Improved non-linear transfer function and neural network methods of flow routing for real-time forecasting

Abstract

Data-based methods of flow forecasting are becoming increasingly popular due to their rapid development times, minimum information requirements, and ease of real-time implementation, with transfer function and artificial neural network methods the most commonly applied methods in practice. There is much antagonism between advocates of these two approaches that is fuelled by comparison studies where a state-of-the-art example of one method is unfairly compared with an out-of-date variant of the other technique. This paper presents state-of-the-art variants of these competing methods, non-linear transfer functions and modified recurrent cascade-correlation artificial neural networks, and objectively compares their forecasting performance using a case study based on the UK River Trent. Two methods of real-time error-based updating applicable to both the transfer function and artificial neural network methods are also presented. Comparison results reveal that both methods perform equally well in this case, and that the use of an updating technique can improve forecasting performance considerably, particularly if the forecast model is poor.

Streamflow Modelling: A Primer on Applications, Approaches and Challenges

“…Methods such as cross-validation have been developed to avoid the overfitting problem (Hsieh, 2009). While ANNs have consistently been shown to outperform older, established versions of traditional Box-Jenkins methods (e.g., Castellano-Méndez et al, 2004;Jain and Kumar, 2007), it has been suggested by Lekkas et al (2001) that state-of-the-art variants of such traditional methods can perform just as well as ANNs in some cases.…”

“…Other hydrologically relevant applications are also common. For instance, ANNs have been used as an empirical downscaling tool for the direct prediction of streamflows from GCM output (Cannon and Whitfield, 2002), and a univariate approach has been employed in some hydraulic modelling cases (e.g., Lekkas et al, 2001;Jain and Kumar, 2007). For further information on the basic principles behind ANN and some common applications see Dawson and Wilby (2001).…”

“…Similarly, Shim et al (2002) found that increasing the number of nodes in the hidden layer increased training time exponentially and did little to reduce model error. Lekkas et al (2001) compared an MLP with one hidden layer to an ANN whose architecture was allowed to evolve during training and found that the latter model outperformed the popular MLP model in simulating discharge. It is worthwhile to note that, although in practice optimal performance may obviously be more challenging to achieve, in principle an MLP with a single hidden layer can model any continuous function (Cybenko, 1989;Hornik et al, 1989)-potentially a tremendous advantage to this approach.…”

Simulation of flood flow in a river system using artificial neural networks

“…Typical applications involve the training of two-to three-layer networks using suitable network architectures like multilayer perceptron, radial basis networks or recurrent networks. The performances of the ANNs in river flow prediction have been found to be comparable with other data driven modelling approaches (Lekkas et al, 2001;Sivakumar et al, 2002).…”

Application of neural network and adaptive neuro-fuzzy inference systems for river flow prediction

“…The performance of ANNs in river flow forecasting has been evaluated by various researchers and compared with other data-driven techniques. The ANN technique has proved its potential in forecasting river flow with promising results (Lekkas et al, 2001;Sivakumar et al, 2002).…”