Algorithmic tuning of spread–skill relationship in ensemble forecasting systems

Abstract

In ensemble weather prediction systems, ensemble spread is generated using uncertainty representations for initial and boundary values as well as for model formulation. The ensuing ensemble spread is thus regulated through what we call ensemble spread parameters. The task is to specify the parameter values such that the ensemble spread corresponds to the prediction skill of the ensemble mean – a prerequisite for a reliable prediction system. In this paper, we present an algorithmic approach suitable for this task consisting of a differential evolution algorithm with filter likelihood providing evidence. The approach is demonstrated using an idealized ensemble prediction system based on the Lorenz–Wilks system. Our results suggest that it might be possible to optimize the spread parameters without manual intervention.