Az ANFIS mesterséges neurális hálózat lehetséges bemeneti adatainak vizsgálata csapadék-lefolyás modellezés esetén

Abstract

Artificial neural networks (ANNs) are gaining popularity across various scientific fields. Thus, studies to analyze the applicability of ANNs have also been appearing within the field of hydrology. In the case of rainfall-runoff modeling, one of the most promising networks is the Adaptive Neuro-Fuzzy Inference System (ANFIS), which effectively combines the learning capability of neural networks and the flexible structure of fuzzy systems. ANFIS can model nonlinear phenomena and identify nonlinear components using, for example, the currently applied Takagi-Sugeno type system. In this study, nine-nine ANFIS-based rainfall-runoff models with different input datasets were compared for the catchments of Torna- and Arany-creek. The models were built using the MATLAB software and the "anfis" function. The input datasets included precipitation, antecedent discharge, antecedent precipitation index, temperature, and potential evaporation. The preprocessing of the data included the examinations of quality and standardization. The models can be divided into two groups: six-six models included antecedent discharge as an input, while three-three models did not use the antecedent discharge data. The sensitivity analysis of the models revealed that an optimal number of initial FIS was 2, and 500 training epochs were sufficient. Based on the results from the nine-nine models, ANFIS-based rainfall-runoff models demonstrated adequate model efficiency during calibration, but their performance decreased during validation. Whether the models used the discharge time-series as an input or not, the best-performing models were those that included all the examined input datasets. However, some models produced outlier values at certain time steps, which could be a result of the computational methods or the structure of the neural network.