RI: Small: Large-scale Probabilistic Forecasting for Energy Systems

一组风力发电场在未来24小时内可能产生多少电力？商业建筑中的居住者将如何相互作用来消耗能源？能够回答这样的预测问题对于发展更可持续的能源基础设施至关重要：如果我们能够提前预测可再生能源的生产和需求，我们就可以更有效、更可靠地调度能源资源，从而大幅减少温室气体排放。不幸的是，这些也是我们需要预测的天生不确定的数量；例如，无论我们的算法有多好，我们都不能完美准确地预测人类的行为。为了使用这样的预测，我们需要能够对这些领域固有的不确定性进行适当的建模。我们需要做出预测，不仅要平均正确，而且要捕捉到复杂的随机波动和预测量之间的相关性。这个项目开发并使用了最近提出的建模框架-稀疏高斯条件随机场-通常使用的马尔可夫随机场的推广。该框架通过利用逆协方差矩阵中的稀疏性有效地对高维分布进行建模。该项目通过极大地加速模型学习，通过扩展现有理论来理解这些模型何时可以有效地学习高维预测器，并通过Copula方法将预测推广到非高斯设置，从而扩展了最新技术。该项目使用这些算法在能源行业的四个关键领域建立预测模型：能源需求、风力发电、家庭和商业建筑的用户使用量以及智能电表的个人能源消耗。该项目具有示范的更广泛的影响。首先，这项研究直接涉及对高效能源管理至关重要的应用领域，在这些领域，即使是微小的进步也可能对可持续性产生相当大的影响。其次，PI利用这项研究将电力系统和机器学习社区更紧密地联系在一起，在机器学习和电力系统场所传播结果，并向能源从业者发布材料和视频讲座。最后，该项目利用这项研究来增加STEM领域的多样性，方法是为研究生和本科生水平上代表性不足的少数群体提供建议，并让高中生和教师进行演讲，说明如何使用计算来解决可持续性问题。

How much power is a set of wind farms likely to generate over the next 24 hours? How will occupants in a commercial building interact to consume energy? Being able to answer prediction questions like these is vital to developing a more sustainable energy infrastructure: If we can predict renewable energy production and demand ahead of time, we can schedule energy resources more efficiently and reliably, leading to significant reductions in greenhouse gas emissions. Unfortunately, these are also inherently uncertain quantities we need to predict; for example, no matter how good our algorithms are, we can't predict human behavior with perfect accuracy. In order to use such predictions, we need to be able to properly model the uncertainty inherent in these domains. We need to make predictions that are not only correct on average, but which capture the complex random fluctuations and correlations between predicted quantities. Only then can we schedule energy resources in a way that accounts for these uncertainties.This project develops and uses a recently-proposed framework for modeling --- sparse Gaussian conditional random fields --- a generalization of the commonly used Markov random field. This framework efficiently models high-dimensional distributions by exploiting sparsity in the inverse covariance matrix. The project extends the state of the art by greatly accelerating model learning, by extending existing theory to understand when these models can effectively learn high-dimensional predictors, and by generalizing the predictions to the non-Gaussian setting through copula methods. The project uses these algorithms to build forecasting models in four crucial domains in the energy sector: energy demand, wind power, user occupancy in homes and commercial buildings, and personal energy consumption from smart meters.The project has exemplary broader impacts. First, the research deals directly with application domains crucial to efficient energy management, where even small advances can have a sizable impact on sustainability. Second, the PI leverages the research to bring the power systems and machine learning communities closer together, disseminating the results at both machine learning and power systems venues, and releasing material and video lectures to practitioners in energy. Finally, the project harnesses the research to increase diversity within STEM fields by advising under-represented minorities at the graduate and undergraduate level, and by engaging High School students and teachers with talks illustrating how computation can be used to address problems in sustainability.