Combining machine learning with computational hydrodynamics for prediction of tidal surge inundation at estuarine ports

Accurate forecasts of extreme storm surge water levels are vital for operators of major ports. Existing regional tide-surge models perform well at the open coast but their low spatial resolution makes their forecasts less reliable for ports located in estuaries. In December 2013, a tidal surge in the North Sea with an estimated return period of 760 years partially flooded the Port of Immingham in the Humber estuary, on the UK east coast. Damage to critical infrastructure caused several weeks of disruption to vital supply chains and highlighted a need for additional forecasting tools to supplement national surge warnings. In this paper, we show that Artificial Neural Networks (ANNs) can generate better short-term forecasts of extreme water levels at estuarine ports. Using Immingham as a test case, an ANN is configured to simulate the tidal surge residual using an input vector that includes observations of surge at distant tide gauges in NW Scotland, wind and atmospheric pressure, and the predicted astronomical tide at Immingham. The forecast surge time-series, combined with the astronomical tide, provides a boundary condition for a local high-resolution 2D hydrodynamic model that predicts flood extent and damage potential across the port. Although the forecasting horizon of the ANN is limited, 6 to 24 hour forecasts at Immingham achieve an accuracy comparable to or better than the UK national tide-surge model and at far less computational cost. Use of a local rather than a larger regional hydrodynamic model means that potential inundation can be simulated very rapidly at high spatial resolution. Validation against the 2013 surge shows that the hybrid ANN-hydrodynamic model generates realistic flood extents that can inform port resilience planning.