Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Flo: A data-driven limited-area storm surge model

Nils Melsom Kristensen, Mateusz Matuszak, Paulina Tedesco, Ina Kristine Berentsen Kullmann, Johannes Röhrs

Abstract

We present Flo, a data-driven storm surge model, covering the North Sea, Norwegian Sea and Barents Sea. The model is built using the Anemoi framework for creating machine learning weather forecasting systems, developed by the European Centre for Medium-Range Weather Forecasts and partners. The model is based on a graph neural network, and is capable of simulating water level due to atmospheric effects (wind stress and inverse barometer effect, i.e. the non-tidally induced part of the total water level; the residual water level) at a horizontal resolution of 4 km and a temporal resolution of 1 hour with a quality comparable to the numerical model on which it was trained. The model was trained using a dataset consisting of 43 years of atmospheric data from the 3-km Norwegian Reanalysis hindcast for mean sea level pressure and winds, and the NORA-Surge hindcast for water level. Evaluation was done by comparing results from hindcast runs of the Flo model against independent observations of more than 90 water level gauges along the European coast, and against the NORA-Surge hindcast. The evaluation shows that Flo produces hindcasts with accuracy similar to the NORA-Surge hindcast, and it is shown that the model can resolve key physical processes. As the NORA-Surge hindcast used for training does not include data assimilation, Flo is not expected to systematically outperform the numerical model when evaluated against observations. Nevertheless, the present work represents an important step towards complementing traditional physics-based storm surge modelling with machine learning approaches and the framework establishes a strong foundation for future developments, particularly for training storm surge models that offer more flexibility for incorporating observations and other additional data sources.

Flo: A data-driven limited-area storm surge model

Abstract

We present Flo, a data-driven storm surge model, covering the North Sea, Norwegian Sea and Barents Sea. The model is built using the Anemoi framework for creating machine learning weather forecasting systems, developed by the European Centre for Medium-Range Weather Forecasts and partners. The model is based on a graph neural network, and is capable of simulating water level due to atmospheric effects (wind stress and inverse barometer effect, i.e. the non-tidally induced part of the total water level; the residual water level) at a horizontal resolution of 4 km and a temporal resolution of 1 hour with a quality comparable to the numerical model on which it was trained. The model was trained using a dataset consisting of 43 years of atmospheric data from the 3-km Norwegian Reanalysis hindcast for mean sea level pressure and winds, and the NORA-Surge hindcast for water level. Evaluation was done by comparing results from hindcast runs of the Flo model against independent observations of more than 90 water level gauges along the European coast, and against the NORA-Surge hindcast. The evaluation shows that Flo produces hindcasts with accuracy similar to the NORA-Surge hindcast, and it is shown that the model can resolve key physical processes. As the NORA-Surge hindcast used for training does not include data assimilation, Flo is not expected to systematically outperform the numerical model when evaluated against observations. Nevertheless, the present work represents an important step towards complementing traditional physics-based storm surge modelling with machine learning approaches and the framework establishes a strong foundation for future developments, particularly for training storm surge models that offer more flexibility for incorporating observations and other additional data sources.
Paper Structure (13 sections, 2 equations, 14 figures, 3 tables)

This paper contains 13 sections, 2 equations, 14 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Training data
  4. Model architecture
  5. Training
  6. Inference
  7. Observation dataset
  8. Results and Discussion
  9. Comparison against NORA-Surge
  10. Evaluating barotropic wave propagation
  11. Comparison against observations
  12. Case study of Storm Xaver
  13. Summary and Conclusions

Figures (14)

  • Figure 1: Histograms showing the distribution of residual water level in the entire NORA-Surge training dataset, for all times and all grid points. Panel (a) shows a 2D year-by-year comparison of residual water level distribution (values on the color scale are logarithmic), whereas panel (b) show combined distribution for the relevant training, validation and test periods, also on a log scale. The blue, green and yellow background colors in panel (a) depict the training, validation and test periods, respectively. In addition, the horizontal dashed red line marks the year 2013 that contains the storm Xaver.
  • Figure 2: The storm surge model domain covering the North Sea, the Norwegian Sea and the Barents Sea is shown by the shaded area.
  • Figure 3: Example of how the mapping of the physical model grid (data grid) space is mapped into the hidden mesh via the encoder in panel (a) and back into data grid via decoder in panel (b).
  • Figure 4: Visualization of how the processor connects the nodes of the hidden mesh via edges at different refinement levels. Panel (a) shows the edges between nodes at refinement level 4-6, where lower number equals longer edges. Panel (b) shows the edges and connections between the nodes at refinement levels 7-10 for the area outlined by the square of dotted lines in panel (a).
  • Figure 5: The training loss for every 100 training steps as a function of steps, as calculated by Anemoi-training. We show the training loss for both of the training runs, together with a smoothed line for each of the two.
  • ...and 9 more figures