The work in this PhD thesis addresses the practical implications of deploying and testing Non-Intrusive Load Monitoring (NILM) and eco-feedback solutions in real-world scenarios. The contributions to this topic are centered around the design and development of NILM frameworks that have been deployed in the wild, supporting long-term research in ecofeedback and also serving the purpose of producing real-world datasets and furthering the state of the art regarding the performance metrics used to evaluate NILM algorithms. This thesis consists of three main parts: i) the development of tools and datasets for NILM and eco-feedback research, ii) the design, implementation and deployment of NILM and eco-feedback technologies in real world scenarios, and iii) an experimental comparison of performance metrics for event detection and event classification algorithms. In the first part we describe the Energy Monitoring and Disaggregation Data Format (EMD-DF) and the SustData and SustDataED public datasets. In second part we discuss the development and deployment of two hardware and software platforms in real households, to support eco-feedback research. We then report on more than five years of experience in deploying and maintaining such platforms. Our findings suggest that the main practical issues can be divided in two categories, technological (e.g., system installation) and social (e.g., maintaining a steady sample throughout the whole study). In the final part of this thesis we analyze experimentally the behavior of a number of performance metrics for event detection and event classification, identifying clusters and relationships between the different measures. Our results evidence some considerable differences in the behavior of the performance metrics when applied to the different problems.
展开▼
