Forecast assimilation: a unified framework for the combination of multi-model weather and climate predictions

Abstract

In this paper we present a unified conceptual framework for the creation of calibrated probability forecasts of observable variables based on information from ensembles of weather/climate model predictions. For the same reasons that data assimilation is required to feed observed information into numerical prediction models, an analogous process of forecast assimilation is required to convert model predictions into well-calibrated forecasts of observable variables. Forecast assimilation includes and generalizes previous calibration methods such as model output statistics and statistical downscaling. To illustrate the approach, we present a flexible variational form of forecast assimilation based on a Bayesian multivariate normal model capable of assimilating multi-model predictions of gridded fields. This method is then successfully applied to equatorial Pacific sea surface temperature grid point predictions produced by seven coupled models in the DEMETER project. The results show improved forecast skill compared to individual model forecasts and multi-model mean forecasts.