Probabilistic Multi-Model Prediction System in Operational Seasonal Forecasting: Basic Concepts and Global-to-Local Scale Applications

Abstract

A multi-model ensemble (MME) approach is a relatively recent contribution to climate forecasting and has been considered as one of effective ways to improve the climate forecast and to quantify forecast uncertainties due to model formulation. As a result, the MME prediction system is currently exploited by a few operational centers that provide seasonal forecasts. However, the majority of the previous studies on MME are focused on deterministic interpretation based on ensemble mean although the superiority of the multi-model concept is more obviously evidenced in a probabilistic framework. Therefore, the advantages of MME may be more clearly illustrated when dealing with probabilistic predictions. Climate forecasts are associated with uncertainty because of stochastic nature of the climate system and the level of the uncertainty can be most appropriately viewed in a quantitative way by using probabilities. In this direction, many studies demonstrated that a probabilistic forecast derived from an ensemble prediction system is of greater benefit than a deterministic forecast by providing useful information on the inherent uncertainty in the climate system. As a result, it has been shown that probabilistic forecasts are of greater value to decision makers and end users than deterministic forecasts. First, this study makes considerable efforts to develop the most appropriate (uncalibrated) probabilistic MME approach for use in an operational global prediction system that combines a large set of models, with individual model ensembles essentially differing in size and its ensemble sizes in hindcast and forecast being inconsistent. Additionally, key issues of probabilistic multi-model ensemble prediction, how to estimate tercile-based categorical probabilities and how to combine the forecast probabilities from multiple models, are discussed. Second, efforts are devoted to improving probabilistic multi-model prediction (PMMP) system by calibrating the single-model predictions using an upgraded multi-variable version of a stepwise pattern projection method and the combination based on skill-based model selection among calibrated single-model predictions to formulate multi-model probabilistic prediction. Finally, in order to advance the concept of MME prediction to local-scale prediction and to explore the utility of such a forecast system for potential end users, a new approach to estimation of forecast uncertainty with the regression-based statistical downscaling and multi-model predictions is suggested and applied to 60 Korean stations.