Multi-step-ahead forecasting of daily reference evapotranspiration using hybrid deep learning models for the Taklamakan Desert oasis

This research focuses on the Taklamakan Desert oasis in southern Xinjiang, China, which represents one of the most hydrologically challenging and climatically extreme agricultural regions in the world. In this study, we address the short-term forecasting of daily reference evapotranspiration (ET 0 ) using six advanced hybrid deep learning models—LSTM, BiLSTM, CNN-LSTM, CNN-BiLSTM, CNN-LSTM-Attention, and CNN-BiLSTM-Attention. Historical daily meteorological data collected from 25 stations were utilized, with records from 1991 to 2020 used for model training and data from 2021 to 2023 reserved exclusively for independent testing. Forecasts were produced for 1-, 3-, and 5-day lead times. Model performance was evaluated using the Global Performance Index (GPI), and interpretability was further enhanced by applying Shapley Additive Explanations (SHAP) to analyze feature contributions under different climatic conditions. BiLSTM-based models demonstrated higher ET 0 forecasting accuracy than long short-term memory networks (LSTM)-based models, while the incorporation of convolutional neural networks (CNN) and attention mechanisms further improved forecasting performance. The CNN-BiLSTM-Attention model consistently exhibited the highest accuracy and robustness across different stations and forecast horizons, making it the most suitable for operational deployment in desert oasis regions. SHAP analysis indicated that temperature and solar radiation are the principal drivers of ET 0 at most stations, highlighting the spatial heterogeneity in feature importance. Thus, this study provides robust and interpretable ET 0 forecasting models along with new hydrological insights, offering practical support for localized water resource management and enhancing confidence in the deployment of precision irrigation models in hyper-arid agricultural systems. • Developed and evaluated six hybrid deep learning models for ET 0 forecasting at 25 Taklamakan oasis sites. • BiLSTM-based hybrids outperformed LSTM in ET0 forecasting; CNN and Attention further improved accuracy. • The CNN-BiLSTM-Attention model had the best accuracy and was robust for 1-, 3-, and 5-day forecasts at all stations. • SHAP analysis showed temperature and solar radiation are key ET0 drivers, aiding local water management.