Coupling hydrological simulation and deep learning for real-time flood defense in mountainous cities: A case study of the June 18, 2023 extreme flood in Guiyang Penshuichi
- 1College of Civil Engineering, Guizhou University, Guiyang, Guizhou, 550025, China
- 2Corresponding author.
Abstract
Study region: This study focuses on Penshuichi Metro Station in Guiyang, southwestern China, where steep slopes and a low-lying urban center promote rapid runoff concentration and water accumulation. Study focus: A framework coupling hydrological simulation and deep learning was developed for real-time inundation forecasting and graded flood defense. The Xinanjiang model generated 5000 rainfall–inundation scenarios incorporating rainfall, terrain, runoff concentration, and drainage variability. A Bidirectional Long Short-Term Memory network with Multi-Head Attention was trained to forecast entrance water depth 10–90 min ahead. The forecasts were combined with inundation duration and pedestrian stability to quantify personnel risk and guide barrier selection. At the 10-min lead time, the model achieved a Nash–Sutcliffe efficiency of 0.991 and a Mean Absolute Error of 0.012 m; its efficiency remained 0.905 at 90 min, outperforming five benchmarks. New hydrological insights for the region: Reconstruction of the June 18 flood revealed an approximately one-hour lag between peak rainfall and peak inundation risk due to rainfall accumulation, slope-driven runoff concentration, and limited drainage capacity. During this delayed response, the personnel risk coefficient continued to rise and ultimately reached 0.943. Scenario analysis further showed that increasing barrier height progressively reduced risk, with a 1.00-m barrier lowering the peak coefficient to 0.150. These findings highlight the value of water-depth forecasting, timely warnings, and early graded-barrier deployment in mountainous cities.
Related
- Ecohydrological coupling and episodic hydrological connectivity in the Edwards Plateau
- Machine Learning-Based Hydrological Drought Prediction Integrating Teleconnections and Hydrological Memory in a Semi-Arid Basin, Algeria
- Projection of hydrological droughts in the upper Hanjiang River under the climate change impact based on stacking machine learning model
- From hydrological laws to catchment models: A process-based model matches deep-learning accuracy on three continents
- Physics‐Informed Fusion Learning for Nowcasting Fault Detection of Hydrological Multi Sensor Network
BibTeX
@article{wang2026coupling,
title = {{Coupling hydrological simulation and deep learning for real-time flood defense in mountainous cities: A case study of the June 18, 2023 extreme flood in Guiyang Penshuichi}},
author = {Jiayang Wang and Liangyu Pan and Xiaoling Tang and Rui Wang and Ning Liu},
journal = {Journal of Hydrology: Regional Studies},
year = {2026},
month = nov,
doi = {10.1016/j.ejrh.2026.104167},
url = {https://doi.org/10.1016/j.ejrh.2026.104167},
note = {CC BY},
}
RIS
TY - JOUR
TI - Coupling hydrological simulation and deep learning for real-time flood defense in mountainous cities: A case study of the June 18, 2023 extreme flood in Guiyang Penshuichi
AU - Jiayang Wang
AU - Liangyu Pan
AU - Xiaoling Tang
AU - Rui Wang
AU - Ning Liu
PY - 2026
DA - 2026/11/02
JO - Journal of Hydrology: Regional Studies
DO - 10.1016/j.ejrh.2026.104167
UR - https://doi.org/10.1016/j.ejrh.2026.104167
AB - Study region: This study focuses on Penshuichi Metro Station in Guiyang, southwestern China, where steep slopes and a low-lying urban center promote rapid runoff concentration and water accumulation. Study focus: A framework coupling hydrological simulation and deep learning was developed for real-time inundation forecasting and graded flood defense. The Xinanjiang model generated 5000 rainfall–inundation scenarios incorporating rainfall, terrain, runoff concentration, and drainage variability. A Bidirectional Long Short-Term Memory network with Multi-Head Attention was trained to forecast entrance water depth 10–90 min ahead. The forecasts were combined with inundation duration and pedestrian stability to quantify personnel risk and guide barrier selection. At the 10-min lead time, the model achieved a Nash–Sutcliffe efficiency of 0.991 and a Mean Absolute Error of 0.012 m; its efficiency remained 0.905 at 90 min, outperforming five benchmarks. New hydrological insights for the region: Reconstruction of the June 18 flood revealed an approximately one-hour lag between peak rainfall and peak inundation risk due to rainfall accumulation, slope-driven runoff concentration, and limited drainage capacity. During this delayed response, the personnel risk coefficient continued to rise and ultimately reached 0.943. Scenario analysis further showed that increasing barrier height progressively reduced risk, with a 1.00-m barrier lowering the peak coefficient to 0.150. These findings highlight the value of water-depth forecasting, timely warnings, and early graded-barrier deployment in mountainous cities.
SN - 2214-5818
ER -