Introduction

      In early days, weather forecast was indeed called probability and issued as probabilistic format. For example, Prof. Cleveland Abbe issued the first official “Weather Synopsis and Probabilities” on February 19, 1871 (NRC, 2006). Later on, due to advance and new knowledge achieved in NWP as well as more observations available, scientists start to use single deterministic value to predict weather (still so today!).

After Lorenz discovered chaotic nature of atmospheric behavior/phenomena in 1960s, some pioneering scientists started to seriously reconsider stochastic approaches in predicting weather and climate. Given the fact that intrinsic uncertainties exist in each steps of a prediction process, we have no way to know the ground truth in an exact fashion. Instead, a complete and faithful description of, say, initial condition and model physics must be in a probabilistic distribution and stochastic in nature within a certain range of uncertainty. As a result of this and the chaotic nature of highly nonlinear numerical models (Lorenz, 1993), there might be a multiple of possible realizations for each forecast. In other words, a complete forecast must also be described in a probabilistic distribution with forecast uncertainty explicitly expressed but not in a single deterministic value!

In 1969, Epstein (1969) first proposed a theoretical Stochastic-Dynamic approach to directly describe forecast error distributions (mean, variance and probability density function) in model equations. Unfortunately, it’s unrealistic to integrate such a system with limited computing power since the number of forecast equations required to be solved is huge for a real atmospheric system. Instead, Leith (1972) proposed a more practical Monte-Carlo approach with limited forecast members. Each forecast member is initiated with randomly perturbed, slightly different initial condition (IC). He pointed out that with as few as eight members, the average of member could give a best estimation of a forecast with adequate accuracy although more members might be needed for forecast variance estimation. With an analytical turbulence equation, Leith showed that Monte-Carlo method is a practical approximation to Epstein’s Stochastic-Dynamical approach. Leith’s Monte-Carlo approach is basically the traditional definition of ensemble forecasting although the content of this definition has been greatly expanded in the last 20 years to include the following: (a) perturbing all uncertain components in a state-of-the-art forecasting system such as physics, numeric and boundary forcing besides perturbing atmospheric ICs (observation and analysis), and (b) flow-dependent IC perturbations with dynamically growing structure rather than statistical, random perturbation (see Part 3).

As computing power increases, operational ensemble forecasting became a reality in early 1990s. Both NCEP and ECMWF (European Center for Medium-Range Weather Forecast) operationally implemented its own global model-based, medium-range ensemble forecast system in December 1992, respectively (Tracton and Kalnay, 1993; Toth and Kalnay, 1993; Mureau et. al., 1993; Molteni et. al., 1996). At the same time, a few people realized that predictability issue is not only relevant to medium-range but also to short-range forecasts and therefore started to research on regional model-based short-range ensemble forecasting (Mullen and Baumhefner, 1994; Mullen and Du, 1994; Brooks et. al., 1995; Du et. al., 1997 and 2000; Mullen et. al., 1999). An operational Short-Range Ensemble Forecasting (SREF) system was under development and evaluation over the North American domain at NCEP since 1995 (Tracton and Du, 1998; Stendsrud et. al., 1999; Hamill and Colucci, 1997 and 1998; Hou et. al., 2001) and became a part of U.S. National Weather Service (NWS) real-time production suite in April 2001 (Du and Tracton, 2001) which is the first real-time operational regional ensemble system among major NWP centers in the world. A time-lagged ensemble forecasting approach was also operationally used for seasonal prediction (9 months) at NCEP from 2004 based on a global model coupled with ocean (Sara et. al., 2006).

Since the initial implementations of NCEP and ECMWF ensemble systems, ensemble approach has been widely accepted and actively pursued at almost all other major NWP centers around the globe such as Houtekamer et. al. (1996), Ebert (2001), Li and Chen (2002), Wang and Kann (2005), Eckel (2005), Chien et. al. (2006), Tennant et. al. (2007), Teixeira et. al. (2007) and Matsueda et. al. (2007). Research on ensemble forecasting also gained its strength since later 1990s and early 2000s and has merged as a hot topic in NWP nowadays (Buizza et. al., 1999a; Mullen and Buizza, 2001; Hansen, 2002; Grimit and Mass, 2002; Bright and Mullen, 2002a and 2002b; Hamill et. al., 2000 and 2004a; Legg and Mylne, 2004; Wandishin et. al., 2005; Eckel and Mass, 2005; Jones et. al., 2007; Yuan et. al., 2005 and 2007c; Jankov et. al., 2007). It’s expected that the ensemble-based probabilistic forecasting will play more and more important role in shaping the future of numerical weather prediction practice and service in years to come.

This lecture series provides a comprehensive review and discusses some general principles on ensemble forecasting to let readers have a big picture about what’s involved in this relatively new and rapidly developing branch of numerical weather prediction (NWP). For technical details, related references are provided at the last part (Part 11), so that interested researchers could study further in depth and join the active research community of this challenging frontier. First, the underline scientific reason why ensemble forecasting is needed was discussed in Part 1. The following parts discussed various aspects related to ensemble forecasting including what ensemble forecasting is aiming for (Part 2), how to build an ensemble prediction system (EPS, Part 3.1, 3.2, and 3.3), what products can be derived from an ensemble (Part 4), what is the role of EPS post-processing (Part 5), how to evaluate the quality of an EPS and its forecasts (Part 6). Due to its increasing importance, how to communicate forecast uncertainty and how to use probability information in users’ decision-making process were illustrated in Part 7. Recent development of downstream applications using meteorological ensembles as inputs was also mentioned in Part 8. Part 9 listed some major differences between the two forecast paradigms -- “single forecast” vs. “ensemble forecasts”. Part 10 mentioned possible future trend of ensemble-related development. Finally, a summary and references are given at Part 11.

Contact  Jun Du