[A-TT35] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT35-P01] A Hierarchical Multi-scale LSTM Model for Improved Water Level Forecasting in the Han River Basin

*JONGHO KIM¹, Trung Duc Tran¹ (1.University of Ulsan)

Keywords:Flood forecasting, Long Short-Term Memory, Hierarchical, Multi-scale, Mutual Information, Bayesian Optimization

Accurate water level forecasting is crucial for effective flood management, reservoir operations, and hydrological planning. This study proposes a Hierarchical Multi-scale Long Short-Term Memory (HMLSTM) model to enhance predictive accuracy by capturing both short-term fluctuations and long-term trends in water level data. The model processes raw time series data at three different temporal resolutions: high-frequency (10-minute intervals), medium-frequency (hourly aggregates), and low-frequency (6-hour aggregates). These resolutions are modeled through separate hierarchical LSTM hidden layers, with each layer targeting specific temporal patterns: the first captures short-term fluctuations, the second extracts medium-scale variations, and the third identifies long-term trends. The outputs from these layers are fused through a feature concatenation layer to produce the final prediction. Mutual Information (MI) is used to select the most relevant input predictors, and Bayesian Optimization (BO) is employed to fine-tune the LSTM hyperparameters. The proposed model is validated using 10-minute interval water level data from 30 stations in the Han River basin, Korea, with predictions made for 36 lead times (6 hours ahead). Experimental results show that the HMLSTM model outperforms benchmark models at most stations, with improvements of 5-10% in Kling-Gupta Efficiency (KGE) and 5-20% in Peak Error (PE), especially for longer lead times. This methodology offers a robust LSTM-based framework for water level prediction, with significant potential for applications in early warning systems and hydrological risk assessment.
Acknowledgment: This work was supported by Korea Environment Industry & Technology Institute (KEITI) through Water Management Program for Drought Program funded by Korea Ministry of Environment (2022003610003), and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(RS-2022-NR070280).