Park et al. (2025) AI-Based Time-Series Ensemble Approach Coupled with a Hydrological Model for Reservoir Storage Prediction in Korea

Identification

• Journal: Water
• Year: 2025
• Date: 2025-11-18
• Authors: Jaeseong Park, Jason Sung-uk Joh, Minha Choi, Taejung Kim, Jaeil Cho, Yangwon Lee
• DOI: 10.3390/w17223296

Research Groups

• Major of Geomatics Engineering, Division of Earth Environmental System Sciences, Pukyong National University, Busan, Republic of Korea
• Research Institute for Geomatics, Pukyong National University, Busan, Republic of Korea
• Department of Water Resources, Graduate School of Water Resources, Sungkyunkwan University, Suwon, Republic of Korea
• Department of Geoinformatic Engineering, Inha University, Inchon, Republic of Korea
• Department of Applied Plant Science, Chonnam National University, Gwangju, Republic of Korea

Short Summary

This study developed an AI-based time-series ensemble framework coupled with a hydrological model to accurately predict reservoir storage rates in South Korea, especially for reservoirs lacking inflow/outflow data. The framework achieved high accuracy, with Mean Absolute Errors of 0.820%p, 1.339%p, and 1.766%p for 1, 2, and 3-day ahead predictions, respectively, outperforming individual models.

Objective

• To enhance reservoir storage prediction accuracy for agricultural reservoirs lacking inflow and outflow data by simulating these variables using a rule-based rainfall–runoff hydrological model (3TM) and incorporating them as input features for AI time-series prediction models.
• To predict agricultural reservoir storage rates using the Temporal Fusion Transformer (TFT) model and compare its performance with Recurrent Neural Network (RNN)-based models (Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU)).
• To further improve prediction accuracy by developing a Bayesian Model Averaging (BMA) ensemble model combining LSTM, GRU, and TFT.

Study Configuration

• Spatial Scale: 46 agricultural reservoirs across South Korea, primarily in the southwestern plains, with effective storage capacities ranging from 7,001,000 cubic meters to 107,000,000 cubic meters.
• Temporal Scale: Daily water level and storage rate data from 2013 to 2024 (12-year period) for training, validation, and testing. Predictions were made for 1, 2, and 3 days ahead.

Methodology and Data

• Models used:
  • Hydrological Model: Three-Tank Model (3TM) for simulating inflow and outflow.
  • AI Time-Series Models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Fusion Transformer (TFT).
  • Ensemble Method: Bayesian Model Averaging (BMA).
  • Optimization Algorithms: Non-Dominated Sorting Genetic Algorithm II (NSGA-II) for 3TM calibration, Optuna library (Bayesian optimization) for AI model hyperparameter tuning.
  • Sensitivity Analysis: Generalized Likelihood Uncertainty Estimation (GLUE).
• Data sources:
  • Reservoir Data: Daily water level and storage rate, effective storage capacity, and catchment area for 46 agricultural reservoirs from the Korea Rural Community Corporation (KRC).
  • Meteorological Data: Daily precipitation, evaporation (converted using a pan coefficient of 0.6), mean wind speed, solar radiation, relative humidity, and mean temperature from the Automated Surface Observing System (ASOS) stations operated by the Korea Meteorological Administration (KMA).
  • Derived Data: Simulated daily inflow and outflow from 3TM, and cosine-transformed Julian day to represent temporal periodicity.

Main Results

• The inclusion of 3TM-simulated inflow and outflow variables significantly improved the prediction accuracy of all AI models. For 1-day ahead predictions, Mean Absolute Errors (MAEs) decreased from 1.116%p (LSTM), 1.025%p (GRU), and 1.047%p (TFT) without 3TM to 0.930%p (LSTM), 0.977%p (GRU), and 0.908%p (TFT) with 3TM.
• The Temporal Fusion Transformer (TFT) model consistently outperformed LSTM and GRU in individual model predictions, especially for longer lead times. For 3-day ahead predictions, TFT with 3TM achieved an MAE of 1.805%p, compared to 2.025%p for LSTM and 2.201%p for GRU.
• During flood control periods, incorporating 3TM variables improved 1-day ahead MAEs by approximately 0.3%p (e.g., TFT from 2.346%p to 2.042%p).
• The Bayesian Model Averaging (BMA) ensemble of LSTM, GRU, and TFT models achieved the highest accuracy, with MAEs of 0.820%p (1-day), 1.339%p (2-day), and 1.766%p (3-day) ahead, outperforming the best single model (TFT). The BMA ensemble also provided more stable predictions during flood control periods.
• Feature importance analysis using the TFT model's attention mechanism identified past reservoir storage rate (43.42%), inflow (11.24%), and outflow (10.78%) as the most critical input variables for prediction.

Contributions

• Proposed a novel framework that integrates a rule-based rainfall–runoff hydrological model (3TM) with AI time-series models to accurately predict reservoir storage rates, particularly for reservoirs lacking explicit inflow and outflow data.
• Demonstrated the significant accuracy improvement achieved by simulating and incorporating missing hydrological variables (inflow and outflow) into AI prediction models, especially during periods of rapid water level fluctuations like flood control seasons.
• Validated the superior predictive performance of the Temporal Fusion Transformer (TFT) model compared to traditional Recurrent Neural Network (RNN) architectures (LSTM and GRU) for multi-horizon reservoir storage forecasting.
• Developed and confirmed the effectiveness of a Bayesian Model Averaging (BMA) ensemble approach, which further enhanced prediction accuracy and robustness beyond individual AI models.
• Quantified the importance of various input features, including the newly simulated inflow and outflow, using explainable AI (XAI) techniques, providing insights into the model's decision-making process.
• Offers a practical solution for proactive reservoir operation, flood damage prevention, and efficient irrigation water management in regions characterized by erratic seasonal rainfall and limited hydrological data availability.

Funding

• Korea Agency for Infrastructure Technology Advancement (KAIA) grant funded by the Ministry of Land, Infrastructure and Transport (Grant RS-2022-00155763).
• Intelligent Technology Development Program on Disaster Response and Emergency Management funded by Ministry of Interior and Safety (MOIS, Korea) (Grant 2021-MOIS37-002).

Citation

@article{Park2025AIBased,
  author = {Park, Jaeseong and Joh, Jason Sung-uk and Choi, Minha and Kim, Taejung and Cho, Jaeil and Lee, Yangwon},
  title = {AI-Based Time-Series Ensemble Approach Coupled with a Hydrological Model for Reservoir Storage Prediction in Korea},
  journal = {Water},
  year = {2025},
  doi = {10.3390/w17223296},
  url = {https://doi.org/10.3390/w17223296}
}