A Proposed Annex Architecture for Enhancing Extrapolation Performance of MLP Neural Networks

Creative Commons License

Yılmaz S., Ekinci Ö.

EngOpt2014 - 4th International Conference on Engineering Optimization, Lisbon, Portugal, 8 - 11 September 2014, pp.149-150

  • Publication Type: Conference Paper / Summary Text
  • City: Lisbon
  • Country: Portugal
  • Page Numbers: pp.149-150
  • Kocaeli University Affiliated: Yes


Limited extrapolation capability of artificial neural networks (ANNs) is noticed and paid attention in several studies. An ANNs based extrapolation model, which includes associated methodologies in a systematic order, is proposed in order to overcome such limitations. Among these methodologies are recursive training, adding new hidden layer units, scaling the input variable in order to bring the inputs into a suitable range and selecting the good performance network architecture after trial and error. Estimation of the peaks flows is an important problem for the design, management and safety of any hydraulic structures. In addition to this problem, the hydrological processes are nonlinear, time varying and spatially distributed. In this study, an annual maximum flows estimation model (AMFE) is developed by means of ANNs, which is trained under the light of these mentioned prescriptions.  In this way, an acceptable extrapolation pattern is achieved for annual maximum flows data of the Cine-Kayirli gauging station in B. Menderes basin,  Turkey. The results are compared with the known statistical maximum flows estimation methods.  

Index Terms-Artificial neural networks, extrapolation, estimation, cascade correlation, recursive training.