Scalable local short-term energy consumption forecasting

Abstract

Smart meter adoption rates are increasing globally and this has contributed to a rapid increase in the type and volume of data: communication, storage, and processing. These recent advances have created new opportunities for smart grid research, particularly in developing effective methods for processing big data. As power system industry moves towards adapting and implementing smart grid functions, energy demands forecasting is mandated at the distribution level to ensure the balance between energy supply and demand. Unlike system-level forecasting, short term energy demand forecasting at the distribution level needs to be highly scalable, due to the needs for collecting and processing energy demand data for a significant number of loads over a short time. This scalability requirement is magnified if the distribution level forecasting is to be performed centrally where system-level forecasting is being performed. In order to address these challenges, this thesis conducts a systematic study of the scalability and performance of time series forecasting techniques on smart meter data for distribution level short-term energy consumption. The conducted study is based on strategies to parallelize standard and online forecasting algorithms. The developed strategies are converted into algorithms to be implemented for performance evaluation. The performance of these algorithms is evaluated using data collected from several loads during different seasons. Test results demonstrate the challenges of including seasonality terms, and model training when using ARIMA based times series forecasting. Additional results show that the online algorithm achieves better scalability and shorter execution times when compared to the standard ARIMA implementation.