Pan Evaporation Forecasting Using Empirical and Ensemble Empirical Mode Decomposition (EEMD) Based Data-Driven Models in the Euphrates Sub-Basin, Turkey

Abstract

Forecasting evaporation, an important variable in the hydrological cycle, is crucial for managing water resources and taking precautions against severe phenomena, such as droughts and floods. In this study, the prediction of daily pan evaporation was carried out in the Euphrates sub-basin, Turkey, which has different climate characteristics and is a critical region for Turkey or neighbouring countries. In this regard, two empirical models, namely the Griffith model and calibrated Hargreaves-Samani, and four ensemble empirical mode decomposition (EEMD) based data-driven models, namely EEMD-Random Forests (EEMD-RF), EEMD-Artificial Neural Network (EEMD-ANN), EEMD-Gradient Boosting Machines (EEMD-GBM), and EEMD-Regression Tree (EEMD-RT) were used for evaporation forecasting. The EEMD and Recursive Feature Elimination (RFE) were implemented as a signal decomposition technique and determination of the importance of the EEMD components, respectively. Although the empirical models yielded satisfactory performance, they predicted low and high evaporation values poorly, in general. The EEMD-RF, EEMD-ANN, and EEMD-GBM models performed better than the EEMD-RT model. The data-driven models, except EEMD-RT, outperformed the empirical models, especially regarding predicting extreme evaporation values. The sensitivity analysis indicated that wind speed, humidity, and maximum temperature could influence evaporation forecasting. This study shows that using data-driven models benefitting from EEMD and RFE can be a good alternative to empirical models for predicting evaporation.