Yu et al. (2026) Comparison of Machine Learning Models in Reservoir Outflow Simulation Under Different Hydrological Conditions

Identification

• Journal: Lecture notes in civil engineering
• Year: 2026
• Date: 2026-01-01
• Authors: Lei Yu, Jing Zhang, Yanfei Yang, Li Zhang, Yan Zhang, Yu Zhang, Luchen Zhang, Zehui Zhou
• DOI: 10.1007/978-981-95-4889-7_38

Research Groups

• Nanjing Hydraulic Research Institute, Nanjing, China
• Bureau of Hydrology, Changjiang Water Resources Commission, Wuhan, China
• Monitoring Center for Hydrological and Water Resources of Poyang Lake, Nanchang, China
• School of Management, Nanjing University of Posts and Telecommunications, Nanjing, China

Short Summary

This study compares the performance of four machine learning models (Random Forest, Gradient Boosting, Long Short-Term Memory, and Bidirectional LSTM) in reservoir outflow simulation under varying hydrological conditions for the Lianghekou reservoir in China. Random Forest achieved the best overall performance (R² = 0.6940), with model accuracy generally declining from wet to dry years while maintaining consistent model ranking.

Objective

• To systematically compare and analyze the performance of four mainstream machine learning models (Random Forest, Gradient Boosting, Long Short-Term Memory, and Bidirectional LSTM) in reservoir outflow simulation under various hydrological conditions (wet, normal, and dry years).

Study Configuration

• Spatial Scale: Lianghekou reservoir, Yalong River Basin, China. The reservoir has a total storage capacity of 10.15 × 10^9 cubic meters.
• Temporal Scale: Decadal-scale (10-day interval) data spanning 64 years (1958–2021).

Methodology and Data

• Models used: Random Forest (RF), Gradient Boosting (GB), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM).
• Data sources: Decadal-scale outflow, inflow, and reservoir storage data (1958–2021) for the Lianghekou reservoir, collected from the Bureau of Hydrology, Changjiang Water Resources Commission.

Main Results

• Random Forest (RF) achieved the best overall performance with an R² value of 0.6940.
• The overall performance ranking from best to worst was: RF (R² = 0.6940), Gradient Boosting (GB) (R² = 0.6794), Bidirectional LSTM (BiLSTM) (R² = 0.6588), and Long Short-Term Memory (LSTM) (R² = 0.6247).
• Model performance exhibited a clear year-type dependency, with prediction accuracy generally declining from wet to dry years.
• The relative ranking of models remained consistent across different hydrological year types (RF > GB > BiLSTM > LSTM).
• Dry years presented the most significant simulation challenges, with all models showing substantially reduced accuracy (e.g., RF R² dropped to 0.6618 in dry years, LSTM R² to 0.5398).
• In wet years, RF achieved R² = 0.7342 and RMSE = 225.63.

Contributions

• Provided a systematic and comprehensive evaluation of four state-of-the-art machine learning models for reservoir outflow simulation under varying hydrological conditions.
• Demonstrated the year-type dependency of model performance, highlighting the increased challenges in simulating reservoir behavior during dry years.
• Proposed a stratified modeling and validation strategy based on watershed hydrological characteristics.
• Offered valuable insights and practical guidance for tailoring model selection to specific application scenarios and hydrological conditions, aiding in the development of intelligent water resource management systems and optimization of reservoir operation strategies.

Funding

• National Key Research and Development Program of China (Grant No. 2022YFC3002705)
• National Natural Science Foundation of China (Grant No. 52409038)
• Natural Science Foundation of Jiangsu Province, China (Grant No. BK20230121)

Citation

@article{Yu2026Comparison,
  author = {Yu, Lei and Zhang, Jing and Yang, Yanfei and Zhang, Li and Zhang, Yan and Zhang, Yu and Zhang, Luchen and Zhou, Zehui},
  title = {Comparison of Machine Learning Models in Reservoir Outflow Simulation Under Different Hydrological Conditions},
  journal = {Lecture notes in civil engineering},
  year = {2026},
  doi = {10.1007/978-981-95-4889-7_38},
  url = {https://doi.org/10.1007/978-981-95-4889-7_38}
}