Statistical Analysis of HWRF Errors for Accuracy Assessment of Coupled Hydrodynamic Modelling Systems

The frequency and destructiveness of coastal storms have required improving the accuracy of numerical prediction models. The coupling of atmospheric, ocean wave and surge and hydrological models on high-resolution numerical domains has improved model accuracy by better representing nearshore/inland geometries and physics. But multiple sources of error remain, from instrument and processing noise in the observational data the models are built upon, to the physical parameterizations that are part of the models, to the stochasticity of the natural processes themselves.

Determining the damage caused by hurricanes using such numerical models requires a statistical evaluation of uncertainty. The US COASTAL Act of 2012 mandated coupled modelling of hurricanes, in order to ascribe building damages as caused by wind or water, at a high level of accuracy. The COASTAL Act modelling system consists of a coupled atmosphere-ocean model providing the wind component (the Hurricane Weather and Research Forecasting model, HWRF), which then forces models of riverine flooding, ocean waves and surge (the water components). Conventionally, time series of observations at fixed in-situ locations such as meteorological stations, wave buoys and tide and stream gauges, and spatiotemporal data along satellite footprints are used to assess the accuracy of a given model. In this study, we statistically evaluate the outputs of deterministic simulations from the HWRF model and an ocean wave model forced by HWRF. The evaluation is done at stationary observation locations and along satellite altimeter tracks for Hurricane Irma (2017). The models’ uncertainty is also determined via analysis of 40 ensemble members, corresponding to 40 sets of initial conditions of driving variables. The high number of ensemble members allows to capture the spread of HWRF prediction errors, ensuring that hydrodynamic models are forced with a wide enough ensemble. We analyze the statistical distributions of time series of model errors, for particular locations and important hurricane variables. We provide an exploratory method to assess the similarity between observations and HWRF model estimates which is general enough to be useful across many geophysical variables. This is particularly important for the error propagation required by complicated coupled model systems, such as for the US COASTAL Act.