The tendency for solutions to diverge in seasonal limited-area forecast models (or nested regional climate models) differs both from short-term forecast models and from global seasonal models. Short-term applications of limited-area models are strongly dependent on initial conditions, whereas such "memory" for initial conditions tends to decline for seasonal and longer simulations because of the continual input of data at the lateral boundaries. Global model solutions are free to diverge based on small differences in initial conditions or model physics, whereas the divergence of solutions in limited-area models may be constrained by specified lateral boundary conditions. Thus, appropriate methods for generating ensembles of seasonal forecasts using limited area models may differ both from short-term limited area models and from seasonal global models.

We have compared methods for generating ensemble simulations of seasonal precipitation using the summer 1993 flood (1 June - 31 July) over the north-central U.S. as a test case. The methods used are:

• Lagged-average-forecast ensemble: Several instances of the MM5 mesoscale mdoel are executed with the same physics and numerics but with differing initial conditions for each ensemble member. Each ensemble member is started at a different initial time with all simulations overlapping for the period of interest. Results for this overlapping period are used as members of the ensemble.
• Perturbed physics ensemble: This method uses MM5 with a single set of physics options, but internal parameters within the convective parameterization are varied to create realizations for the ensemble. All simulations use the same initial conditions.
• Mixed physics ensemble: A number of simulations are performed with MM5 using a variety of physics options but the same initial conditions. The resulting simulations are then evaluated as an ensemble.
• Multi-model ensemble: The various models participating in the PIRCS 1-B experiment, which considered the 1993 flood period, are considered as individual realizations of an ensemble.

Each ensemble was evaluated using a variety of error measures such as mean square error, equitable threat score, etc. Results show that the multi-model and mixed-physics ensembles had the largest spread; notably, the spread obtained by using different cumulus parameterizations was as large as the spread obtained by using completely different models. The lagged-average and mixed-physics ensembles had much lower spread and appear to be less useful as ensemble forecasts than the other two types of ensembles.