Monthly streamflow estimation using wavelet-artificial neural network model: A case study on Camlidere dam basin, Turkey


UYSAL G., Sorman A. U.

9th International Conference on Theory and Application of Soft Computing, Computing with Words and Perception (ICSCCW), Budapest, Hungary, 22 - 25 August 2017, vol.120, pp.237-244 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 120
  • Doi Number: 10.1016/j.procs.2017.11.234
  • City: Budapest
  • Country: Hungary
  • Page Numbers: pp.237-244
  • Keywords: monthly streamflow estimation, neural netwok, multi-layer perceptron, wavelet transform
  • Anadolu University Affiliated: Yes

Abstract

Data driven techniques have become well-known application in hydrology in which physical processes are highly nonlinear. They require detailed analyses of different input combinations, selecting the appropriate model structures, assigning the optimization parameters etc. Besides, the model performance are also highly correlated with additional analysis techniques. In this study, the value of using different data sets such as air temperature, precipitation, evaporation and streamflow records, evapotranspiration around the basin are investigated to estimate monthly inflows using a multi-layer perceptron network model. Since the noise always exists in the time-series data, Discrete Wavelet Transform (DWT) is applied for data decomposition. Caml. dere dam basin, which is one of the vital water supply reservoir of the capital city of Turkey, Ankara, is selected as an application area. The model sets are employed using 1960 -2016 monthly observed data. The reliability of the modelled flows are verified with: coefficient of determination (R-2), Nash-Sutcliffe model efficiency (NSME), root mean square error (RMSE) and mean absolute error (MAE). According to the results, instead of increasing input vector number, application of data pre-processing have more impact to capture especially high flows. Decomposed discharge data together with meteorological other inputs perform 0.85 - 0.73 both for R-2 and NSME for training and testing periods, respectively. (c) 2018 The Authors. Published by Elsevier B.V.