Quantifying Irrigation Water Demand through Optimized Daily Vapor Pressure Deficit Forecasting Using LSTM and Metaheuristic Algorithms

This study presents a robust computational framework that integrates long short-term memory (LSTM) neural networks with four advanced metaheuristic optimization algorithms: Nondominated Sorting Genetic Algorithm II (NSGA-II), Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Gray Wolf Optimizer (GWO) for daily forecasting of vapor pressure deficit (VPD) in hyperarid climates. The framework was applied to four meteorological stations in Kuwait to predict atmospheric moisture demand, a key driver of evapotranspiration and irrigation scheduling. Key LSTM hyperparameters were systematically optimized to improve model accuracy, stability, and structural efficiency across distinct climatic settings. Model evaluation employed multiple statistical metrics, including root-mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2), alongside residual autocorrelation analysis and cross-station generalization heatmaps. GWO- and GA-optimized models demonstrated superior performance, achieving the best overall average accuracy and offering strong consistency across error metrics and bias stability. Unlike traditional single-station or static models, this framework supports spatially transferable VPD prediction without retraining, providing a scalable solution for operational water resources planning in data-scarce, arid environments. The results underscore the effectiveness of combining deep sequential learning with nature-inspired optimization for developing generalizable, real-time forecasting systems. By enabling anticipatory irrigation decision-making based on high-resolution VPD forecasts, the proposed approach offers a practical tool for climate-resilient agricultural and water infrastructure management in vulnerable regions.