Storm surges, due to hurricanes or extra-tropical depressions, are awesome natural hazards for which improved forecasting procedures can be used to mitigate against their potential destructive capacity. Many nations now possess operational forecasting systems that combine numerical models with real time observations of sea level. Such models, usually based on hydrodynamic equations, have a long history in coastal flood warning. Whilst they have been very successful, and form the backbone of current operational forecast procedures, these models are inherently limited by inaccuracies in bathymetry, meteorological forcing and parameterisations of sub-grid scale processes. The same modelling methods are often used to simulate the behaviour of extreme sea level in response to future climate scenarios. The science that underpins the next generation of storm and hurricane surge forecasting tools is vitally important and is an international concern.

There are now real opportunities for alternative, data-driven methods of surge prediction using artificial intelligence (e.g. neural networks), and rule-based models (RBMs). RBMs can provide a high-level linguistic representation of the mapping between input and output variables in a prediction problem, allowing for more understandable models which give an insight into important underlying relationships. Such models can also be extended to incorporate both the fuzzy and probabilistic uncertainty typically present in hydrological and oceanographic applications.

This talk will examine each component of a tide-surge forecast procedure, the opportunities for involving artificial intelligence systems, and the practicalities within an operational framework. Areas where forecast improvements are possible are a better understanding of the physics and scales of surge generation, and the nature of non-linear interaction with tides, as well as subtle problems of model validation against observational data. The value of complementary, non-deterministic techniques is – in many cases – enhanced for regions lacking powerful super-computer systems. This ease of use is a strong reason for integrating deterministic and artificial intelligence methods.