Understanding residential electricity consumption patterns and forecasting trends based using machine learning and optimization

Energy is the most important factor driving the progress and development of modern society since Industrial Revolution, and electricity is closely related with human activities. As the global greenhouse gas content continues to increase, reducing carbon emissions has become an important theme in the current global development. Based on this, The Intergovernmental Panel on Climate Change (IPCC) has set different pathways that limit global warming to 1.5 °C and some specific emission standards in different countries. Fossil fuels (coal, oil and gas) are by far the largest contributor to global climate change, while traditional thermal power generation is an important part of fossil fuels, so reducing non-essential electricity consumption can help to reduce fossil fuel emissions from the perspective of demand-side management. To achieve this goal, it is necessary to analyze people's electricity consumption patterns, the social-economical characteristics of electricity end users, and influencing factors that affect electricity consumption. Therefore, this thesis focuses on three aspects of electricity consumption. First, we focus on developing novel machine learning-based approaches for characterizing residential electricity consumption patterns and predicting electricity consumption pattern based on household characteristics. In addition, natural disasters are also important factors that lead to changes in electricity usage, such as earthquake, tsunami and pandemic. COVID-19 has caused a huge impact on people's lives and therefore lead to significant changes in electricity consumption. Second, we study the effect of COVID-19 on electricity consumption, and predict daily consumption over 1-6 weeks based on historical consumption data and population flow indicators in seven America cities by establishing a set of Bayesian structured time series models. Third, we study personalized electricity customer segmentation problem. Identifying different group of customers with both similar electricity usage and social-economical background is key to implement demand-side management programs. We proposed a two-stage constrained clustering framework based on mixed integer linear programming and clustering algorithm, Lagrangian relaxation procedure are proposed to efficiently solve large problem instances. A realistic case studies are analyzed to demonstrate how proposed framework can be used to support personalized customer segmentation problem.