the XAJ model, LSTM neural network, and XAJ-LSTM model. Results indicate that the XAJ-LSTM-RPV model outperformed the other models, achieving a certainty coefficient of 0.9, corresponding to an A-level forecasting accuracy. Additionally, during the prediction trial, the model maintained a certainty coefficient above 0.75 over a 10-hour forecasting period, achieving B-level accuracy. These findings demonstrate that the XAJ-LSTM-RPV model, combining RPV with hydrological and neural network methods, significantly enhances forecasting accuracy and enables precise multi-step advance predictions, providing a valuable reference for flood forecasting in the basin.
Year: 2024