The Dynamics of Error Growth and Predictability in a Model of the Gulf Stream. Part II: Ensemble Prediction

Abstract

For any forecasting system, the ability to reliably estimate the skill of a forecast in advance (i.e., at the time the forecast is issued) is clearly desirable. In this paper the potential of ensemble prediction for estimating both the skill of forecasts of the Gulf Stream and the predictability of the ocean is examined. Using ensemble prediction methods the author has investigated how effective different types of perturbations are for perturbing the initial conditions of the ensemble members. The perturbations considered include the singular vectors, finite-time normal modes, and adjoint finite-time normal modes of a linearized version of the forecast model. The relationship between the skill of a forecast and the spread of an ensemble of forecasts about a reference forecast (the?control?) is examined as a function of (a) the type of perturbations used to perturb the ensemble members and (b) various different measures of forecast skill and ensemble spread. Assuming that the forecast model is perfect the author finds that a statistically significant relationship exists between skill and spread for forecast periods beyond one week. Specifically, a low (high) spread in the ensemble members relative to a control forecast is accompanied by a high (low) control forecast skill. In a nonperfect model, a statistically significant relation still exists between skill and spread, but it tends to deteriorate after forecast times of about a week. In general, singular vectors and linear transformations of the adjoint finite-time normal modes are most effective for perturbing ensemble members and yield statistically significant relationships between skill and spread over a wide range of skill and spread values. The skill?spread relationships identified appear to be insensitive to the details of the ensemble experiment.