Stochastic Parameterization of Deep Convection in Short-Range Ensemble Weather Forecasts

用于进行天气预报和模拟地球气候的数值模式不能明确地表示所有空间和时间尺度上的运动。相反，必须使用参数化方案来考虑那些具有短尺度的运动。一个模型将包含若干这样的方案，每一个方案都与已省略的一个特定的小规模物理过程有关。这些将包括大气边界层的湍流、重力波和深层潮湿对流。对流方案是本项目的重点，因为已知现有的深对流表示是造成天气预报和气候模型中一些最大和最顽固的系统误差的原因。传统上认为参数化方案是确定的。因此，方案的输入来自本地的当前状态，已解析的规模流，输出对于给定的输入是唯一的。其原理是，小尺度运动可以被统计地考虑，并且对它们的总体平均效应的估计被反馈到大尺度。近年来，这种确定性假设受到了挑战：忽略小尺度运动的波动分量可能是无效的，它能够与分辨率尺度流动强烈相互作用。有充分的证据表明，忽视这种波动不仅在理论上不能令人满意，而且可能对模型性能产生重大影响。一个有吸引力的替代方法是使用随机动态参数化方法，其目的是解释波动。随机方案的原理是反馈特定小尺度状态的影响。根据小尺度运动的统计模型随机选择状态。2005年6月，欧洲中期天气预报中心举办了一个专题研讨会。在《论文集》（第七页）中，目前的情况是这样总结的：“随机动态参数化是一个相对较新的概念，但它有可能对天气和气候预报的所有领域产生重大影响。”迄今为止，关于随机方法的研究一直令人鼓舞，并倾向于分为两类。在一类中，随机成分的处理相对简单，但作为一个完整的预测系统的一部分，并受到广泛的测试。例如，目前正在研究的方法包括纳入MOGREPS（英国气象局全球和区域综合预报系统），这是一个用于业务天气预报的新系统。在另一类中，为随机组件构建了详细的模型，但测试通常相当有限，并且在某种理想化的配置中进行。Plant & Craig的随机对流参数化就是一个例子。这个项目的关键问题，也是社区的一个主要问题，是努力构建随机变异的详细模型是否值得，或者简单的处理是否足够。为了回答这个问题，首先有必要在一个完整的操作预测系统中实施一个详细的、最先进的随机参数化，其次将其性能与更简单的变异性处理进行比较。在这里，Plant & Craig方案将作为MOGREPS的一部分实施，并与操作系统并行评估其性能。在实施方案时，为了进行适当的比较，有必要确定随机变率相关的长度和时间尺度。该项目将在一个通用的背景下建立这些尺度和它们的敏感性（即，这些结果将不是特定于Plant & Craig方案）。这是因为相关尺度的知识对于任何随机动态模型的使用都是重要的。

The numerical models that are used to perform weather forecasts and to simulate the earth's climate are incapable of representing explicitly the motions on all space and time scales. Rather, those motions with short scales must be taken into account by using parameterization schemes. A model will contain a number of such schemes, each relating to a particular small-scale physical process that has been omitted. These will include turbulence in the atmospheric boundary layer, gravity waves and deep, moist convection. The convection scheme is the focus in this project since existing representations of deep convection are known to be responsible for some of the largest and most stubborn systematic errors in weather forecasting and climate modelling. Parameterization schemes have traditionally been assumed to be deterministic. Thus, the input to a scheme is taken from the current state of the local, resolved-scale flow and the output is unique for a given input. The philosophy is that the small-scale motions can be considered statistically and an estimate of their ensemble-mean effect is fed back to the large scale. In recent years, this deterministic assumption has been challenged: it may not be valid to neglect the fluctuating component of the small-scale motions, which is capable of interacting strongly with the resolved-scale flow. There is good evidence to suggest that neglect of such fluctuations is not just theoretically unsatisfactory but that it may have significant impacts on model performance. An attractive alternative is to use a stochastic-dynamic parameterization method, which aims to account for the fluctuations. The philosophy of a stochastic scheme is to feed back the effects of a particular small-scale state. The state is chosen at random based on a model for the statistics of the small-scale motions. In June 2005, a workshop on the subject was organized by the European Centre for Medium-range Weather Forecasts. In the Proceedings (p. vii) the current situation is summarized thus: 'Stochastic-Dynamic Parameterization is a relatively new concept, yet one that has potential to impact significantly on all areas of weather and climate forecasting.' Studies to date on the stochastic approach have been consistently encouraging and have tended to fall into two distinct categories. In one category, the stochastic component is treated in relatively simply, but is included as part of a full forecast system and subject to extensive testing. Examples include methods currently being investigated for inclusion in MOGREPS (Met Office Global and Regional Ensemble Prediction System), a new system for operational weather forecasting. In another category, detailed models are constructed for the stochastic component, but testing is typically rather limited and occurs in somewhat idealized configurations. An example is the stochastic convective parameterization of Plant & Craig. The key question for this project, and a major issue for the community, is whether efforts to construct detailed models of the stochastic variability are worthwhile, or whether a simple treatment might be sufficient. In order to answer that question, it is necessary first to implement a detailed, state-of-the-art stochastic parameterization into a full operational forecast system, and second to compare its performance with simpler treatments of variability. Here, the Plant & Craig scheme will be implemented as part of MOGREPS and its performance assessed in parallel with the operational system. When implementing the scheme, and to allow for appropriate comparisons, it is necessary to determine the length and time scales over which the stochastic variability is to be correlated. The project will establish these scales and their sensitivities in a generic context (i.e., these results will not be specific to the Plant & Craig scheme). This is because knowledge of the correlation scales is important for the use of any stochastic-dynamic model.