Calibrated probabilistic ET0 forecasting with lightweight deep learning across northern and southern Xinjiang: seven-station evaluation and irrigation implications

Dachun Chen, Chunfang Yue

Accurate and calibrated reference evapotranspiration (ET0) forecasts are essential for irri-gation scheduling, yet many ET0 studies report deterministic point forecasts and evaluate only a limited number of forecast origins. This study compared deterministic and proba-bilistic ET0 forecasts at seven NCEI stations in northern and southern Xinjiang, including Kashi, Hotan, Korla and Turpan as external stations. Daily records from 1 January 1984 to 16 May 2025 were divided chronologically into 70% training, 10% validation and 20% testing. Four model families were evaluated: DLinear, ENMAMBA, CNNBiLSTM and Crossformer. H7, H30 and H180 forecasts were assessed at all feasible test origins using moving-block bootstrap inference and Holm correction. Rolling conformal calibration generated probabilistic forecasts evaluated by dense-grid CRPS, PICP80, PICP90, MPIW80 and MPIW90. A 2024 drip-irrigated cotton record was used to test how forecast strategies interact with actual irrigation deliveries and generic sandy-loam parameters. DLinear achieved the lowest H30 and H180 RMSE in the northern and external station groups. Overall PICP80/PICP90 was close to nominal, but spring and summer intervals remained under-covered. The irrigation simulation did not reproduce observed root-zone moisture; its decision implications are therefore treated as a bounded sensitivity analysis rather than a field-calibrated recommendation.