Neural Network Hydrological Modelling: An Evolutionary Approach
Neural networks are now extensively used for rainfall–runoff modelling. Considered a ”black box” approach, they allow non-linear relationships to be modelled without explicit knowledge of the underlying physical processes. This method offers many advantag
- PDF / 521,591 Bytes
- 12 Pages / 441 x 666 pts Page_size
- 2 Downloads / 236 Views
Neural Network Hydrological Modelling: An Evolutionary Approach A.J. Heppenstall, L.M. See, R.J. Abrahart, and C.W. Dawson
Abstract Neural networks are now extensively used for rainfall–runoff modelling. Considered a “black box” approach, they allow non-linear relationships to be modelled without explicit knowledge of the underlying physical processes. This method offers many advantages over earlier equation-based models. However, neural networks have their own set of issues that need to be resolved, the most important of which is how to best train the network. Genetic algorithms (GAs) represent one method for training or breeding a neural network. This study uses JavaSANE, a system that advances on traditional evolutionary approaches by evolving and optimising individual neurons. This method is used to evolve good performing neural network rainfall–runoff solutions for the River Ouse catchment in the UK. The results show that as lead times increase, the JavaSANE networks outperform conventional feedforward networks trained with backpropagation. Keywords Genetic algorithms · neural networks · rainfall–runoff modelling
23.1 Background Hydrological systems are complex entities containing numerous non-linear processes which are often interrelated at different spatial and temporal scales. This complexity presents substantial problems in both understanding and modelling hydrological processes. This is particularly problematic in the area of rainfall–runoff modelling (Zhang and Govindaraju, 2000). The development of mathematical models to understand A.J. Heppenstall School of Geography, University of Leeds, Woodhouse Lane, Leeds, LS2 9JT, UK L.M. See School of Geography, University of Leeds, Woodhouse Lane, Leeds, LS2 9JT, UK R.J. Abrahart School of Geography, University of Nottingham, Nottingham, NG7 2RD, UK C.W. Dawson Department of Computer Science, Loughborough University, Loughborough, LE11 3TU, UK
R.J. Abrahart et al. (eds.), Practical Hydroinformatics. Water Science c Springer-Verlag Berlin Heidelberg 2008 and Technology Library 68,
319
320
A.J. Heppenstall et al.
this system occupies a substantial amount of the literature where efforts are concentrated on attempting to capture the characteristics of the underlying physical processes through the use of equations of mass and momentum (Jain and Srinivasulu, 2004). However, many of these models have varying degrees of success being limited by the need for large quantities of high-quality data as well as detailed calibration and optimisation methods. Due to the constraints of these physically – based models, attention has shifted to the use of “black box” methods such as neural networks. Neural networks can be used to develop relationships between input and output variables that form part of a process without having explicit knowledge of the underlying physics. These models are now extensively used to perform rainfall–runoff modelling (see other chapters in this book on neural network modelling for many good examples). This uptake is due to the advantages th
Data Loading...
