Modelling long-term monthly temperatures by several data-driven methods using geographical inputs

This paper investigates the ability of five different data-driven methods, artificial neural networks (ANNs), adaptive neuro-fuzzy inference system (ANFIS) with grid partition (GP), ANFIS with subtractive clustering (SC), support vector regression (SVR) and gene expression programming (GEP), in predicting long-term monthly temperatures by using data from 50 stations in Iran. The periodicity component (month of the year), station latitude, longitude and altitude values were used as inputs to the applied models to predict the long-term monthly temperatures. The overall accuracy of the SVR model was found to be better than that of the other models. The GEP model gave the worst estimates. The maximum determination coefficient (R-2) values were found to be 0.996, 0.999, 0.997 and 0.959 for the ANN, ANFIS-GP, ANFIS-SC and GEP models in Karaj and Qazvin stations, respectively. The highest R-2 value (0.999) of SVR model was found for the Tabas station. The minimum R-2 values were respectively found as 0.988, 0.946 and 0.985 for the ANFIS-GP, ANFIS-SC and SVR models in Bandar Abbas station while the ANN and GEP models gave the minimum R-2 values of 0.982 and 0.886 in the Abadan and Kerman stations, respectively. The results indicated that the long-term monthly temperatures of any site can be successfully estimated by data-driven methods applied in this study using geographical inputs. The interpolated maps of temperatures were also obtained by using the optimal SVR model and evaluated in the study. The temperature maps showed that the highest temperatures were occurred in the southeastern and central parts of the Iran.