An analysis of the error space of a high-resolution implementation of the GHER hydrodynamic model in the Mediterranean Sea

An ensemble of 250 model setups covering the Mediterranean Sea is built by perturbing various parameters: the bathymetry, the initial conditions, atmospheric forcing fields (air temperature, cloud coverage, wind), and internal model parameters (diffusion coefficients). The ensemble is then forwarded in time using the GHER hydrodynamic model, allowing to obtain information about the expected error associated with the forecast in a natural way. The evolution of this error is analyzed. In particular, we examine the time evolution and stationarity of its spatial average, and the spatial distribution of the error at different instants, by means of its first to fourth order moments, and of empirical orthogonal functions. We verify whether the a posteriori error distribution is Gaussian using the Anderson-Darling test. From these results, we are able to assess what parameters and forcing fields are most critical for the forecast. Qualitative conclusions are obtained throughout the text, in accordance with our expectations. Moreover, quantitative estimations of the expected error are also given.