Assessment of Artificial Intelligence–Based Models and Metaheuristic Algorithms in Modeling Evaporation: Fid-Bau Portal

Assessment of Artificial Intelligence–Based Models and Metaheuristic Algorithms in Modeling Evaporation

Zounemat-Kermani, Mohammad / Kisi, Ozgur / Piri, Jamshid / Mahdavi-Meymand, Amin

Evaporation ( $E p$ ) has a vital importance for the management and development of water resources projects. In this study two scenarios are considered in prediction of monthly pan evaporation. The first scenario challenges the ability of three artificial intelligence–based models [neural network autoregressive with exogenous input (NNARX), gene expression programming (GEP), and adaptive neuro-fuzzy inference system (ANFIS)]. The second scenario investigates the capability of five different metaheuristic algorithms [particle swarm optimization (PSO), firefly algorithm (FA), artificial bee colony (ABC), continuous ant colony optimization (CACO), and genetic algorithm (GA)] integrated with the ANFIS model in $E p$ modeling. Meteorological factors (monthly air temperature, solar radiation, relative humidity, and wind speed data) of two stations in Turkey were used as inputs to the models. Various statistic measures [root-mean-square error (RMSE), mean absolute error (MAE), and determination coefficient ( $R^{2}$ )] and diagnostic analysis (Taylor diagram) were deployed to evaluate and compare the performance of the models. The results of the first scenario show that the ANFIS model gave better performance in Gaziantep Station, whereas the NNARX model performed better in estimating $E p$ values in Adiyaman Station. In the second scenario, it was observed that the PSO and GA algorithms performed better in comparison to the other algorithms in Gaziantep and Adiyaman stations, respectively. The non-parametric Kruskal–Wallis test denoted that there is a significant difference (alpha of 0.05) between the observed versus predicted amounts of monthly $E p$ for the NNARX, ANFIS, and GEP. However, there is no sign of significant difference in predicting monthly $E p$ between the applied metaheuristic algorithms.