Understanding forecast value in complex decision-making systems

应用气象预报产生社会经济价值是世界天气研究计划（WMO-WWRP）的一个主要重点，其中“气候服务”的增长尤其迅速。然而，尽管这些新兴服务明显以“最终用途”为重点，但在气象预报的价值评估中，受影响的人类或环境系统的复杂性往往被忽视。这是理解上的一个重大差距，因为决策取决于复杂的人类或环境系统对天气输入的反应，其中许多可能是非直觉的。该项目解决了这一研究缺口，并将在如何评估气象预报价值以供决策方面引发一个阶段性变化。亚季节到季节性的数值天气预报为未来数周甚至数月的概率气象预报提供了可能。虽然在提高预报质量方面作出了相当大的努力，但对预报价值（希望根据预报采取行动的决策者的利益；Murphy 1993）的理解仍然有限。迄今为止，用户价值通常是通过在固定时间范围内应用的静态“成本-损失”框架来看待的（即，在某个时间点上“行动”/“不行动”的二元决策；Murphy 1993, Richardson 2000）。尽管这种方法是通用的，并提供了有用的见解，但它在许多现实世界的决策中受到严重限制。特别是，天气状态不一定与复杂受影响系统的特定状态唯一地联系在一起（Brayshaw 2018），而且决策也经常：-取决于预测误差随时间推移的先前“轨迹”-包含多层行动（即，在任何给定时间，可能在多个前置时间内采取一整套行动）-随着新预报的可用而更新。具有这些属性的决策存在于许多部门（例如，能源、电信）。在这个项目中，CASE的合作伙伴负责运营英国约90%的固定线路电信基础设施，将提供决策环境的具体示例。在英国，估计每年330亿英镑的净经济贡献可归因于电信基础设施（Kelly, 2015），但固定线路电信的暴露性质导致了重大的天气风险（BT年度报告2013-2018）。因此，熟练地预测“故障”（即，提高故障率预测质量）和管理资源以进行后续维修（即，提取预测值）是一个关键问题，管理决策需要跨越许多时间范围，包括短期操作（周）和长期计划（年）。本项目将涉及两个研究领域：1。开发新方法，以最大限度地提高基础设施应用影响预测的质量。例如，这包括将基于模式的预测方法与基于网格点的预测方法进行对比。理解复杂决策系统中预测价值和预测质量之间的联系。这将在复杂程度和时间尺度的层次结构中利用一系列决策示例（例如，日常操作与季节性计划）。访问业务数据和确定决策协议通常是理解预测值的主要障碍。然而，通过CASE合作伙伴，学生将有机会获得与英国大部分固定线路电信网络相关的历史数据和决策协议。除了开发出新颖和通用的科学方法外，该项目还有望为英国基础设施的一个关键方面提供改进天气管理技术的设计信息。

The application of meteorological forecasts to produce socio-economic value is a major focus of the World Weather Research Programme (WMO-WWRP), with "climate services" particularly seeing rapid growth. However, despite the apparent "end use" focus of these nascent services, the complexity of the impacted human- or environmental- system is often neglected in value assessments of meteorological forecasts. This is a significant gap in understanding, as decision-making is contingent on the responses of complex human- or environmental- system to weather inputs, many of which can be non-intuitive. This project addresses this research gap and will lead a step-change in how meteorological forecast value is assessed for decision making.Subseasonal-to-seasonal numerical weather prediction offers the potential for skillful probabilistic meteorological forecasts weeks- to months- ahead. While considerable effort has been devoted to improving forecast quality, an understanding of forecast value (the benefit to a decision-maker wishing to act on a forecast; Murphy 1993) remains limited. To date, user-value has been typically viewed through a static "cost-loss" framework applied at a fixed time-horizon (i.e., a binary decision of "act"/"do not act" at some time-point; Murphy 1993, Richardson 2000). Although this approach is general and provides useful insights, it is severely limited for many real-world decisions. In particular, weather states cannot necessarily be uniquely linked to particular states of a complex impacted system (Brayshaw 2018) and decisions also often:- depend on the preceding "trajectory" of forecast errors over time- contain multi-leveled actions (i.e., at any given time, a whole set of actions may be taken across multiple lead times)- updated as new forecasts become available.Decisions with these properties are found in many sectors (e.g., energy, telecommunications). In this project, a CASE partner responsible for operating ~90% of the fixed-line telecommunications infrastructure in Great Britain will provide a concrete example of a decision-making context. In the UK, an estimated net economic contribution of £33bn/year is attributable to telecommunications infrastructure (Kelly, 2015) but the exposed nature of fixed-line telecommunications leads to significant weather risk (BT annual reports 2013-2018). Skillfully predicting "faults" (i.e., improving fault rate forecast quality) and managing resources for their subsequent repair (i.e., extracting forecast value) is therefore a key issue, with management decisions taken across many time-horizons for both short-term operations (weeks) and longer-term planning (years). This project will address two research areas:1. Developing novel methods to maximize the quality of impact forecasts for infrastructure applications. This includes, for example, contrasting pattern-based forecasting methods against grid-point based forecasts.2. Understanding the link between forecast value and forecast quality in complex decision-making systems. This will draw on a range of decision-making examples across a hierarchy of complexity and time-scale (e.g., day-to-day operations vs. seasonal planning).Accessing business data and identifying decision protocols is typically a major obstacle for understanding forecast value. However, through the CASE partner, the student will have access to both historic data and decision protocols pertaining to a large part of the UK fixed-line telecommunications network. Alongside the novel and general scientific methods developed, the project is therefore also expected to contribute to informing the design of improved weather-management techniques for a key aspect of UK infrastructure.