Hybrid Deep Learning for Climate-Driven Atmospheric Irrigation Potential Forecasting: A Case Study for Ankara

Accurate forecasting of atmospheric irrigation potential under variable climate conditions is essential for sustainable water management in semi-arid regions. In this study, monthly atmospheric irrigation potential for Ankara, Türkiye was analyzed using a hybrid residual model that combines a Long Short-Term Memory (LSTM) network with an Extreme Gradient Boosting (XGBoost) algorithm. Daily climate data from the National Aeronautics and Space Administration Prediction of Worldwide Energy Resources (NASA POWER) archive for 2001–2024 and reference evapotranspiration (ET₀) values calculated using the Food and Agriculture Organization (FAO) 56 Penman–Monteith equation formed the basis of the predictors. The LSTM model learned the temporal structure of the climatic inputs, and XGBoost was then applied to represent the remaining residual errors by incorporating extended climatic and seasonal features. Testing on the 2022–2024 period, the hybrid model produced a Root Mean Square Error (RMSE) of 24.4 mm, a Mean Absolute Error (MAE) of 20.3 mm and a Coefficient of Determination (R²) of 0.87. Feature-importance results showed that Month Sine and last-step ET₀ had the strongest influence, followed by relative humidity, wind speed and shortwave radiation. Although the accuracy gains over the standalone models were moderate, the hybrid design delivered more stable forecasts and a clearer view of variable contributions, offering practical value for climate-driven irrigation pressure assessment in semi-arid environments.