Document Type : Research Paper

Authors

• Moein Tosan ¹
• Afshin Shayeghi ²
• Javad Teymouri ³
• Aydin Bakhtar ³

¹ Department of Irrigation and Reclamation Engineering, College of Agriculture and Natural Resources, University of Tehran, Karaj, Iran.

² Department of Geography & Environmental Sustainability, University of Oklahoma, Norman, OK, USA

³ Department sof Civil Engineering, University of Texas at Arlington, Arlington, TX, USA

Abstract

Reliable multi-step-ahead forecasting of reference evapotranspiration (ETo) is critical for proactive water resource management, yet understanding the temporal memory of hydrological systems remains a challenge for black-box deep learning models. This study presents a novel, interpretable forecasting framework integrating temporal SHapley additive exPlanations (SHAP) with advanced recurrent neural networks to predict daily ETo up to 7 days in advance across three contrasting climatic zones in Iran; Birjand (arid), Mashhad (semi-arid), and Gorgan (humid). By benchmarking long short-term memory (LSTM), Bidirectional LSTM (BiLSTM), and CNN-LSTM architectures, it is demonstrated that model complexity does not always guarantee superiority; the standard LSTM proved remarkably robust, achieving high short-term accuracy (R2 > 0.93 for 1-day forecast) in arid regions. However, a distinct humid-climate penalty was observed, with forecast accuracy degrading more rapidly in Gorgan due to stochastic cloud dynamics. The application of Temporal SHAP revealed climate-specific memory effects: in arid zones, Wind Speed exhibited a persistent influence extending back 5 days, acting as a long-term driver of evaporative demand, whereas humid regions were governed by short-term radiative pulses. Furthermore, analysis of extreme events and drought propagation showed that while the model successfully captures heatwave-driven peaks, its reliability decreases under severe evaporative stress (standardized ETo Anomaly > 2). Cross-spatial generalization tests confirmed that models trained on arid data transfer effectively to humid regions (R2 = 0.95), but the reverse transfer fails to capture extreme advective forcing. This study provides a transferable, physically interpretable blueprint for developing early warning systems in data-scarce regions.

Keywords

• Multi-step forecasting
• Explainable AI (XAI)
• Temporal SHAP
• Climatic memory