A data assimilation procedure for operational prediction of storm surge in the northern Adriatic Sea

• Published in: Continental shelf research 2006-03, Vol.26 (4), p.539-553
• Main Authors: Lionello, P., Sanna, A., Elvini, E., Mufato, R.
• Format: Article
• Language: English
• Subjects: Adjoint model; Data assimilation; Marine; Operational; Prediction; Storm surge
• ISSN: 0278-4343; 1873-6955
• DOI: 10.1016/j.csr.2006.01.003

A procedure designed for the operational prediction of storm surge in northern Adriatic Sea is presented. Its main purpose is to provide an early warning for the flood of the Venice city centre. The procedure uses a hydro-dynamical shallow water model (forced by the wind stress and sea level pressure fields) and assimilates hourly sea level observations (available at the “Aqua Alta” research platform, 15 km offshore the Venetian littoral) in order to produce an optimal surge forecast. A cost function describing the discrepancy between model results and observations is defined and the adjoint and the conjugate gradient methods are used for the computation of the cost function gradient and the search of its minimum, respectively. Each operational simulation is split into an analysis period (whose optimal length has been found to be 3 days) and a forecast period (prediction has been considered up to the 3-day range). During the analysis, the observations are assimilated in the model in order to identify the optimal initial condition for the surge prediction carried out for the following forecast period. A penalty term, which improves the stability of the assimilation procedure, has been identified and included in the cost function. The results show that the procedure is capable of compensating for the errors (due to both inaccurate meteorological forcing and model shortcomings) and effectively improves the reliability of the storm surge forecast. Moreover, this study proves that the main source of error for the short and medium range forecast, in such a weakly dissipative system, resides in an inaccurate initial condition, which can be improved by a variational data assimilation procedure.