Using Decision Tree and K-Means to Improve ANFIS for Predicting Missing Flow Data in Çoruh Basin

Complete and continuous flow data provides contributions to the water resources projects appropriate design. However, both financial and technical factors prevent the regular collection of data and such a situation leads to missing data problems. Input variables are from five stations in the Çoruh Basin in the northeast province of Türkiye, and output variables are from one station. In the study, 75% (171 months of data) of the 228 months flow data between the years 1993–2011 are used in the training phase and 25% (57 months of data) in the testing phase. In the ANFIS models, Decision Tree (DT) and K-means (KM) methods are used in selecting the input variables and determining the number of membership functions. In addition, artificial neural network (ANN), multiple linear regression (MLR), and ANFIS models with randomly determined cluster numbers are also used to compare the model performance of these methods. The results show that using DT and KM methods, the ANFIS models generate more reliable results than other models. During the study, the regression coefficient (R2), weighted mean square error (WMSE), and Wilcoxon (Z) values are taken into account for reliability measure. The DT–K-means–ANFIS model achieved the highest accuracy with R2 = 0.98 and WMSE = 5.89 during the testing phase, demonstrating superior performance to the ANN and MLR models. It is shown that the most successful models can be determined in a shorter time by using DT and KM methods before generating ANFIS models.