.png){kind=icon}
返回
Data and knowledge-driven model for flood peak runoff forecasting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jher.2026.100695.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Accurate forecasting of flood runoff peaks during rainstorms remains challenging because prediction errors usually increase near flood thresholds and peak discharge. Most deep learning models learn patterns from data only and do not explicitly emphasize peak-critical errors during training. Therefore, we propose a data and knowledge-driven (DK-TCIT) model that integrates Time-Distributed Convolutional Neural Networks (TD-CNN) for local feature extraction, Informer with ProbSparse attention for global temporal dependencies, and Temporal Convolutional Networks (TCN) for local-global sequence modeling. A key innovation is a knowledge-guided loss function that embeds expert knowledge of flood dynamics, assigning higher learning priority to the critical peakflow region detected from observed flood thresholds. DK-TCIT was evaluated on two basins in China (ChangHua and TunXi) using a 12-hour input window to predict the next 6 h of runoff. Results show that DK-TCIT consistently outperformed ConvLSTM, CNN, SLSTM, TD-CNN-LSTM, STALSTM, Informer, and TCN across all metrics. In TunXi, it achieved RMSE reductions of 31-42% and NSE improvements of 26-41% compared with the best baseline model, while similar gains were obtained in the ChangHua basin. The proposed loss function also surpassed Mean Squared Error (MSE), Mean Absolute Error (MAE), and standard Huber loss, with the largest gains observed around peak runoff conditions. These findings indicate that combining hybrid spatiotemporal learning with explicit peak-focused supervision improves short-term flood peak forecasting and provides a practical solution for flood hazard management applications.
Keyword:
Flood peak forecasting
Spatiotemporal modeling
Hybrid deep learning
Knowledge-driven loss function
Hydrological forecasting
期刊
J
IF:
2.3
论文数:
17
被引数:
0
机构
引用论文
暂无论文信息