Coupled dominant factors analysis, dual attention deep learning, and uncertainty quantification for long-term pan evaporation ensemble prediction in the Wuding River Basin, China

The Wuding River, a major tributary of the Yellow River in China. This study aims to improve the accuracy and reliability of long-term pan evaporation (Epan) prediction in arid and semi-arid regions. A novel prediction framework integrating dominant factors analysis, intelligent modeling, and uncertainty quantification is developed. Path analysis identifies the Epan’s dominant drivers, while a dual-attention long short-term memory (DA-LSTM) Epan prediction model is constructed to capture key feature-temporal dependencies. The DA-LSTM model is compared with support vector regression and linear regression, and integrates into an ensemble scheme. An improved C-Vine Copula-based multi-model processor (CMMCP) is introduced to quantify Epan prediction uncertainty. The framework is evaluated using monthly data from 1980 to 2013 at upstream, midstream, and downstream stations in the Wuding River Basin. The Epan influence mechanisms have regional differences, and thermal factors dominate Epan in the upstream region, while thermal and dynamic factors contribute downstream. Path analysis indicates that surface temperature, sunshine duration, precipitation, and wind speed are key positive drivers, whereas atmospheric pressure exerts a negative influence, all of which exhibit significant total effects. The DA-LSTM achieves superior accuracy (KGE > 0.90), and the CMMCP reduces prediction interval widths by 12.9–26.7% compared to MMCP while maintaining 90% coverage, confirming its enhanced reliability for regional Epan prediction and water resource management. • The dominant influencing factors of the pan evaporation were analyzed. • A novel pan evaporation prediction model based on the dual attention deep learning. • The dual attention deep learning model achieved superior predictive performance. • The improved ensemble prediction method effectively revealed prediction uncertainty.