Han et al. (2025) Baseflow Separation for Improving Dam Inflow Prediction Using Data-Driven Models: A Case Study of Four Dams in South Korea

Identification

• Journal: Water Resources Management
• Year: 2025
• Date: 2025-10-15
• Authors: Heechan Han, Heeseung Park, Donghyun Kim
• DOI: 10.1007/s11269-025-04286-4

Research Groups

• Department of Civil Engineering, Chosun University, Gwangju, South Korea
• Institute of Water Resource System, Inha University, Incheon, South Korea

Short Summary

This study developed and evaluated data-driven models (Deep Neural Network and Random Forest) coupled with a baseflow separation process to improve dam inflow prediction accuracy in four South Korean dams, demonstrating that baseflow separation significantly enhances predictive performance.

Objective

• To evaluate whether coupling a baseflow separation process with data-driven models (Deep Neural Network and Random Forest) improves the accuracy of dam inflow prediction.

Study Configuration

• Spatial Scale: Four major dam watersheds in South Korea (Soyang River, Andong, Imha, and Chungju). Watershed areas range from 1627.4 square kilometers to 2482.9 square kilometers.
• Temporal Scale: Daily data from January 2000 to December 2020 (21 years). Training data covered 15 years (January 2000 to September 2014), and testing data covered the remaining 30% (2015 to 2020). Predictions were evaluated for lead times ranging from 1 to 7 hours.

Methodology and Data

• Models used:
  • Data-driven algorithms: Deep Neural Network (DNN) and Random Forest (RF).
  • Baseflow separation: One-parameter digital filter method (Lyne and Hollick 1979; Nathan and McMahon 1990; Arnold and Allen 1999) with a filtering parameter (α) of 0.925.
  • Software: Python 3.6.13 with tensorflow-1.14.0 and keras-2.3.1 libraries. A web-based hydrograph analysis tool (WHAT) was used for baseflow calculation.
• Data sources:
  • Daily dam inflow data: Water Resources Management Information System (WAMIS).
  • Daily meteorological data (precipitation, maximum/minimum air temperature, humidity): Weather Data Service Portal of the Korea Meteorological Administration (KMA), collected from eight weather stations within each watershed.

Main Results

• The model incorporating baseflow and direct flow as separate inputs (Model 1) consistently outperformed models that did not include baseflow separation (Model 4) for all four dams.
• Compared to Model 4, Model 1 showed significant performance improvements:
  • For RF: Root Mean Squared Error (RMSE) improved by 2–28%, Correlation Coefficient (CC) by 4–43%, and Kling-Gupta Efficiency (KGE) by 22–63%.
  • For DNN: RMSE improved by 2–18%, CC by 5–44%, and KGE by 17–64%.
• The Random Forest (RF) algorithm generally exhibited better predictive performance than the Deep Neural Network (DNN) algorithm across most lead times.
• Optimal lead times for prediction varied by dam and algorithm (e.g., 3 hours for Andong, Imha, and Soyang River dams for both algorithms; 2 hours for RF and 5 hours for DNN for Chungju Dam).
• During the test period (2015-2020), both algorithms achieved CC and KGE values of approximately 0.6 or higher. RMSE values ranged from 0.51 to 0.55 meters (Andong), 1.54–1.57 meters (Soyang River), 0.52–0.53 meters (Imha), and 2.16–2.40 meters (Chungju).
• While both algorithms accurately simulated the temporal patterns and low flow rates, they consistently underestimated peak flow values, indicating a limitation in predicting high-flow hydrological events.

Contributions

• Demonstrated that integrating a baseflow separation process significantly improves the accuracy of dam inflow prediction using data-driven models (DNN and RF).
• Provided a comprehensive case study across four diverse dam watersheds in South Korea, validating the applicability and effectiveness of the proposed methodology in a real-world context.
• Highlighted the importance of optimizing lead times for input data, showing that optimal lead times are specific to both the algorithm and the geographical characteristics of the station.
• Positioned data-driven algorithms as viable and effective alternatives to complex physical models for dam inflow prediction, offering a simpler yet accurate approach for water resource management.

Funding

• This study was supported by a research fund from Chosun University (2023).

Citation

@article{Han2025Baseflow,
  author = {Han, Heechan and Park, Heeseung and Kim, Donghyun},
  title = {Baseflow Separation for Improving Dam Inflow Prediction Using Data-Driven Models: A Case Study of Four Dams in South Korea},
  journal = {Water Resources Management},
  year = {2025},
  doi = {10.1007/s11269-025-04286-4},
  url = {https://doi.org/10.1007/s11269-025-04286-4}
}