Evaluating the performance of a hybrid machine learning method for modeling evapotranspiration

Evapotranspiration (ET) is a hydrological variable that is very important in hydrology and agriculture. Traditionally, ET estimation process is based on empirical models and conducted over monthly and daily data. Technological developments have increased the collection of climate data related to ET. Commonly, processing data with a finer temporal resolution using classical methods is complex. However, artificial intelligence has become an alternative tool to model ET, when large datasets are available. One of the most used techniques to model ET is neural networks. Models developed using neural networks have shown good performance in ET estimation, mainly when a multilayer perceptron is employed. However, it is recognized that multilayer perceptron architectures could present difficulties related to their learning mechanisms. Thus, this paper presents the implementation of a hybrid machine learning method based on multilayer perceptron and metaheuristic algorithms. The experimental results show that hybridized methods can improve the ET estimation process. In this study, the performance comparison is based on several state-of-the-art metaheuristic methods including the Runge Kutta Optimization Algorithm, the Crow Search Algorithm, the Fuzzy Logic, and the Grey Wolf Optimizer Algorithm.