Modeling daily discharge responses of a large karstic aquifer using soft computing methods: Artificial neural network and neuro-fuzzy


Kurtulus B., Razack M.

JOURNAL OF HYDROLOGY, cilt.381, ss.101-111, 2010 (SCI İndekslerine Giren Dergi) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 381
  • Basım Tarihi: 2010
  • Doi Numarası: 10.1016/j.jhydrol.2009.11.029
  • Dergi Adı: JOURNAL OF HYDROLOGY
  • Sayfa Sayıları: ss.101-111

Özet

This paper compares two methods for modeling karst aquifers, which are heterogeneous, highly non-linear, and hierarchical systems. There is a clear need to model these systems given the crucial role they play in water supply in many countries. In recent years, the main components of soft computing (fuzzy logic (FL), and Artificial Neural Networks, (ANNs)) have come to prevail in the modeling of complex non-linear systems in different scientific and technologic disciplines. In this study, Artificial Neural Networks and Adaptive Neuro-Fuzzy Interface System (ANFIS) methods were used for the prediction of daily discharge of karstic aquifers and their capability was compared. The approach was applied to 7 years of daily data of La Rochefoucauld karst system in south-western France. In order to predict the karst daily discharges, single-input (rainfall, piezometric level) vs. multiple-input (rainfall and piezometric level) series were used. In addition to these inputs, all models used measured or simulated discharges from the previous days with a specified delay. The models were designed in a Matlab (TM) environment. An automatic procedure was used to select the best calibrated models. Daily discharge predictions were then performed using the calibrated models. Comparing predicted and observed hydrographs indicates that both models (ANN and ANFIS) provide close predictions of the karst daily discharges. The summary statistics of both series (observed and predicted daily discharges) are comparable. The performance of both models is improved when the number of inputs is increased from one to two. The root mean square error between the observed and predicted series reaches a minimum for two-input models. However, the ANFIS model demonstrates a better performance than the ANN model to predict peak flow. The ANFIS approach demonstrates a better generalization capability and slightly higher performance than the ANN, especially for peak discharges. (C) 2009 Elsevier B.V. All rights reserved.