Hybrid residual modeling of pan evaporation in hyper-arid climates: Benchmarking interpretable neural architectures against physical drivers

Abstract

Study Region Kuwait, Arabian Peninsula, Western Asia This study proposes an interpretable hybrid residual learning framework that integrates stepwise linear regression (SWLR) with four neural network architectures: Feedforward Neural Networks (FNN), Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNN), and Echo State Networks (ESN). The framework is designed to model daily pan evaporation based on routine meteorological inputs. It was applied to two meteorological stations in Kuwait representing distinct arid microclimates. By separating the deterministic linear component from nonlinear residual behavior, the models provide enhanced transparency while improving predictive performance. Among the configurations tested, SWLR–FNN and SWLR–LSTM hybrids demonstrated consistent accuracy ( R ² > 0.98; RMSE < 0.65 mm/day at KIA), stability under synthetic input perturbation (retaining >97 % accuracy under 20 % noise), and effective generalization across spatial domains. In contrast, the CNN and ESN hybrids, while novel in this context, exhibited lower robustness and greater residual dispersion. A comprehensive evaluation encompassing distributional agreement, residual diagnostics, and transferability, highlighted the strengths and limitations of each model in replicating the complex, nonlinear behavior of evaporation under extreme environmental conditions. The proposed framework offers a practical and interpretable approach to evaporation modeling in data-scarce regions. Model robustness and adaptability support its broader application in arid-zone hydrological forecasting, water allocation planning, and drought risk resilience. • Proposed novel SWLR-guided deep models for arid-zone evaporation modeling • Integrated CNN, LSTM, ESN, and FNN with SWLR for residual error learning • Validated cross-station generalizability and robustness of hybrid models • Introduced a benchmarking approach using physical evaporation indicators • Demonstrated SWLR–FNN and SWLR–LSTM superiority for arid-zone evaporation modeling

Topics & Concepts