The Met. Office has developed a variational assimilation for its Unified Model forecast system, which contains a grid‐point mode) that is run operationally in global, mesoscale, and stratospheric configuration. Key characteristics of the design are: a development path from three‐dimensional to four‐dimensional variational assimilation; global and limited‐area configurations; variational analysis of perturbations; and a carefully designed, well conditioned background term. The background term is implemented using a sequence of variable transforms to independent balanced and unbalanced variables, to vertical modes, and to spectral coefficients. The coefficients used are based on statistics from differences of one‐ and two‐day forecasts valid at the same time. The covariance model represents many of the features seen in the covariances of forecast differences. The three‐dimensional variational data assimilation (3D‐Var) system was implemented in the operational global forecast system on 29 March 1999. In parallel trials, the 3D‐Var system gave a 2.7% improvement in a composite skill score (verified against observations and weighted according to the importance of each field).

S u M M ARYPrevious studies and different methods of estimating short-range forecast errors are summarized. Zonally and temporally averaged statistics based on differences of one-and two-day forecasts valid at the same time are presented and an attempt is made to explain many of the features by reference to dynamical concepts.Vertical correlation length-scale tends to increase with horizontal correlation scale but to be very short in the Tropics; horizontal scale is longest in the Tropics and in the stratosphere. The variations in vertical correlation are much more pronounced for largely balanced variables such as rotational wind and temperature than they are for divergent wind or humidity. The extratropics are dominated by an equivalent barotropic mode with the level of the maximum wind amplitude (and the zero crossing of the temperature correlation) being determined by the tropopause. Surface pressure is negatively correlated with low-level temperature as expected (except over the Antarctic plateau where it is positively correlated); it is also negatively correlated with temperatures near/above the tropopause in the extratropics.The covariance model used in The Met. Office Global Three-Dimensional Variational (3D-Var) Data Assimilation system represents the variation of vertical covariances with latitude reasonably well, but the longer horizontal scales in the stratosphere are not currently reproduced. The implied covariances used operationally have been modified so that the correlation length-scales, both horizontal and vertical, are somewhat shorter than those direct from the forecast differences. Recent changes to the representation are briefly described, with an indication of their impact on the forecasts. The impacts are significant relative to other changes tested, and the covariance model has played a major role in the successful implementation and subsequent improvement of our 3D-Var system. * We currently take it to be 1.0; Rabier et u1. (1998) scale their standard deviations by 0.9.

SUMMARYThe Met. Office has developed a variational assimilation for its Unified Model forecast system, which contains a grid-point model that is run operationally in global, mesoscale, and stratospheric configurations. Key characteristics of the design are: 0 a development path from three-dimensional to four-dimensional variational assimilation; 0 global and limited-area configurations; 0 variational analysis of perturbations; 0 and a carefully designed, well conditioned background term.The background term is implemented using a sequence of variable transforms to independent balanced and unbalanced variables, to vertical modes, and to spectral coefficients. The coefficients used are based on statistics from differences of one-and two-day forecasts valid at the same time. The covariance model represents many of the features seen in the covariances of forecast differences.The three-dimensional variational data assimilation (3D-Var) system was implemented in the operational global forecast system on 29 March 1999. In parallel trials, the 3D-Var system gave a 2.7% improvement in a composite skill score (verified against observations and weighted according to the importance of each field).

A new humidity analysis variable has been successfully introduced into the Met Office's variational data assimilation system. The new variable uses a transformation consisting of a nonlinear normalization and a link with temperature increments that is a function of background (forecast) humidity. The normalization (which is the more important aspect) makes the new variable's errors more symmetrical and thus better represented by the variational cost function. As in the previous operational system, water vapour and cloud are combined into a single total water variable for the analysis step. The transform is now operational in both global and limited-area systems. The forecast improvements -to the mass as well as the humidity fields -are largest in the Southern Hemisphere and there is a better fit of both background and analysis to humidity-sensitive satellite channels. The results suggest that the transformation is particularly beneficial for the use of satellite data. The transform reduces the problem of negative humidities in the analysis which were more prevalent over the ocean.

A key attribute of a probabilistic forecast system is its reliability: the degree to which forecast probabilities agree with outcome frequencies. Here, we focus on short‐lead‐time (12 h) reliability in the nonlinear background forecasts of the Ensemble of Data Assimilations (EDA) from the European Centre for Medium‐Range Weather Forecasts (ECMWF). A ‘reliability budget’, derived from consistency arguments, is used to separate the mean‐squared departures of the ensemble mean (relative to observations) into bias, ensemble variance and observation‐error contributions, along with a residual that indicates a deficiency in reliability. At these short lead times, the residual is found to be sensitive, in a local manner, to the assignment of observation errors and to the parametrization of ‘stochastic physics’, which accounts for the deficit in a model's error growth rate. In particular, the results highlight the importance of the stochastic physics parametrization to represent error growth rates fully in convective regions and suggest that current stochastic physics may be too active in subtropical anticyclones, where the mid‐tropospheric meteorology is largely characterized by time‐mean descent and radiative cooling. Other results demonstrate how the reliability budget can be used to tune observation errors, which leads to an improvement in diagnosed background reliability. Although there remains some ambiguity in the attribution of deficiencies, the budget represents a useful additional tool that can help stimulate improvements in model stochastic error representation and observation‐error estimates. Such improvements should help facilitate the development of more reliable ensemble forecasts in future.