Regulating Irrigation Water Supply in Arid Agricultural Basins Using a Computational Machine-Learning Framework for Forecasting Vapor Pressure Deficit

Accurate forecasts of vapor pressure deficit (VPD) are essential for irrigation water management, especially for water-scarce nations, because they enable precise estimation of crop water demand and support the optimization of irrigation schedules to minimize water loss. In this study, VPD forecasting models were constructed across five stations in Kuwait using three machine-learning algorithms: support vector machine (SVM), Gaussian process regression (GPR), and nonlinear autoregressive networks (NAR). Daily vapor pressure data sets from selected weather stations from 2007 to 2024 were used. The data were divided into training and validation sets using a cross-validation scheme. An initial partial autocorrelation analysis revealed significant autocorrelation in VPD data with an 11-day lag. Visual inspection of data revealed no long-term trend and validated the utility of the autoregressive model. Three statistical metrics were used to evaluate model performance on the validation set: coefficient of determination (R2), mean absolute error (MAE), and root-mean square error (RMSE). The SVM and NAR models outperformed GPR in forecasting VPD across all stations. Both achieved, at Rabyah station, a coefficient of determination (R2) of 0.95, where SVM recorded the lowest MAE of 3.295 hPa, and NAR followed with 3.342 hPa. These models consistently reproduced historical VPD patterns with high accuracy and generalizability. The models efficiently and consistently replicated historical VPD data across all stations. This study endorses the SVM and NAR models for future hydroclimatological research, specifically for modeling and forecasting VPD. The findings of the current study can aid authorities and policymakers in formulating climate adaptation strategies more effectively based on accurate VPD predictions.