Water Quality and Sediment Behaviour of the Future: Predictions for the
21st Century (Proceedings of Symposium HS2005 at IUGG2007, Perugia, July
2007). IAHS Publ. 314, 2007, 38-44
Neuroevolution
methodologies applied to sediment forecasting
ALISON J. HEPPENSTALL1, LINDA M. SEE1 & ROBERT J. ABRAHART2
1 School of Geography, University of Leeds, Leeds LS2 9JT, UK
2 School of Geography, University of Nottingham, Nottingham NG7 2RD, UK
Abstract
Sediment forecasting represents a significant modelling challenge. This is due
to the combined effects of suspended sediment transfer and throughput being source
limited and subject to hysteresis effects. Recent approaches to modelling and
forecasting have involved the use of neural networks. Despite yielding good
results, this method has its own set of limitations, for example lack of
guidance in parameter setting and the potential to overtrain. This paper
reports on the application of a neuroevolutionary toolbox, JavaSANE. This toolbox
is applied to two catchments in Puerto Rico that have been previously studied
by Kisi (2005), who used a range of different methods including a neuro-fuzzy
approach and neural networks to model suspended sediment in these catchments. These
experiments are replicated using JavaSANE and compared to the results reported
in Kisi (2005). These results show that JavaSANE produces estimates that are
better or comparable to those of Kisi (2005).
Key words genetic
algorithm; JavaSANE; neural network; neuroevolution; Puerto Rico; sediment