Privacy-preserving and Learning-efficient Energy Data Analysis for Advanced Metering Infrastructure

Abstract

The deployment of Advanced Metering Infrastructure (AMI) enables the collection of fine-grained, real-time energy consumption data, supporting demand-side management applications such as residential energy management, electricity pricing, and microgrid operation. Accurate understanding of load patterns is essential for these applications, making short-term load forecasting and load super-resolution two fundamental analytical tasks in smart grid research.

Although deep learning methods have achieved state-of-the-art performance in both tasks, their practical deployment remains challenging. Individual households often lack sufficient training data, centralized data sharing raises privacy concerns, and existing models frequently exhibit poor generalization. In particular, most load super-resolution methods require high-resolution training data and are restricted to fixed resolution scaling factors.

This thesis aims to improve the deployability of deep learning techniques for AMI data analytics. First, a privacy-preserving short-term load forecasting framework is proposed, allowing residential users to collaboratively train models through decentralized clustering and federated learning without sharing raw consumption data, while an asynchronous communication mechanism enhances robustness. Second, an unsupervised load super-resolution framework is developed that eliminates the reliance on high-resolution training data by training solely on low-resolution load profiles and incorporating adaptive noise identification. Finally, generalization is improved through a meta-learning-based forecasting framework for rapid adaptation to new users with limited data, and a flexible load super-resolution method capable of reconstructing load profiles at arbitrary temporal resolutions from a single training instance. Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed approaches.