Leveraging Data Analytics Towards Activity-based Energy Efficiency in Households

Aiming for sustainable development means reconsidering the access to energy sources in industrialized countries, which are not faced with contingency scenarios that are implemented in emergent and newly-developed countries, to allow equal access to energy sources for all and thwart environmental degradation. The global penetration of renewable energy sources to replace fossil fuel and nuclear power plants means adjusting to stochastic energy production. The expected yield will be dependent on very different weather and landscape conditions and will represent a challenge for countries with continuous access to energy sources and where energy is often considered a public utility. Tracking wastage, improving the scheduling, and the processes that consume energy, would allow us to match the demand and the supply of energy. This will be particularly crucial during peak time, where meeting the high demand incurs the ramping up of mostly unclean additional power plants or introducing power system instability. The digitalization of the energy sector has started with the roll- out of smart meters to record the electricity consumption at a finer granularity and are aimed to replace the biannual or yearly dispatch of utility companies’ employees to read the meter. Considerable research efforts have been directed at analyzing aggregated loads from these smart meters or at developing algorithms for disaggregating households’ total electricity consumption to isolate single appliances’ traces. However, less focus has been set on assessing the potential of using sub-metered data for improving the energy efficiency in households. This was primarily linked to the fact that the necessary datasets were not widely available, due to the difficulty and the costs in instrumenting households for acquiring the consumption data from appliances. The objective of this thesis is to investigate how to leverage and improve existing disaggregated datasets to develop data-driven techniques to improve the energy efficiency within residential homes. Starting from smart meter data, we segmented households into groups with similar electricity consumption pattern based on their peak consumption, to identify hurtful consumption patterns in the perspective of utility companies, for which they could launch targeted mitigation campaigns. However, improving the energy efficiency in the residential sector requires to change individuals’ relationship to- wards their electricity consumption. These behaviors are closely re- lated to the activities that are carried out throughout the day and can be supported by the usage of consumer electronics, such as appliances. Therefore, we turned to analyzing the behaviors inside house- i holds that triggered the usage of electricity by studying a large disaggregated dataset and developed learning techniques to extract activity patterns. We first addressed the challenge of determining when appliances are actively used by households’ residents, from when they are off or idle and incurring standby consumption by developing GMMthresh, an automatic thresholding method, which is agnostic of the appliance’s type, brand and model, but instead relies on the statistical distribution of its power consumption. Due to the lack of event-based and activity labels in existing datasets to allow us to validate our learning technique, we leveraged crowdsourcing concepts to provide an expert-annotated dataset to enrich the existing datasets through our Collaborative Annotation Framework for Energy Datasets (CAFED). We conducted two in-depth studies to quantify the performance of regular users against expert users in labeling energy data on CAFED. We provided analysis tools and methods that can be generalized to crowdsourcing systems for improving the quality of the workers’ contributions. Using the expert-annotated labels, we validated GMMthresh with expert manually labeled data. Then, we developed a method for learning temporal association rules for identifying activities involving the usage of appliances within households. Our pipeline includes our thresholding algorithm and a novel search algorithm for determining time windows for the association rules efficiently and in a data-driven manner. The contributions of this thesis rely on exploiting energy data and developing novel techniques towards identifying activity patterns and their scheduling, which could then become part of an ambient intelligence system that would smarten existing homes. The methods we developed are not restricted to the energy research, as they can be applied to sensor data, where for example inertial sensors also require machine learning algorithms to filter out background noise from actual movement. Similarly, our work on the crowdsourcing of time series opens new perspectives for extending the range of data that can be annotated by the crowd and provides design insights and mitigation techniques for improving the quality of the labeling on collaborative platforms. Finally, our temporal association rules mining framework is not limited to energy time series but can be applied to search for temporal windows and understanding the scheduling of any time series dataset.