Actionable deep learning under uncertainty for carbon-centric building operations

经营性碳是指用于保持建筑物温暖、凉爽、通风、照明和供电的所有能源所产生的温室气体排放。运营碳占全球温室气体排放量的28%。减少全球温室气体排放对应对气候变化至关重要。为此，能源资产和设备的使用应考虑到碳排放。能源系统的进步以及可再生能源和能源储存的整合使得在能源结构具有低碳能源的情况下运行它们成为可能。给定一个特定的地理位置，能源应该以最纯粹的形式消耗，以显着减少建筑存量的运行碳足迹。还可以考虑一些其他因素，如峰值需求和运营成本。这带来了一个多目标决策问题，该问题旨在优化资产/设备运营，从而重塑能源消耗概况，以获得更好的需求侧响应。为了解决这个问题，需要一个智能的、主动的解决方案，以考虑碳强度和电价等相关变量的未来变化，从而做出实时控制决策。该项目旨在重塑建筑消费概况，通过能源资产和设备的智能运营，如储能和灵活需求（如电动汽车），最大限度地跟踪电网的低碳发电。这将涉及开发一种用于建筑能源管理的深度学习方法（例如，深度强化学习），将实时监测和预测相关变量（如碳强度）考虑在内。在制定建筑控制决策时，我们将在碳和能源效率方面纳入不确定性并优化多个目标。此外，我们还将确定可能影响和促进决策过程的重要变量（例如，时间、天气、气候、空气质量）。该研究预计将对建筑存量产生更好的需求侧响应，并带来一系列效益，包括减少温室气体排放、能源成本和峰值需求。

Operational carbon refers to GHG emissions that come from all energy sources used to keep our buildings warm, cool, ventilated, lighted, and powered. Operational carbon contributes to 28% of global GHG emissions. Reduced global GHG emissions are crucial to combating climate change. For this purpose, the energy assets and appliances should be used considering carbon emissions. The advancement of energy systems and the integration of renewable generations and energy storage have made it possible to operate them when the energy mix has low carbon generations. Given a particular geographic location, the energy should be consumed in its purest form to significantly reduce the operational carbon footprint of the building stock. Some other factors may also be considered such as peak demand and operational cost. This brings out a multi-objective decision-making problem that is set to optimize the asset/appliance operations, thus reshaping the energy consumption profile for a better demand-side response. To address this problem, a smart, proactive solution is needed to account for the future changes of pertinent variables such as carbon intensity and electricity price in producing real-time control decisions.This project aims to reshape the building consumption profile to maximally track the low-carbon generation at the grid through smart operations of energy assets and devices such as energy storage and flexible demand (e.g., EVs). This will involve the development of a deep learning approach (e.g., deep reinforcement learning) for building energy management factoring in real-time monitoring and prediction of pertinent variables such as carbon intensity. We will incorporate uncertainties and optimize multiple objectives in the space of carbon and energy efficiency when making building control decisions. Further, we will also identify important variables (e.g., time, weather, climate, air quality) that can affect and contribute to the decision-making process. The research is expected to result in a better demand side response for the building stock and a range of benefits including reduced GHG emission, energy cost, and peak demand.