Feature Extraction and Data Gap Analysis of US Residential Electricity Consumption

Increasing electricity consumption, especially in residential buildings, poses significant challenges to the management of the electric grid. These challenges can be mitigated by gains in energy efficiency and by the development of intelligent energy management applications to anticipate and prevent peaks. Both energy efficiency and energy management require a deep understanding of detailed individual load levels within a house. In this work, a large real world residential dataset of US residential buildings is developed and used as the foundation for addressing the data gaps that impair current state-of-the-art electricity consumption analysis algorithms. This dataset is the first known database that includes non-intrusive load monitoring (NILM), meaning that individual appliance consumption within each house, including any electric vehicles garaged in that residence is measured. The interactions between electricity consumption, weather and size of houses is investigated and compared to current state-of-the-art methods to demonstrate the performance and more accurate prediction of future usage. The approach extracts important time periodicities, performs a deep data loss analysis and develops a long-short term memory (LSTM) model based forecast for next half hour consumption in a house. The results demonstrate the impact of adding finer level data of NILM information to improve the LSTM forecasts by over 5%.