Spatiotemporal analysis and machine learning prediction of reference evapotranspiration in Khenchela, Algeria: comparison of MLR, GRNN, and LSTM models

Abstract

Reference evapotranspiration (ET₀) is a key parameter for water management in semi-arid regions with variable climates. This study analyzed the spatiotemporal dynamics of annual ET₀ in the Khenchela region of north-eastern Algeria (2000–2024). ET₀ was computed using the FAO-56 Penman–Monteith (PM) method. Spatial patterns were mapped using Inverse Distance Weighting (IDW). Meteorological data from 16 stations were used to train three models: Multiple Linear Regression (MLR), Generalized Regression Neural Network (GRNN), and Long Short-Term Memory (LSTM) to predict ET₀. The regional mean annual ET₀ increased by 7.2% from 2010 to 2019 decadal average (1 490 mm/year) to the 2020-2024 period (1597 mm/year), contributing to a cumulative 25-year increase of 7% from 2000 to 2009 baseline with hotspots in Babar 2 reaching ~2194 mm/year. The Mann–Kendall test confirmed significant upward trends (p < 0.05) driven by rising temperatures and declining relative humidity. All models performed well (R² > 0.965, RMSE < 0.49 mm/day, RSR < 0.20), with LSTM showing superior accuracy (R² > 0.987, RMSE < 0.232 mm/day, NSE ≈ 0.991, WI > 0.909). The superior performance of LSTM is attributed to its inherent capability to capture temporal autocorrelation and long-term dependencies in climatic time-series data. These findings support adaptive irrigation and drought mitigation in semi-arid regions of northern Africa.

Keywords: Algerian semi-arid region, climate change, evapotranspiration, FAO-56 Penman-Monteith, machine learning, semi-arid regions, water resource management.