Neural network architectures and overtopping predictions

D. C. Wedge, David Ingram, C. G. Mingham, D. McLean, Z. A. Bandar

Research output: Contribution to journalArticlepeer-review


Overtopping of seawalls presents a considerable hazard to people and property near the coast and accurate predictions of overtopping volumes are essential in informing seawall construction. The methods most commonly used for the prediction of time-averaged overtopping volumes are parametric regression and numerical modelling. In this paper overtopping volumes are predicted using artificial neural networks. This approach is inherently non-parametric and accepts data from a variety of structural configurations and sea-states. Two different types of neural network are considered: multi-layer perceptron networks and radial basis function networks. It was found that the radial basis function networks considerably out-perform both the multi-layer perceptron networks and the curve-fitting (parametric regression) regime, and approach bespoke numerical simulations in accuracy. Unlike numerical simulation, the neural network approach gives generic prediction across a range of structures and sea-states and therefore incurs considerably less computational cost.
Original languageUndefined/Unknown
Pages (from-to)123-133
Number of pages11
JournalProceedings of the ICE - Maritime Engineering
Issue numberMA3
Publication statusPublished - 2005


  • coastal engineering
  • sea defences

Cite this