Robust, Trustworthy and Explainable Predictive Models for Low Carbon Power and Energy

电力和能源系统正处于过渡状态，这是因为需要对发电来源以及供暖和运输方面的最终用途进行脱碳[1]。这些变化是在遗留配电基础设施的背景下发生的，这些基础设施不是为低碳技术(LCT)设计的，几乎没有监控[2]。因此，随着大规模数字化，这种情况已经开始改变，这将需要更先进的机器学习方法来进行预测或摘要生成，以基于数据做出运营和规划决策。然而，模型(例如，当代的神经网络体系结构)正变得越来越复杂，数据流处于更高的维度--所做的决策需要被证明是合理的，而这种证明需要基于对模型输出的理解。数据质量和模型精度可能会受到操作噪声[3]或过时测量的影响，数据流也可能完全脱机[4]，但模型预测仍是可预期的。在非ML级别提供模型可解释性和可解释性，为运营最终用户提供了快速和主动地询问模型输出的便利，最终允许他们信任或放弃预测，从而使他们的业务受益。该博士将调查和开发支持采用新的预测分析的工具，以支持低碳电力和能源系统-受益于ScottishPower的多个业务部门：零售(例如需求预测)、可再生能源(例如最大限度地减少削减)、交易(例如失衡预测)和网络(例如支持拥堵管理)。博士学位的主要目标将是研究预测模型在业务上采用的障碍是如何表现出来的，并确定能够解决这些障碍的支持决策支持工具。博士学位的设想产出包括：(1)将由ScottishPower定义和提供的与问题案例研究相关的新模型；(2)在ScottishPower的IT系统上部署与新模型相关的软件演示工具，旨在利用ScottishPower的业务(例如Azure或AWS)提出的实际部署架构进行最终用户评估；(3)支持ScottishPower的数字中心计划，以确定模型部署战略，以造福于他们更广泛的业务。支持这些目标的研究将在以下领域进行：1.如何在模型预测的应用用例背景下，以人类可读的术语而不是统计或数据科学术语来解释模型预测。还将通过应用自然语言处理模型，调查利用维修报告[2、5、6]或标准文件等现有业务数据为这些解释生成背景的能力。模型预测中的可信度如何传播到决策中，以及如果模型可以提供其他信息，如可信度或与过去情景的关系，这些决策可能会如何改变[7]。3.如何保证模型的预测，即使是在数据不完整的情况下，也要通过缺失数据的填补方法来保证。这将反过来与目标1相关，因为输入的输入值将不可避免地改变模型输出。如何根据成本或收益而不是误差指标来解释模型预测能力。

Power and energy systems are in a state of transition, invoked by the need to decarbonise both the sources of power generation, but also its end use in terms of heating and transportation [1]. These changes are taking place against a backdrop of legacy power distribution infrastructure which was not designed for Low Carbon Technologies (LCT) and featured little or no monitoring [2]. Consequently, this has started to change with large scale digitalisation which will require more advanced machine learning methods for prediction or summary generation to make operational and planning decisions based on data. However, models (e.g., contemporary Neural Network architectures) are becoming ever more complex and data streams are in higher dimensions - decisions made need to be justified, and that justification needs to be based on the understanding of model output. Data quality and in turn model accuracy may be compromised by operational noise [3] or out of date measurements, and data streams may also be offline altogether [4], but model predictions are still expected. Providing model explainability and interpretability at a non-ML level provides the operational end-user with the facility to interrogate model outputs rapidly and initiatively, ultimately permitting them to trust or discard the predictions to the benefit of their business.This PhD will investigate and develop tools that will support the adoption of new predictive analytics to support a low carbon power and energy system - with benefit across multiple business units in ScottishPower: retail (e.g. demand forecasting), renewables (e.g. minimising curtailment), trading (e.g. imbalance forecasting) and networks (e.g. supporting congestion management). The main aims of the PhD will be to look at how the barriers to operational adoption of predictive models manifest and identify the supporting decision support tools that can resolve these. The envisaged outputs from the PhD include: (1) new models associated with problem case-study to be defined and provided by ScottishPower; (2) the deployment of software demonstrator tools related to the new models on ScottishPower's IT system intended for end-user evaluation utilising the real deployment architecture proposed by ScottishPower's business (e.g. Azure or AWS); (3) support ScottishPower's Digital Hub initiative in determining a strategy for model deployment to the benefit of their wider business.The research underpinning these objectives will be in the following areas:1. How model predictions can be explained in the context of their application use case and in human readable terms rather than statistical or data science terms. The ability to harness existing operational data such as maintenance reports [2, 5, 6] or standards documents to generate context for these explanations will also be investigated through the application of Natural Language Processing models.2. How confidence in model predictions propagates through to decisions made and how these decisions might be altered if models could provide additional information such as confidence level or relation to past scenarios [7]. 3. How to guarantee model predictions are undertaken every time even in the face of incomplete data through missing data imputation methods. This will in turn, relate to objective 1 as imputed input values will inevitably alter model outputs.4. How model predictive capability can be interpreted in terms of cost or benefit rather than an error metric.