Search

Pan evaporation (E p) is an efficient and practical tool for planning and managing water resources, understanding the water balance in hydrological processes, and developing irrigation systems, particularly in regions with limited water resources. Accurate prediction of E p using easy-to-measure meteorological variables is beneficial for any region. The current study aimed to explore the applicability of Copula-based Bayesian Model Averaging (CBMA) for improving probabilistic E p predictions in different climates of Iran. To this end, the parameters of the Adaptive Neuro-Fuzzy Interface System (ANFIS) were optimized using four meta-heuristic algorithms of Seagull Optimization Algorithm (SOA), Crow search Algorithm (CA), Firefly Algorithm (FFA), and Particle Swarm Optimization (PSO) for finding global solutions. The ANFIS-SOA, ANFIS-CA, ANFIS-FFA, ANFIS-PSO, and ANFIS models were implemented as inputs for employing ensemble CBMA and Bayesian Model Averaging (BMA) methods. Daily meteorological variables of average air temperature (T a), sunshine hours (SH), relative humidity (RH), wind speed (WS), and E p from six stations from 2000 to 2003 were applied. Evaluation of five Posterior Distribution Functions (PDFs) and three copula functions showed that Gumbel marginal distribution and Gumbel–Hougaard copula function provide the smallest Kolmogorov–Smirnov statistic indicator at the 5% significance level for all predictive models. The results established that the ensemble CBMA approach exhibited the highest prediction accuracy in all climates, followed by the BMA model, superior to the other individual models. The average root mean square error (RMSE) of the ensemble CBMA model was lower than BMA, ANFIS-SOA, ANFIS-CA, ANFIS-PSO, ANFIS-FFA, ANFIS, by 20.35%, 43.19%, 51.28%, 56.74%, 61.15%, and 64.36%, respectively. Furthermore, the uncertainty analysis indicated the E p predictions were more confident after applying CBMA. In conclusion, we highly recommend applying the ensemble CBMA model to improve E p 's prediction performance by standalone and hybrid machine learning models. The combination of T a -SH was suggested to be used for robust predicting daily E p in the regions of hot desert, cold desert, and cold semi-arid and T a -RH in the humid subtropical area regarding the convenience of data acquisition. • Copula Bayesian Model Averaging is adopted to predict probabilistic pan evaporation. • Copula functions are evaluated to capture joint and marginal distributions. • The assumption of Gaussian distribution in BMA led to decreasing performance accuracy. • The predictions of the ensemble CBMA model outperform at different climate regions. • The T a and SH variables have important roles to satisfy E p predictions. [ABSTRACT FROM AUTHOR]