Rainfall–runoff modeling using adaptive neuro-fuzzy inference system (ANFIS) and genetic algorithm (GA). Issue 10 (2nd September 2022)

• Record Type: Journal Article
• Authors: Vakili, Shabnam; Mousavi, Seyed Morteza
• Is Part Of: Water Supply. Volume 22:Issue 10(2022)
• Page Start: 7460
• Page End: 7475
• Publication Date: 2022-09-02
• Subjects: adaptive neuro-fuzzyi Inference system (ANFIS) -- genetic algorithm (GA) -- modeling -- rainfall–runoff relations
• DOI: 10.2166/ws.2022.318
• Languages: English
• ISSNs: 1606-9749

Abstract

Abstract: Nonlinear properties and natural uncertainties in the rainfall–runoff process, the necessity of extensive data, and the complexity of the physical models have caused researchers to use methods inspired by nature such as artificial neural networks, fuzzy systems, and genetic algorithms (GA). The main purpose of this study was to estimate runoff employing Adaptive Neuro-Fuzzy Inference System (ANFIS) and GA models using accessible, applicable, and easily available climatic data. The results of the two models were compared to provide an easy but reliable model to estimate evaporation. The models were utilized to estimate the runoff in Sivand river basin located in Fars province in central Iran. The results were compared considering a range of model performance indicators as mean absolute error (MAE), Nash–Sutcliffe efficiency coefficient (NSE), root mean square error (RMSE), and correlation coefficient (R 2 ). According to the results presented, ANFIS with lower RMSE and MAE and higher correlation coefficient and NSE between the observed and predicted values provided higher accuracy in comparison to GA. Also, it was clear that using ANFIS, an increase in the number of membership functions and running cycles of the model decreased the error such that the results in the studied stations using were improved by 42, 44 and 11%, respectively by increasing the number of membership functions and run rounds. Also, it was observed that the nonlinear models performed better than the linear models when applying GA such that non-linearizing the model improved the results of the GA model in the three studied stations by 27.5%, 17%, and 9.5%, respectively. Meanwhile, considering RMSE amounts the best results from ANFIS were 23%, 54.6%, and 35.7% better than the best results from GA in the three stations, respectively. According to the results of the study runoff can be estimated appropriately by utilizing ony meteorological data and there is no need for more complex and interdependent data. A sensitivity analysis was conducted too by removing rainfall and evaporation parameters in two different scenarios. The ANFIS model showed the lowest sensitivity to the absence of those parameters especially evaporation in scenario 3 with RMSE = 0, 0, and −0.005 for Chambian, Dashtbal, and Tang Balaghi stations, respectively. The results of the study justifies using ANFIS employing only meteorological data to estimate runoff in areas when scant data are available. HIGHLIGHT: In this study, the effect of meteorological parameters on runoff estimation was investigated with different membership functions and rounds of ANFIS in an adaptive fuzzy neural inference system (ANFIS) and linear and nonlinear models in the genetic algorithm (GA). To the best of our knowledge, a study of the number of membership functions and rounds of ANFIS and the linearity and nonlinearity of the GA model using minimal meteorological data is a novel endeavor. Most researchers have focused on complex physiographic characteristics, which are difficult to access, while according to the results of this study runoff these can be estimated appropriately by utilizing minimum meteorological data. Also, as it is explained in the "sensitivity analysis" section, reliable results can be obtained by using ANFIS even after removing the evaporation parameter from the input data. Graphical Abstract