An approach for tuning ensemble prediction systems

Reliable ensemble prediction systems (EPSs) are able to quantify the flow-dependent uncertainties in weather forecasts. In practice, achieving this target involves manual tuning of the amplitudes of the uncertainty representations. An algorithm is presented here, which estimates these amplitudes off-line as tuneable parameters of the system. The tuning problem is posed as follows: find a set of parameter values such that the EPS correctly describes uncertainties in weather predictions. The algorithm is based on approximating the likelihood function of the parameters directly from the EPS output. The idea is demonstrated with an EPS emulator built using a modified Lorenz'96 system where the forecast uncertainties are represented by errors in the initial state and forecast model formulation. It is shown that in the simple system the approach yields a well-tuned EPS in terms of three classical verification metrics: ranked probability score, spread-skill relationship and rank histogram. The purpose of this article is to outline the approach, and scaling the technique to a more realistic EPS is a topic of on-going research.