Applicability assessment of advanced time series deep learning models for 6-hour ahead flood forecasting

Haeun Jung; Hyunil Kim; KiJeon Nam; Sangdan Kim

doi:10.3741/JKWRA.2026.59.10.1275

All Issue

2026 Vol.59, Issue 10 Preview Page Next Page

Applicability assessment of advanced time series deep learning models for 6-hour ahead flood forecasting 6시간 선행 홍수 예측을 위한 최신 시계열 딥러닝 모델 적용성 평가

31 October 2026. pp. 1275-1284

PDF XML

Abstract

This study evaluates the performance of three time series deep learning models, Long Short-Term Memory (LSTM), TimesNet, and TimeMixer, for 6-hour ahead water level forecasting. The models are assessed in terms of flood warning level classification performance and peak flood reproduction. In addition, the effects of Optuna-based hyperparameter optimization and the use of radar-based basin-averaged rainfall data are investigated. The results indicate that TimeMixer consistently achieves the most stable and superior forecasting performance across the two study sites. TimesNet shows competitive performance only under certain conditions, while LSTM fails to adequately capture rapid water level changes during rising stages. The impact of hyperparameter optimization also differs across sites and models, with TimeMixer showing the most stable improvements. Furthermore, the application of radar rainfall data improves flood warning level classification performance for TimesNet and TimeMixer, whereas its effect on LSTM remains limited. In terms of peak flood reproduction, TimeMixer demonstrates stable performance in both peak magnitude and timing. TimesNet shows better peak magnitude estimation in some events but exhibits larger delays in peak timing compared to TimeMixer. Overall, TimeMixer shows the most balanced performance in terms of predictive accuracy and reproducibility, and is considered a promising model for potential application in the field of flood forecasting.

Keywords

Flood forecasting

Peak flood event reproduction

Long short-term memory

TimesNet

TimeMixer

본 연구에서는 6시간 선행 수위 예측을 대상으로 시계열 딥러닝 모델인 Long Short-Term Memory (LSTM), TimesNet, TimeMixer의 성능을 비교·분석하였다. 홍수 위험 단계 분류 성능과 첨두 홍수 재현성을 주요 평가 기준으로 설정하였으며, Optuna 기반 하이퍼파라미터 최적화와 레이더 기반 유역 평균 강우 자료의 도입 효과를 함께 분석하였다. 분석 결과, TimeMixer는 두 대상 유역에서 가장 안정적이고 우수한 예측 성능을 나타냈다. TimesNet은 일부 조건에서만 경쟁력 있는 성능을 보였으며, LSTM은 급격한 수위 변화 구간을 충분히 재현하지 못하였다. 하이퍼파라미터 최적화 효과는 지점 및 모델에 따라 상이하게 나타났으며, TimeMixer에서 가장 안정적인 성능 향상이 확인되었다. 또한 레이더 강우 자료는 TimesNet과 TimeMixer의 홍수 위험 단계 분류 성능 향상에 기여한 반면, LSTM에서는 제한적인 효과를 보였다. 첨두 홍수 재현성 분석 결과, TimeMixer는 첨두 발생 시점과 첨두 수위 크기 모두에서 안정적인 재현 성능을 보였으며, TimesNet은 일부 사건에서 첨두 수위 재현은 우수하였으나 시간 지연이 크게 나타났다. 결과적으로 TimeMixer는 예측 정확도와 재현성 측면에서 가장 균형 잡힌 성능을 보였으며, 향후 홍수예보 분야에 활용 가능한 유망한 모델로 판단된다.

키워드

홍수예보

첨두 홍수 재현성

Long Short-Term Memory

TimesNet

TimeMixer

References

Chailert, T., Trigg, M.A., Altahhan, A., and Pournaras, E. (2025). “Flash flood forecasting and the role of catchment response time: Predictions via deep learning models.” International Federation for Information Processing Networking 2025 Conference, Limassol, Cyprus, pp. 19-25.

Chen, D., Zhou, R., Pan, Y., and Liu, F. (2022). “A simple baseline for adversarial domain adaptation-based unsupervised flood forecasting.” arXiv preprint, arXiv:2206.08105.

Daramola, S., Munoz, D.F., Munoz, P., Saksena, S., and Irish, J. (2025). “Predicting the evolution of extreme water levels with long short‐term memory station‐based approximated models and transfer learning techniques.” Water Resources Research, Vol. 61, No. 3, e2024WR039054.

