Liu et al. (2026) Global patterns of reservoir normal water storage capacities estimated by statistical and machine learning models

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-10-19
• Authors: Ping Liu, Li Tang, XiaoHui Sun
• DOI: 10.1016/j.jhydrol.2026.136648

Research Groups

• Taiyuan University of Technology, Taiyuan, China
• Wuhan University of Science and Technology, Wuhan, China
• Tsinghua University, Beijing, China

Short Summary

This study addresses a critical data gap by estimating Normal Water Storage (NWS) for 7250 global reservoirs using an ensemble of statistical and machine learning models. The Gradient Boosted Decision Trees model performed best, revealing a global mean NWS Ratio of 72% and significant spatial variations crucial for water resource management and hydrological modeling.

Objective

• To estimate Normal Water Storage (NWS) for 7250 global GRanD reservoirs and analyze its spatial patterns across individual, basin, and national scales.

Study Configuration

• Spatial Scale: Global, covering 7250 reservoirs from the GRanD database. Analysis performed at individual reservoir, 164 major river basin, and 148 national scales. Training data utilized 801 reservoirs from the USA, China, Japan, Spain, India, and South Africa.
• Temporal Scale: Satellite-based water area time series (GRSAD) spanning 36 years (1984–2020) were used for the percentile threshold model.

Methodology and Data

• Models used:
  • Linear Regression (LR) models (unified and categorized by reservoir size, climate, purpose, and regulation type).
  • Satellite-based Percentile Threshold model (categorized by storage capacity, climate, purpose, and regulation type).
  • Machine Learning (ML) models: Deep Neural Networks (DNN), Extreme Gradient Boosted Trees (XGBoost), Random Forests (RF), and Gradient Boosted Decision Trees (GBDT).
• Data sources:
  • Global Reservoir and Dam database (GRanD) (7250 reservoirs with attributes including total storage capacity, upstream catchment area, surface water area, residence time, and purpose).
  • Reported Normal Water Storage (NWS) data for 801 reservoirs (collected from national and regional inventories).
  • Global Reservoir Dataset derived through Deep Learning (GRDL) (area-storage curves).
  • Global Reservoir Surface Area Dataset (GRSAD) version 3 (monthly water area time series, 1984–2020, derived from Landsat imagery).
  • Koppen-Geiger climate classification map (for climate categories).
  • Annual flow at dam site (from GRanD database).

Main Results

• The Gradient Boosted Decision Trees (GBDT) model, utilizing seven input features (GBDT-7), was identified as the best-performing model.
• GBDT-7 achieved a coefficient of determination (R2) of 0.95, a Pearson's correlation coefficient (CC) of 0.98, and a mean absolute percentage error (MAPE) of 12%.
• The estimated total global Normal Water Storage (NWS) capacity for the 7250 GRanD reservoirs is approximately 4660 cubic kilometers (km³).
• The global mean Normal Water Storage Ratio (NWSR), defined as the ratio of NWS to total storage capacity, is 72%, with an ensemble uncertainty interval ranging from 61.56% to 73.98%.
• Significant spatial heterogeneity in NWSR was observed:
  • Medium-sized reservoirs exhibit higher NWSR (median 0.76) compared to smaller and larger reservoirs.
  • Flood control reservoirs generally have lower NWSR (median 0.73).
  • Reservoirs in arid regions tend to have lower NWSR (median 0.72) than those in tropical regions (median 0.74).
• At the national scale, the USA, China (551 km³ NWS), Canada, Russia, and Brazil account for substantial portions of the global NWS capacity. China and Brazil dedicate over 80% of their capacities to NWS.
• At the basin scale, the Parana, Yangtze, and Mississippi basins allocate a substantial proportion (82% to 88%) of their storage capacity for NWS, while the Nile basin exhibits a lower NWSR of 0.2.

Contributions

• Developed the first comprehensive global dataset of Normal Water Storage (NWS) for 7250 GRanD reservoirs, bridging a critical data gap in global reservoir capacity information.
• Demonstrated the superior performance of machine learning models, particularly Gradient Boosted Decision Trees, in accurately estimating NWS by capturing complex, non-linear relationships between reservoir characteristics and operational storage.
• Revealed detailed global spatial patterns and heterogeneity of the NWS ratio (NWSR) at individual, basin, and national scales, providing novel empirical insights into reservoir operations and water management practices worldwide.
• Provided a vital empirical input for water resource management, large-scale hydrological modeling, and climate change adaptation, enabling more realistic simulations of human water regulation and downstream discharge dynamics in prominent models such as CaMa-Flood, LISFLOOD, and SWAT.
• Made the developed dataset and machine learning reference code publicly available, serving as a valuable technical reference and baseline for future research and decision-making in water resource management.

Funding

• The Key Research and Development Program of Shanxi Province (Grant No. 202202020101007).

Citation

@article{Liu2026Global,
  author = {Liu, Ping and Tang, Li and Sun, XiaoHui},
  title = {Global patterns of reservoir normal water storage capacities estimated by statistical and machine learning models},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136648},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136648}
}