Accurate Medium-Term Forecasting of Farmland Evapotranspiration Using Corrected Next-Generation Numerical Weather Prediction

Accurate medium-term evapotranspiration (ET) forecasting is critical for irrigation scheduling and hydrological assessments. To address biases in numerical weather prediction (NWP) systems, we developed a hybrid GWO_XGB model integrating Extreme Gradient Boosting (XGBoost) with Gray Wolf Optimizer (GWO) for bias correction. Using the corrected data, we evaluate four hybrid models—Support Vector Machine (SVM) and XGBoost, each optimized with either GWO or Grasshopper Optimization Algorithm (GOA)—for 1- to 10-day ET forecasts across 11 farmland stations in Europe and North America (2003–2014). The results showed that the GWO_XGB model demonstrated the best comprehensive performance (average RMSE = 0.476 mm d−1, R2 = 0.829), while the GWO_SVM model performed the weakest (average RMSE = 0.572 mm d−1, R2 = 0.761). Forecast accuracy of Rs and VPD declined with lead time, with the 1-day forecasts being most accurate (RMSE range: 2.005–3.061 MJ mm d−1). Using calibrated NWP data, the highest 1-day forecast accuracy was achieved (average RMSE = 0.715 mm d−1), with GWO_XGB remaining the best (1–3 days average RMSE = 0.667 mm d−1; 10-day cumulative forecast RMSE = 0.698 mm d−1). Overall, the GWO_XGB model combined with NWP calibration provides reliable short- to medium-term ET forecasts for agricultural water management.