Ensemble forecasting is still in its infant stage and a new frontier of NWP family and has many areas to be yet developed. Below are listed a few potential areas.

(a) Complimentary role of lower-resolution (low-res) ensemble and higher-resolution (hi-res) single forecast. Hi-res single run has smaller-scale detail spatial features and is more accurate for short range in general but lacks of uncertainty information, while low-res ensemble provides uncertainty information but is less accurate for short range and lacks of smaller-scale features. How to best combine low-res ensemble with hi-res single run is an important and practical issue which needs to be further explored (Roebber et. al., 2004; Kong et. al., 2006 and 2007). Hybrid Ensembling approach is one of such efforts, which superposes forecast variances from low-res ensemble on a single hi-res run by adding the difference between hi-res and low-res control forecasts to each ensemble members to improve the overall performance of an ensemble (Du, 2004). This method is found very useful and effective in improving an ensemble and has been operationally implemented for both global and regional EPSs at NCEP. Similar approach could also be applied to ETKF-based data assimilation (Gao et. al., 2007).

(b) Adaptive ensemble systems. Flow-dependent coupling among different scale EPSs might be desired. For example, a three-tier (global, regional and local) flow-dependent adaptive system could work as follows. A regional EPS varies its membership and model spatial resolution based on large-scale flow situation guided by a global EPS: if the global EPS projects less predictable flow or high-impact potential over the region, more members and higher resolution should be used to run the regional ensemble; otherwise, fewer members with coarser resolution could be used by the regional EPS. A global or larger-domain EPS should also determine where to run and which possibly distinct clusters/analyses to initiate a hi-res local EPS focusing on individual high-impact event or local flow for special needs (Molteni et. al., 2001). However, it should be carefully studied how to effectively couple multi-scale systems with each other to provide users maximum useful forecast information: does the smaller-scale EPS act purely as a downscaling tool of the larger-scale EPS by sharing same ICs, LBCs and perturbations or should each system have independent inputs to maximize diversity by using different ICs, LBCs and perturbations? Adaptive or manual-interactively manipulated perturbations in both IC and model configuration depending on specific weather systems or features of your interest, seasons and geographic regions might be found effective too in a local-scale ensemble forecasting (Homar et. al., 2006).

(c) Coupling with data assimilation (DA) process. With ETKF or similar technique, ensemble forecasting and DA can be coupled as one unified NWP component: ensemble provides truly flow-dependent background error to DA, while DA provides multiple analyses to initiate the ensemble model integrations with more realistic IC perturbations truly representing “error of the day” (see Part 3 (5)).

(d) Flow-dependent dynamical post-processing methods are needed as discussed in Part 5. How to combining ensemble approach with traditional statistical approach such as ensemble MOS to produce better forecast guidance should be a fruitful area for exploration too.

(e) Using ensemble in adaptive observation (Bishop and Toth, 1999; Bishop et. al. 2001; Majumdar et. al., 2002; Palmer et. al., 1998). How to effectively use ensemble information for adaptive observation or targeting is still in its infant stage and more research and field experiments are needed. Current limited efforts are mainly focusing on large-scale winter storms of relatively higher predictability. Using regional ensemble to target warm season convective systems and heavy precipitation is obviously important but a more challenging task (Du et. al, 2007b). When forecast error is propagating at a mixture of both phase-speed and group-speed, it’s obviously a more complex situation to be investigated in targeting technique.

(f) Ensemble dynamics. A complete and mature theoretical framework is needed for ensemble forecasting which basically doesn’t exist at the moment. It’s little known and researched about the error dynamics. We need to precisely describe how error evolves and propagates in the governing equations of a model to establish a theoretical ensemble dynamics (Farrell, 1990; Nicolis, 2004; Vannitsem and Toth, 2002).

(g) To better serve user community, broad efforts are needed to understand, communicate and work with specific end-users to develop optimal economic value based decision-making strategies by scientifically incorporating forecast uncertainty information. The newly established American Meteorological Society’s Ad Hoc Committee on Uncertainty in Forecasts is a good first step of this.

(h) Last but not least, a highly interactive, flexible and user-friendly visual display software capability needs to be developed for easy manipulating large amount of ensemble data and generating various ensemble products as well as performing ensemble verification. Help from software engineers is obviously necessary on this effort.