Cui et al. (2026) Coupled dominant factors analysis, dual attention deep learning, and uncertainty quantification for long-term pan evaporation ensemble prediction in the Wuding River Basin, China

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-11
• Authors: Zhen Cui, Caihong Hu, Gan Miao, Chengshuai Liu, Shentang Dou
• DOI: 10.1016/j.ejrh.2026.103323

Research Groups

• School of Water Conservancy and Transportation, Zhengzhou University, Zhengzhou, China
• Yellow River Institute of Hydraulic Research, Zhengzhou, China
• Yellow River Laboratory, Zhengzhou, China
• Engineering Technology Research Center of Intelligent Water Conservancy of Henan Province, Zhengzhou, China

Short Summary

This study develops a novel framework integrating dominant factors analysis, dual-attention deep learning, and uncertainty quantification to improve long-term pan evaporation (Epan) ensemble prediction in the Wuding River Basin, China. The framework, utilizing a DA-LSTM model and an improved C-Vine Copula-based multi-model processor (CMMCP), significantly enhances Epan prediction accuracy and reliability by reducing uncertainty.

Objective

• To interpret the regional heterogeneity of Epan driving mechanisms and identify dominant influencing factors using path analysis, providing physically meaningful inputs for deep learning models.
• To develop a novel dual-attention Long Short-Term Memory (DA-LSTM) Epan prediction model and evaluate its performance against benchmark models.
• To incorporate Copula functions into the Multi-Model Condition Processor (MMCP) framework, creating a CMMCP method for ensemble Epan prediction, to quantify prediction uncertainty arising from model structural differences.

Study Configuration

• Spatial Scale: Wuding River Basin, a major tributary of the Yellow River in China, with a watershed area of 30,261 km². Data from three meteorological stations: Jingbian (upper reach), Hengshan (middle reach), and Suide (lower reach).
• Temporal Scale: Monthly data spanning 1980–2013 (34 years). The training period was 1980–2004, and the validation period was 2005–2013.

Methodology and Data

• Models used:
  • Path analysis (for dominant factor identification)
  • Dual-Attention Long Short-Term Memory (DA-LSTM) neural network (main prediction model)
  • Long Short-Term Memory (LSTM) neural network (benchmark)
  • Support Vector Regression (SVR) (benchmark)
  • Multiple Linear Regression (LR) (benchmark)
  • Temporal Attention-based LSTM (TA-LSTM) (for comparison)
  • Feature Attention-based LSTM (FA-LSTM) (for comparison)
  • Multi-Model Condition Processor (MMCP) (ensemble prediction, benchmark)
  • C-Vine Copula-based Multi-Model Condition Processor (CMMCP) (improved ensemble prediction)
• Data sources:
  • National Meteorological Science Data Center of China (http://data.cma.cn/)
  • Monthly pan evaporation (Epan) observed by E601 pan.
  • Monthly meteorological variables: mean surface temperature (TEMmean), maximum surface temperature (TEMmax), minimum surface temperature (TEMmin), atmospheric pressure (PRS), mean relative humidity (RHUmean), minimum relative humidity (RHUmin), cumulative precipitation (PRE), mean wind speed (WINmean), minimum wind speed (WINmin), and sunshine duration (SSD).
  • Data quality control included manual verification and interpolation correction for missing data (missing rate < 10% for Epan, < 1% for other variables) and outlier identification using the Pauta Criterion (3σ Criterion).

Main Results

• Regional Heterogeneity of Epan Drivers: Epan influence mechanisms exhibited spatial heterogeneity across the basin. In the upstream region, thermal factors (e.g., mean surface temperature, sunshine duration) primarily drove Epan. In the midstream and downstream regions, Epan was driven by the synergistic effects of both thermal and dynamic factors (e.g., mean surface temperature, mean wind speed).
• Dominant Factors: Path analysis identified mean surface temperature, maximum surface temperature, minimum surface temperature, sunshine duration, precipitation, and mean wind speed as primary positive drivers of Epan. Atmospheric pressure was identified as the main negative factor inhibiting Epan.
• DA-LSTM Performance: The DA-LSTM model achieved superior Epan prediction accuracy (KGE > 0.900) compared to the standard LSTM model and traditional models like SVR and LR. For instance, at Jingbian station, DA-LSTM achieved a KGE of 0.901, a correlation coefficient (R) of 0.962, a relative error (RE) of -1.68%, a mean absolute error (MAE) of 21 mm, and a root-mean-square error (RMSE) of 28 mm during the validation period.
• CMMCP Ensemble Prediction: The CMMCP method, using the P-III distribution as the optimal marginal distribution, significantly reduced prediction interval widths by 12.9–26.7% compared to the traditional MMCP method (e.g., 81 mm vs 93 mm at Jingbian) while maintaining a high coverage rate (CR values close to or exceeding the 90% confidence level). CMMCP also showed lower continuous ranking probability score (CRPS) values (6.2–13.6% reduction), indicating superior overall probabilistic prediction performance and more reliable quantification of prediction uncertainty.

Contributions

• Innovative integration of path analysis, a dual-attention-based LSTM model, and a C-Vine Copula-based Multi-Model Condition Processor (CMMCP) into a comprehensive framework for long-term pan evaporation prediction.
• Provided novel insights into the spatial heterogeneity of dominant driving mechanisms for pan evaporation in arid and semi-arid regions, enhancing the physical interpretability of Epan processes.
• Quantitatively improved uncertainty reduction and prediction skill compared to existing ensemble forecasting approaches and single-model methods, particularly by relaxing the normality assumption in ensemble prediction through the use of C-Vine Copula functions.

Funding

• Science fund for Distinguished Young Scholars of Henan Province (232300421017)
• National Natural Science Foundation of China (52441902)

Citation

@article{Cui2026Coupled,
  author = {Cui, Zhen and Hu, Caihong and Miao, Gan and Liu, Chengshuai and Dou, Shentang},
  title = {Coupled dominant factors analysis, dual attention deep learning, and uncertainty quantification for long-term pan evaporation ensemble prediction in the Wuding River Basin, China},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103323},
  url = {https://doi.org/10.1016/j.ejrh.2026.103323}
}