Accurate estimation of evapotranspiration (ET) is crucial for e ffective water resources management and agricultural planning, particularly in arid and semi-arid regions. However, existing prediction approaches often su ffer from limited generalization and systematic bias under complex climatic variability. This study proposes a novel iTransformer_ET framework that integrates improved temporal encoding and attention mechanisms to enhance ET time series prediction. The prediction model was optimized by combining a multisource meteorological dataset covering the entire Arid Zone of Northwest China (NWC). Comprehensive evaluation against state-of-the-art statistical, machine learning, and deep learning models demonstrates that iTransformer_ET achieves the highest predictive accuracy, with R-2 = 0.9562, root mean squared error (RMSE) = 0.3701 mm /day, mean absolute percentage error (MAPE) = 24.35%, and an mean directional accuracy (MDA) of 1.0000. Notably, it yields the smallest bias magnitude (PBIAS = -13.07%), indicating minimal systematic deviation, and the highest explained variance score (EVS = 0.9819), confirming strong robustness and stability across varying ET ranges. The proposed approach o ffers a reliable and scalable solution for ET prediction, with potential applications in climate impact assessment, precision irrigation scheduling, and sustainable water resources management.