Estimation of Long-Term Monthly Temperatures by Three Different Adaptive Neuro-Fuzzy Approaches Using Geographical Inputs

This paper investigates the accuracy of three different adaptive neuro-fuzzy inference systems (ANFISs), ANFIS with grid partition (ANFIS-GP), ANFIS with substructive clustering (ANFIS-SC), and ANFIS with fuzzy c means (ANFIS-FCM) in estimation of long-term monthly air temperatures. Data of 71 stations in Turkey are used in the applications. The periodicity (month of the year) and geographical variables (latitude, longitude, and altitude) are used as inputs to the models. ANFIS models are also compared with artificial neural networks (ANNs) and multilinear regression (MLR). Three ANFIS methods provide superior accuracy to ANN and MLR methods, and the ability of the ANFIS-GP is observed to be superior to the ANFIS-SC and ANFIS-FCM models. Among the ANFIS methods, the worst estimates are obtained from the ANFIS-FCM method. The maximum determination coefficients (R2) are observed as 0.998, 0.995, and 0.995 for the ANFIS-GP, ANFIS-SC, and ANFIS-FCM at S.Urfa, Tunceli, and Usak stations, individually. The minimum R2 values are individually found as 0.902 and 0.921 for the ANFIS-GP and ANFIS-SC at Sinop station, whereas the ANFIS-FCM model gives a minimum R2 of 0.934 at Yalova and Yozgat stations in the testing stage. The outcomes show that long-term monthly air temperatures can be effectively assessed by the ANFIS-GP method using geographical inputs. The interpolated air temperature maps are likewise produced by using the ideal ANFIS-GP and are assessed in this study. The temperature maps demonstrate that the most noteworthy measures of temperatures happen in the southeastern, northeastern, western, and northwestern parts of Turkey.