History or past memory always sheds lights to our future path if it’s interpreted and applied properly and is, therefore, an important aspect of weather forecasting (Cao, 2002). In a 3D-EPS, past-memory or history dimension is also considered besides varying IC and model. Direct Time-Lagged ensemble is a typical approach to bring this history dimension in. The degree of consistency from run to run in immediate past should be a measure of forecast uncertainty: high (less) consistency indicates high (low) predictability of an event. This agrees with forecaster’s experience: a forecaster is usually more (less) confident when he sees high-consistency (jumpiness) between runs from cycle to cycle. A main concern of using the past-time dimension is that forecast quality degrades with the age of a forecast: older forecasts perform worse than newer forecasts. However, as model and IC quality improves, this might not be true any more if the past time covered is not too old but only immediate cycles. For example, it’s not uncommon to observe that a 48hr model prediction could be more accurate than a 24hr prediction. As frequency of running a model at NWP centers increases (e.g., four times per day is a normal practice for many models at NCEP and even more frequently for special models such as NOAA RUC - Rapid Update Cycle - model which is run every hour), the information contained in those immediate past cycles could be huge and needs to be utilized more cost-effectively. An advantage of this dimension is no sacrifice in model resolution. All members are integrated with the highest possible full resolution as the single high-resolution run with no extra computing cost. However, not much research was done to seriously evaluate how to combine a 2-D ensemble with Direct Time-Lagged ensemble into a new 3-D ensemble to improve overall ensemble performance and provide more useful information to users. The situation might improve when more people realize the importance of this past-time dimension in ensembling.

    In real world operation, no single EPS (or single type of ensemble product) is universal and satisfies all needs but multiple scale ensemble systems are needed to serve a variety of forecasting purposes. Those multiple systems should work interactively and seamlessly with each other in some kind of adaptive ways (Subsection 2.7). Each system has its own uniqueness in construction and addresses its own unique problems. For example, a climate EPS focuses on the trend of climate change such as due to greenhouse gas-induced global warming or natural variability issues; a seasonal EPS on month to year scale of dominant weather mode such as warm or cold, wet or dry etc.; a global EPS on median-range 3-14 day’s large-scale flow pattern and serves early warming purpose; a regional EPS on short-range 1-3 day’s detail weather events with more focusing on surface weather elements; a relocatable storm-scale local EPS on 6-24hr details of a particular individual high-impact storm over a specific region of interest such as severe convective storm outbreak, fire weather, hurricane and disastrous event (natural or human-caused); and micro-scale ensemble such as ensemble cloud, turbulence and PBL (planet boundary layer) schemes. Different scale EPS obviously needs different strategies in perturbing ICs and model. For example, both environment and vortex (structure and intensity) need to be perturbed for hurricane prediction (Zhang and Krishnamurti, 1999; Cheung and Chan, 1999a and 1999b); how and what to perturb in ICs (warm bubbles?), might physics play more important role than IC, and how to assimilate special observations like Doppler Radar data into ICs etc. are all issues need to be studied in a convective storm-scale ensemble; fire weather, dispersion ensemble might focus more on near-surface elements and PBL winds and structures; and longer-range forecasts need to consider ocean-atmosphere coupled EPS which lower boundary forcing such as SST etc. should be important; …, just to mention a few.