Storm surge AI: machine learning can help with coastal protection

Contents

Researchers at the University of Hamburg have succeeded in training an artificial intelligence system to predict the frequency and height of storm surges over the next ten years quite reliably. The knowledge gained from this could be useful for coastal protection, dyke construction and a safe infrastructure, according to a press release. The results of the study were published in the journal Geophysical Research Letters.

Three locations in particular were investigated: Cuxhaven in Germany, Esbjerg in Denmark and Delfzijl in the Netherlands. The team led by Dr. Daniel Krieger from the Center for Earth System Research and Sustainability trained a neural network using weather data and the traditional climate computing model MPI-ESM developed by the Max Planck Institute for Meteorology.

Reliable predictions

For the analysis, the team used the hourly water levels of a location, which have been measured for decades. For Cuxhaven, there were around 700,000 measured values, collected since 1940. The researchers used these values, weather maps and air pressure data to feed a statistical model whose algorithm is self-learning. The model was only given 80 percent of the data. The rest remained secret so that the model could be tested later. The team then linked 10-year forecasts with the AI model to obtain specific forecasts for one location.

According to the university, the forecasts proved to be reliable. For example, an average of 11.6 storm surges per year were recorded in Cuxhaven over the last ten years. The model came up with 12.8 storm surges per year for the same period, with an upper and lower tolerance range of 1.6 storm surges. The best prediction quality was achieved for Delfzijl – where the agreement with the observations was 83 percent.

Storm surges are higher

Forecasts for individual years, on the other hand, proved to be less reliable according to the study. The forecasts for periods of seven to 10 years, on the other hand, were significantly better. "Until 2029, the value remains similar with 12 storm surges per year," says Krieger. The height of the floods is different: "While the highest annual storm surge in the past ten years averaged 2.5 meters, our model indicates an average of three meters for the next five years."

According to the university, it has so far been possible to use climate models to calculate whether more storms will occur in the North Sea area in the future – but not how they will affect certain coastal locations. The fact that this is now possible provides insights for coastal protection and dyke construction, and is of great importance considering rising sea levels.

Only one second per forecast

The calculation time is also remarkable. This is just one second for a forecast. This is several hundred times faster than conventional climate models. According to the paper, the challenges are that training can only be carried out for locations with long historical measurement data. In addition, the interactions between tides and storm surges are not fully understood. Another factor is strong currents coming from the Atlantic.

(mki)