Study region: The Yazd and Ramsar stations are located in hybrid arid and humid climates in Iran, respectively. Study focus: This research study develops a complementary expert system for accurately fore- casting reference evapotranspiration (ET0) over one, three, and seven-day horizons by integrating Machine Learning (ML) models with a novel multivariate decomposition technique. Initially, significant input predictor lags were established through cross-correlation, and the War Strategy Optimization (WSO) algorithm was used for optimal Feature Selection (FS) and determining Multivariate Variational Mode Decomposition (OMVMD) parameters. Each predictor was decomposed into Intrinsic Mode Functions (IMFs) to enhance the temporal characteristics of the data. The study introduced the FS-OMVMD-RF, FS-OMVMD-KNNR, and FS-OMVMD-ETE models, utilizing Random Forest (RF), K-Nearest Neighbors Regressor (KNNR), and Extra-Trees Ensemble (ETE) techniques for daily ET0 estimation. These hybrid models were benchmarked against al- ternatives combining Time-Varying Filter-based Empirical Mode Decomposition (TVF-EMD) with individual models. New hydrological insights for the region: Results demonstrated that the developed models signifi- cantly enhance ET0 forecasting capabilities across multiple time scales. Notably, the FS-OMVMD- ETE model achieved the highest accuracy for ET0 (t + 7) at Yazd and Ramsar stations. The analysis indicated that in a hyper-arid climate, the U2 feature has the greatest impact on fore- casting, while in a humid climate, Tmean is the most influential factor
