An evolutionary deep learning model integrated with signal reconstruction for long-term hourly forecast of flood in the upper basin of the Gangjiang River
Liu, Zhenxin ; Wen, Tianfu LU ; Zhou, Jianxu and Zhang, Linus LU
(2026)
In Journal of Hydroinformatics
p.2026132-2026132
- Abstract
- Deep learning has advanced flood forecasting through its strong nonlinear modeling capabilities. However, accuracy declines and robustness remain insufficient at longer forecast lead times. Therefore, an evolutionary deep learning model for flood forecasting was developed by integrating time-varying filter-based empirical mode decomposition (TVFEMD), sample entropy-based signal reconstruction (SE), and a CNN-BiLSTM network in this study. Meanwhile, the model's hyperparameters were optimized using a modified whale optimization algorithm (MWOA). Taking the Zhangshui Basin (upper Ganjiang River, China) as a case study, performance and robustness were evaluated and compared between the established model, namely TVFEMD-SE-CNN-MWOA-BiLSTM, with... (More)
- Deep learning has advanced flood forecasting through its strong nonlinear modeling capabilities. However, accuracy declines and robustness remain insufficient at longer forecast lead times. Therefore, an evolutionary deep learning model for flood forecasting was developed by integrating time-varying filter-based empirical mode decomposition (TVFEMD), sample entropy-based signal reconstruction (SE), and a CNN-BiLSTM network in this study. Meanwhile, the model's hyperparameters were optimized using a modified whale optimization algorithm (MWOA). Taking the Zhangshui Basin (upper Ganjiang River, China) as a case study, performance and robustness were evaluated and compared between the established model, namely TVFEMD-SE-CNN-MWOA-BiLSTM, with six other benchmark models from both overall flood forecasting and specific flood event prediction. For 6-, 9-, and 12-h lead times, the TVFEMD-SE-CNN-MWOA-BiLSTM model consistently delivered the best forecast performance. Compared with the results of the CNN-MWOA-BiLSTM model, it showed average improvements of 0.04 in NSE, 0.037 in KGE, and reductions of 33.82 m3/s in RMSE and 7.62 m3/s in MAE. Compared with results of the BiLSTM model, improvements were even greater: +0.147 (NSE), +0.153 (KGE), −72.22 m3/s (RMSE), and −25.42 m3/s (MAE). The model significantly enhances flood forecasting accuracy and robustness across lead times, providing a more reliable foundation for river basin flood control and risk management. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/6c1e4c56-5998-448b-81a4-3dd79fa49387
- author
- Liu, Zhenxin
; Wen, Tianfu
LU
; Zhou, Jianxu
and Zhang, Linus
LU
- organization
- publishing date
- 2026-02-01
- type
- Contribution to journal
- publication status
- epub
- subject
- in
- Journal of Hydroinformatics
- pages
- 20 pages
- publisher
- IWA Publishing
- ISSN
- 1464-7141
- DOI
- 10.2166/hydro.2026.132
- language
- English
- LU publication?
- yes
- id
- 6c1e4c56-5998-448b-81a4-3dd79fa49387
- date added to LUP
- 2026-03-09 10:27:03
- date last changed
- 2026-03-19 02:52:13
@article{6c1e4c56-5998-448b-81a4-3dd79fa49387,
abstract = {{Deep learning has advanced flood forecasting through its strong nonlinear modeling capabilities. However, accuracy declines and robustness remain insufficient at longer forecast lead times. Therefore, an evolutionary deep learning model for flood forecasting was developed by integrating time-varying filter-based empirical mode decomposition (TVFEMD), sample entropy-based signal reconstruction (SE), and a CNN-BiLSTM network in this study. Meanwhile, the model's hyperparameters were optimized using a modified whale optimization algorithm (MWOA). Taking the Zhangshui Basin (upper Ganjiang River, China) as a case study, performance and robustness were evaluated and compared between the established model, namely TVFEMD-SE-CNN-MWOA-BiLSTM, with six other benchmark models from both overall flood forecasting and specific flood event prediction. For 6-, 9-, and 12-h lead times, the TVFEMD-SE-CNN-MWOA-BiLSTM model consistently delivered the best forecast performance. Compared with the results of the CNN-MWOA-BiLSTM model, it showed average improvements of 0.04 in NSE, 0.037 in KGE, and reductions of 33.82 m3/s in RMSE and 7.62 m3/s in MAE. Compared with results of the BiLSTM model, improvements were even greater: +0.147 (NSE), +0.153 (KGE), −72.22 m3/s (RMSE), and −25.42 m3/s (MAE). The model significantly enhances flood forecasting accuracy and robustness across lead times, providing a more reliable foundation for river basin flood control and risk management.}},
author = {{Liu, Zhenxin and Wen, Tianfu and Zhou, Jianxu and Zhang, Linus}},
issn = {{1464-7141}},
language = {{eng}},
month = {{02}},
pages = {{2026132--2026132}},
publisher = {{IWA Publishing}},
series = {{Journal of Hydroinformatics}},
title = {{An evolutionary deep learning model integrated with signal reconstruction for long-term hourly forecast of flood in the upper basin of the Gangjiang River}},
url = {{http://dx.doi.org/10.2166/hydro.2026.132}},
doi = {{10.2166/hydro.2026.132}},
year = {{2026}},
}