Accurate modeling of evaporation is essential for sustainable water resource management, especially in arid and semi-arid regions. This study proposes a hybrid deep learning framework for simulating monthly evaporation at the Sidi-M’Hamed Ben Aouda (SMBA) reservoir in Algeria using long-term hydroclimatic data (1978–2023). Six modeling scenarios were developed based on Long Short-Term Memory (LSTM) networks—LSTM, RF-LSTM, WTC-LSTM, EMD-LSTM, EMD-RF-LSTM, and EMD-WTC-LSTM—to evaluate the effects of feature selection and signal decomposition strategies on predictive performance. Among all configurations, the Random Forest–based model (RF-LSTM) achieved the highest predictive accuracy (NSE ≈ 0.77), followed by the wavelet-based WTC-LSTM (NSE ≈ 0.75). In contrast, classical EMD-based hybrids exhibited lower stability and weaker performance. These findings suggest that robust feature engineering, achieved through Random Forest or Wavelet Transform Coherence, plays a more significant role in enhancing predictive accuracy than increasing structural complexity. The proposed RF-enhanced deep learning framework, therefore, offers a reliable and interpretable approach for monthly evaporation forecasting in arid environments. • A deep learning framework for simulating evaporation at the SMBA basin is proposed • Long-term hydroclimatic data spanning 1978–2023 were used • Six modeling scenarios, all based on LSTM neural networks, were developed. • The models were trained on 80% of the data and tested on the remaining 20% • Among all configurations, the RF-LSTM model delivered the best results
Achite et al. (Fri,) studied this question.