CPS: Medium: Federated Learning for Predicting Electricity Consumption with Mixed Global/Local Models

这项提议旨在将联合学习与电力系统相结合，利用来自众多设备的分布式数据来更好地预测电力消耗并降低发电成本。我们的目标是利用电力领域越来越常见的数据来源，即每15分钟记录一次用电量的智能电表的激增。我们将开发机器学习方法，根据这些数据预测未来一天的用电量。这种学习必须在对不愿与中央当局共享信息的最终用户提供隐私保障的情况下进行。由于电力市场的结构，使这些预测比目前的做法更准确，可以以更低的成本和更可持续的环境方式生产电力。我们建议训练递归神经网络用于时间序列预测，而不是与公用事业共享来自每个用户的完整数据集，而是通过公用事业和消费者之间的重复交互来保护消费者数据的隐私。要使这一愿景成为现实，需要两项科学进步。首先，我们必须开发有效、样本高效和快速的“嵌套”联邦学习方法，这种方法可以处理同时是局部和全局的模型。我们需要一个全球模型来捕捉家庭用电的共同模式，但我们也需要一个本地模型来捕捉每个家庭的独特特征。需要的第二个进步是用于从时间序列数据中学习的神经体系结构，该体系结构能够捕获数据中的长期相关性。事实上，电力消耗表现出长期的依赖关系，人类行为是复杂的，因此任何潜在的模式总是受到无法直接建模的噪声的破坏。这里提出的方法的发展可以降低整个美国的电力成本。更间接的是，它可以为安装智能电表的倡议提供额外的动力，这些智能电表可以在更高的粒度测量电力消耗，同时向消费者保证他们的数据是安全的。最后，它可以使引入依赖天气的可再生发电变得更容易，这为预测时空电力供需平衡带来了一系列新的挑战，与公用事业公司为适应不确定的可再生能源发电预测而设计的消费者需求响应激励相关。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This proposal aims to integrate federated learning with power systems, leveraging distributed data from numerous devices to better predict electricity consumption and lower the cost of generation. Our goal is to take advantage of data sources which are becoming more common in the power domain, namely the proliferation of smart meters which record electricity consumption at 15- minute intervals. We will develop machine learning methods which predict electricity consumption at the day-ahead scale from this data. Such learning must be done with privacy guarantees for end-users who are hesitant to share information with a central authority. Due to the way power markets are structured, making these predictions more accurately than current practice allows electricity to be produced at lower cost and in a more environmentally sustainable way. We propose to train recurrent neural networks for time series prediction without sharing the full data sets from each user with the utility, but rather through repeated interactions between the utility and the consumer which preserve the privacy of consumer data. To make this vision a reality will require two scientific advances. First, we must develop effective, sample efficient, and fast methods for "nested" federated learning which can deal with models that are simultaneously local and global. We need a global model to capture common patterns of electricity consumption across households, but we also need a local model to capture the idiosyncratic features of each household. The second advance required is a neural architecture for learning from time series data which is capable of capturing long-term dependencies in the data. Indeed, electricity consumption exhibits long-term dependencies and human behavior is complex so that any underlying pattern is always corrupted by noise which cannot be modeled directly.The development of the methods proposed here could reduce the cost of electricity throughout the United States. More indirectly, it could provide additional steam to initiatives to install smart meters which can measure electricity consumption at a higher level of granularity, while at the same time assuring consumers that their data is safe. Finally, it could make it easier to introduce weather dependent renewable generation, which creates a new set of challenges for predicting spatiotemporal electricity supply-demand equilibria associated with consumer demand response incentives designed by utilities to adapt to uncertain renewable generation forecasts.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.