Prediction of pan evaporation across diverse climates and scenarios using temporal attention clockwork recurrent neural networks coupled with long short-term memory

Accurate prediction of evaporation is crucial for effective water resource management, particularly in regions facing water scarcity. This study investigates evaporation dynamics at two distinct locations with different climates in Iran (Minab and Ramsar stations) using machine learning methods, including simple Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), ClockWork Recurrent Neural Network (CWRNN), Hybrid temporal attention RNN-LSTM, and CWRNN-LSTM models under 11 different scenarios. Our key findings include: (1) the temporal attention CWRNN-LSTM model achieved R² values of 0.982 at Minab and 0.985 at Ramsar, indicating a strong correlation between predicted and observed evaporations; (2) the model produced low Root Mean Square Error (RMSE) values of 0.412 and 0.255, respectively, reflecting its high accuracy; and (3) compared to conventional standalone models, the hybrid models improvemed R² by up to 32.4% and reduced RMSE by 65.7%. By capturing underlying trends and variations in evaporation dynamics, the temporal attention CWRNN-LSTM model could serve as a robust tool for improving water resource management strategies.