Application of data assimilation for improved operational water level forecasting on the northwest European shelf and North Sea

Abstract

For the Netherlands, accurate water level forecasting in the coastal region is crucial, since large areas of the land lie below sea level. During storm surges, detailed and timely water level forecasts provided by an operational storm surge forecasting system are necessary to support, for example, the decision to close the movable storm surge barriers in the Eastern Scheldt and the Rotterdam Waterway. In the past years, a new generation operational tide-surge model (Dutch Continental Shelf Model version 6) has been developed covering the northwest European continental shelf. In a previous study, a large effort has been put in representing relevant physical phenomena in this process model as well as reducing parameter uncertainty over a wide area. While this has resulted in very accurate water level representation (root-mean-square error (RMSE) ∼7–8 cm), during severe storm surges, the errors in the meteorological model forcing are generally non-negligible and can cause forecast errors of several decimetres. By integrating operationally available observational data in the forecast model by means of real-time data assimilation, the errors in the meteorological forcing are prevented from propagating to the hydrodynamic tide-surge model forecasts. This paper discusses the development of a computationally efficient steady-state Kalman filter to enhance the predictive quality for the shorter lead times by improving the system state at the start of the forecast. Besides evaluating the model quality against shelf-wide tide gauge observations for a year-long hindcast simulation, the predictive value of the Kalman filter is determined by comparing the forecast quality for various lead time intervals against the model without a steady-state Kalman filter. This shows that, even though the process model has a water level representation that is substantially better than that of other comparable operational models of this scale, substantial improvements in predictive quality in the first few hours are possible in an actual operational setting.