Improving the performance of daily pan evaporation (Evp) prediction using the ensemble empirical mode decomposition combined with deep learning models

This study presents a novel approach for the prediction of daily pan evaporation (Evp) based on identifying an optimal combination of model inputs. The Kardeh Dam catchment area in northeastern Iran, where significant evaporation occurs, was investigated in this study. Initially, an appropriate combination of inputs was identified through the gamma test and genetic algorithm (GTGA). Using the ensemble empirical mode decomposition (EEMD), each input data was converted into intrinsic mode functions (IMFs), which were then used as input to long short-term memory (LSTM) and convolutional neural network (CNN) models. The model inputs considered in this work for predicting pan evaporation included maximum and minimum temperature, precipitation, and evaporation in earlier stages. Each input variable was transformed into nine IMFs, simplifying the complex pattern of input variables and leading to improved performance of both CNN and LSTM models. The prediction results using the CNN exhibited RMSE, MAE, and SI values equal to 0.33 mm, 0.24 mm, and 0.06, respectively. Using the LSTM, these values were equal to 0.043 mm, 0.11 mm, and 0.016, respectively. The proposed methodology can be applied in various regions for improved evaporation prediction, offering decision-makers and researchers a clearer understanding of future evaporation trends to effectively manage water resources and prevent wastage of water, particularly in arid and semi-arid regions.