Intelligent Data Analysis for Energy Management

Abstract

Predictive data analysis has been identified as essential to support intelligent energy management for better energy sustainability and efficiency. Previous studies have showcased that predicted energy information can benefit consumers economically by optimising energy usage while assisting energy suppliers in efficiently planning power distribution and implementing DR energy management. Recent advances in the Internet of Things (IoT) and Information and Communication Technologies (ICT) simplify the collection of desired energy data streams for further informatics analysis. With such energy data, machine learning (ML) prevails to effectively infer future knowledge associated with online energy resource scheduling, e.g., renewable energy generation, load demands and electricity prices. Although some early efforts have been dedicated to incorporating ML into energy management, computation resource limitations and data scarcity are two pressing challenges for on-site predictive energy analysis. Due to privacy concerns, users prefer on-premise model establishment instead of placing the training task in the cloud and sharing sensitive energy data. But most ML algorithms rely heavily on solid computational resources and vast amounts of labelled data to succeed. Users are often unable to fulfil the requirements in real-world scenarios. To this end, this thesis uses different perspectives to propose several affordable solutions for performing on-demand intelligent data analysis on local resource-constrained devices. Also, three algorithm-specific training frameworks have been developed to solve data shortage by leveraging easily obtainable but extensive data sources based on transfer learning and federated learning. We implement our design under practical settings for photovoltaic (PV) power prediction and non-intrusive load monitoring (NILM) as case studies to fully evaluate their performances.