Extended‐range predictions with ecmwf models: Influence of horizontal resolution on systematic error and forecast skill

Abstract The influence of horizontal resolution on model systematic error and skill scores are examined for both winter and summer seasons from a set of extended‐range integrations of the European Centre for Medium Range Weather Forecasts (ECMWF) numerical weather prediction model. the forecast data are composed of 24 30‐day integrations, two per month from successive 12 GMT operational analyses, from April 1985 to March 1986 at four (triangular truncation) horizontal spectral resolutions of T21, T42, T63 and T106. Systematic errors in wind, temperature, heat and momentum fluxes are discussed, both in the tropics and extratropics. Dynamical reasons for differences between various model resolutions and between simulations and observations are proposed. Diagnosis of model systematic error is aided by comparing zonal‐mean diagnostics of the observed flow in the northern hemisphere winter and southern hemisphere winter. the impact of resolution on regional simulations of rainfall during the Indian and African monsoons is also discussed. In the extratropical troposphere, the behaviour of the T21 model is quite distinct from the higher‐resolution models. the southern hemisphere flow is weak, and in both hemispheres horizontal momentum fluxes are severely underestimated. the systematic errors of the T42, T63 and T106 models are quite similar to each other in the extratropical troposphere. Errors at these higher resolutions resemble the observed difference (in zonal‐mean wind, temperature and eddy fluxes) between diagnostics in the northern hemisphere winter and southern hemisphere winter, suggesting that orographic forcing in this version of the ECMWF model is inadequate, despite the inclusion of envelope orography. In the extratropical stratosphere, the behaviours of the T21 and T42 models resemble each other, both having a cold polar bias, apparently associated with inadequate dynamical heating and therefore relaxation to radiative equilibrium. This in turn is related to insufficient Rossby wave focusing into the polar vortex at T21 and T42. the T63 and T106 resolution integrations have much smaller stratospheric cold‐pole biases. In the tropics, there are serious systematic errors at all resolutions; however, some of these mean errors increase with increasing resolution. These include the global divergent and nondivergent wind errors, and more regional simulations such as the monsoon flow, where simulation of diabatic heating associated with convective activity is crucial. Since the local values of moisture flux convergence at T106 are effectively much noisier than the larger‐scale values of such fluxes used to drive the convection scheme at T21, it is speculated that diabatic heating fields are more likely to project onto the relevant tropical meteorological modes at low model resolution. On the other hand, it is shown that the ability to resolve local orographic features is important in accurately simulating local tropical precipitation maxima over land. Despite having smaller systematic errors in the northern hemisphere, 10‐day mean T21 model anomaly correlation coefficient scores were worse than those of higher‐resolution models, at least up to day 20. However, there was no indication that the higher‐resolution models had significantly different extended‐range skill characteristics. On the other hand, the increase of tropical systematic error with increasing horizontal resolution may indicate that the potential of higher‐resolution models in extended‐range prediction may be underestimated in this study. the asymptotic spread between forecasts initialized 24 hours apart indicated that all models substantially underestimated low‐frequency variability (by up to 50% in the extended summer period). Bearing in mind the computational burden of longer timescale integrations with complex numerical models of the atmosphere, the evidence presented in this paper suggests that the T106 resolution may be unnecessarily high for both extended‐range forecasting, and for climate simulation studies. However, at present, the cost of integrating at T63 resolution may also be excessive for many climate studies. It would appear from our analysis, that in the troposphere the behaviour of the T42 model is comparable with the higher‐resolution models, and is a satisfactory compromise for many purposes, particularly bearing in mind the impact of resolution on tropical systematic errors. However, it would appear that the extreme sensitivity of the 1985/86 ECMWF model climate drift to resolution in the range between T21 and T42, and the apparent inability of the ECMWF T21 model to simulate the correct internal nonlinear dynamics of the extratropics, make questionable its use for climate studies at resolution lower than T42.