Performance-Based Comparative Forecasting of Near-Future Evapotranspiration Using Statistical, Machine-Learning and Deep Learning Methods: A Case Study of Lake Burdur, Türkiye

Creative Commons License

Göztaş M., Oruç Ünal N., Yıldız D., Yıldız D.

Atmosphere, cilt.17, sa.7, ss.1-27, 2026 (Hakemli Dergi)

Özet

In this study, daily reference evapotranspiration (ET0) values for the period 2025–2030 for Lake Burdur, located in the Mediterranean climate zone and within the Burdur closed basin, were estimated using nested architecture focused on high accuracy. The ET0 target corresponds to the FAO-56 Penman–Monteith reference evapotranspiration variable provided by the Open-Meteo Historical Weather API, and it is treated throughout as a standardized measure of atmospheric evaporative demand rather than as actual lake-surface evaporation or basin water loss. For this purpose, daily mean air temperature, relative humidity, shortwave surface radiation, and evapotranspiration data for the period 1984–2024 were obtained from the Open-Meteo platform. In the first stage of the study (Model 1), separate SARIMAX (statistical), XGBoost (machine learning), and LSTM (deep learning) models were applied for temperature, relative humidity, and radiation series; the model with the highest validation mean for each variable was selected. Accordingly, LSTM (Mean  = 0.967) was determined to be the most successful model for temperature, SARIMA(X) (Mean  = 0.812) for relative humidity, and XGBoost (Mean  = 0.845) for the radiation variable, which is non-linear, has strong autocorrelation, and exhibits distinct seasonality. In the second stage (Model 2), these best climate predictions were used as independent variables for evapotranspiration, and LSTM provided the highest success for evapotranspiration (Mean  = 0.941). Trend analyses revealed that the increase in temperature and evapotranspiration and the decrease in relative humidity observed in the past period will continue in the near future. The uncertainty analysis conducted using the Monte Carlo/resampling approach on historical data showed that the 95% prediction intervals largely protected the upward trend in evapotranspiration against random fluctuations. These intervals reflect residual-based uncertainty under the fitted model rather than the full predictive uncertainty of future basin evapotranspiration. The findings indicate that designing model selection appropriate to the structure of the variables within a nested prediction framework significantly improves forecast accuracy and can provide a viable decision support input for sustainable water management in Mediterranean basins experiencing water scarcity.