From theory to practice: towards scalable and smart energy monitoring based on the Internet of Things
The Internet of Things (IoT) is an ecosystem in which physical objects are being connected to the Internet via underlying technologies such as embedded systems, communication technologies, sensor networks, Internet protocols and applications. IoT has played a remarkable role in various fields (e.g., transportation, healthcare, industrial automation, construction, and energy) to improve the quality of our lives. Due to the unprecedented growth of IoT devices, several machine learning techniques have been applied to gain more useful insights and help us make quicker and better decisions. Despite of the benefits of IoT applications, there are many remaining challenges. Particularly, the scalability of IoT system and effective data-driven approaches for IoT applications are two main challenges in the research community and industry.
In this dissertation, we focus on addressing research challenges in energy monitoring system by using the framework of IoT coupled with different machine learning techniques for time series data. We first examine and develop an IoT architecture of smart energy system from hardware design, firmware development, network communication, real-time data platform, and web/mobile applications. The early adoption of this system has been deployed in IoT laboratory at Department of Computing, Macquarie University. Since smart meters only provide an aggregate energy reading, we also develop the smart IoT devices (i.e., smart sockets) to be able to collect data from each electric appliance with a higher sample rate (e.g., less than or within one minute time intervals). The proposed design of our smart sockets is more affordable for energy consumers and it can be integrated into smart homes and smart offices in the future.
However, the most challenging study in energy load monitoring is how to automatically classify appliance label as well as determine energy consumption of each appliance. There are two fundamental approaches in this field, including Non-Intrusive Load Monitoring (NILM) and Intrusive Load Monitoring (ILM). NILM is a group of statistical and machine learning techniques to determine energy consumption of each household appliance based on readings from smart meters while also give useful feedback to energy consumers. ILM is a framework to use smart plug connected directly to each electric appliance to monitor energy consumption. Our research is inspired by ILM and data-driven techniques in order to resolve appliance classification problem in smart energy monitoring. Particularly, we propose a novel hybrid and semi-supervised approaches for automatic appliance identification. The experiments of these models are conducted based on power data collected from our IoT system.
We then address industrial issues in which data is collected from smart meters. More specifically, we study unsupervised learning techniques to determine typical customer load profiles based on real dataset provided by an Australian electricity retailer. By getting more information of energy consumers of the energy retailer, we then perform tree-based models to solve another real-world problem in energy, namely consumer price index classification. The consumer price index, which is a price indicator related to electricity consumption of customers, helps energy retailers make critical decisions on pricing strategy.
This dissertation is our significant attempts to solve practical and real-world problems in research and industry, especially the energy sector, through the advances in IoT technologies and machine learning. Furthermore, we provide the details of how to implement the outcome of our studies in real life, which is often a missing piece in the existing researches so far. At the end of this dissertation, we discuss the future smart grid based on the development of our smart energy monitoring; and the expansion of our IoT systems to smart irrigation for micro garden and smart parking in the context of smart cities.
