Li et al. (2026) LSTM-based flood-stage forecasting under limited flood-event samples: Effects of forecasting scheme, spatial input configuration, and input-window length in the Pajiang River basin

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-09-17
• Authors: Shanshan Li, Zhaoli Wang, Bensheng Huang, Daoyi Chen, Liangxiong Chen, Jiachao Chen, Lilan Zhang, Haibo Peng
• DOI: 10.1016/j.ejrh.2026.103993

Research Groups

• Guangdong Research Institute of Water Resources and Hydropower
• Institute for Ocean Engineering, Shenzhen International Graduate School, Tsinghua University
• School of Civil Engineering and Transportation, South China University of Technology
• Graduate School of Engineering, Kyoto University

Short Summary

This study evaluates the performance of LSTM-based flood-stage forecasting under limited flood-event samples in the Pajiang River basin. The results show that recent local stage observations provide the most consistent predictive information across flood events and training-data levels.

Objective

• Evaluate the effect of forecasting scheme on downstream flood-stage prediction under limited flood-event samples.
• Investigate how spatial input configuration, input-window length, and training-data availability interact with forecast skill and stability.

Study Configuration

• Spatial Scale: The study focuses on the Pajiang River basin in Guangdong Province, China.
• Temporal Scale: The study uses hourly data from 2012 to 2024, with a focus on flood events.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) networks were used for forecasting.
• Data sources: Hourly stage and discharge measurements from Jiangkouxu, Feilaixia, and Damiaoxia gauge stations.

Main Results

• Multi-output forecasting scheme (M2) generally provided the best balance between accuracy and stability across flood events and training-data levels.
• Recursive forecasting scheme (M1) deteriorated at longer lead times due to error accumulation.
• Input-window length had limited effects for local-input models but greater and less consistent effects when upstream information was included.

Contributions

• This study provides new insights into the effect of forecasting scheme, spatial input configuration, and input-window length on downstream flood-stage prediction under limited flood-event samples.
• The results support multi-output forecasting based on recent local stage observations under event-scarce conditions similar to the Pajiang River basin.

Funding

• This research was funded by [project name], [program name], and [reference code].

Citation

@article{Li2026LSTMbased,
  author = {Li, Shanshan and Wang, Zhaoli and Huang, Bensheng and Chen, Daoyi and Chen, Liangxiong and Chen, Jiachao and Zhang, Lilan and Peng, Haibo},
  title = {LSTM-based flood-stage forecasting under limited flood-event samples: Effects of forecasting scheme, spatial input configuration, and input-window length in the Pajiang River basin},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103993},
  url = {https://doi.org/10.1016/j.ejrh.2026.103993}
}