10.1029/2024WR039054

Do, H.S., Kim, J., Cha, E.J., Chang, E.C., Son, S.W., and Lee, G. (2023). “Long-term change of summer mean and extreme precipitations in Korea and East Asia.” International Journal of Climatology, Vol. 43, No. 7, pp. 3476-3492.

10.1002/joc.8039

Han River Flood Control Office (HRFCO) (2010). Research report on expanding detailed indicators for water resource assessment. Publication Registration No. 11-1611492-000060-01, p. 83.

Hochreiter, S., and Schmidhuber, J. (1997). “Long short-term memory.” Neural computation, Vol. 9, No. 8, pp. 1735-1780.

10.1162/neco.1997.9.8.1735

Huang, Y., Zou, R., Wang, Y., Aslam, L., and Dong, R. (2026). “M2FMoE: Multi-Resolution multi-view frequency mixture-of-experts for extreme-adaptive time series forecasting.” arXiv preprint, arXiv:2601.08631.

10.1609/aaai.v40i26.39362

Korea Institute of Civil Engineering and Building Technology (KICT) (2017). Development of flood/snowstorm disaster forecasting and warning platform based on hydrological radar (5th year). Publication Registration No. KICT 2017-066, pp. 39-47.

Korea Ministry of Environment (ME) (2025). 2024 Hydrological survey report. Publication Registration No. 11-1480000-100081-10, pp. 2050, 2361.

Le, X.H., Ho, H.V., Lee, G., and Jung, S. (2019). “Application of long short-term memory (LSTM) neural network for flood forecasting.” Water, Vol. 11, No. 7, 1387.

10.3390/w11071387

Liu, J., Shen, C., O'Donncha, F., Song, Y., Zhi, W., Beck, H.E., Bindas, T., Kraabel, N., and Lawson, K. (2025). “From RNNs to Transformers: Benchmarking deep learning architectures for hydrologic prediction.” Hydrology and Earth System Sciences, Vol. 29, No. 23, pp. 6811-6828.

10.5194/hess-29-6811-2025

Martel, J.L., Arsenault, R., Turcotte, R., Castañeda-Gonzalez, M., Brissette, F., Armstrong, W., Mailhot, E., Pelletier-Dumont, J., Lachance-Cloutier, S., Rondeau-Genesse, G., and Caron, L.P. (2025). “Exploring the ability of LSTM-based hydrological models to simulate streamflow time series for flood frequency analysis.” Hydrology and Earth System Sciences, Vol. 29, No. 19, pp. 4951-4968.

10.5194/hess-29-4951-2025

Ramesh, T.R., Mathias, R., Yadav, L.N., Jha, S.K., Reddy, A.R., and Karthika, V.R. (2025). “AI-powered early warning systems for urban flood management: Integrating climate models and real-time sensor data.” International Journal of Environmental Sciences, Vol. 11, No. 23s, pp. 74-81.

Wang, S., Wu, H., Shi, X., Hu, T., Luo, H., Ma, L., Zhang, J., and Zhou, J. (2024). “Timemixer: Decomposable multiscale mixing for time series forecasting.” International Conference On Learning Representations, Vienna, Austria, Vol. 2024, pp. 38626-38652.

Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., and Long, M. (2022). “Timesnet: Temporal 2d-variation modeling for general time series analysis.” arXiv Preprint, arXiv:2210.02186.

Ye, L., Li, X., Li, J., Zhang, C., and Zhou, H. (2026). “The impact of spatial resolution on hourly flood modeling in large watersheds.” Hydrology and Earth System Sciences, Vol. 30, No. 10, pp. 2995-3018.

10.5194/hess-30-2995-2026

Zhou, Q., Teng, S., Situ, Z., Liao, X., Feng, J., Chen, G., Zhang, J., and Lu, Z. (2023). “A deep-learning-technique-based data-driven model for accurate and rapid flood predictions in temporal and spatial dimensions.” Hydrology and Earth System Sciences, Vol. 27, No. 9, pp. 1791-1808.

10.5194/hess-27-1791-2023

Information

Publisher :KOREA WATER RESOURECES ASSOCIATION
Publisher(Ko) :한국수자원학회
Journal Title :Journal of Korea Water Resources Association
Journal Title(Ko) :한국수자원학회 논문집
Volume : 59
No :10
Pages :1275-1284
Received Date : 2026-07-09
Accepted Date : 2026-08-20
DOI :https://doi.org/10.3741/JKWRA.2026.59.10.1275

Journal of Korea Water Resources Association ISSN:2799-8746(Print) 2799-8754(Online) 한국수자원학회 논문집

All Issue