Multi-Seasonality Efficient Residential Load Forecast using Decomposition Algorithm

Home Energy management system using forecasting based energy scheduling strategies require accurate as well as computationally efficient forecasting models for scalable adoption of the technology. Though several decomposition forecasting techniques are prevalent in existing literatures, they are usually computationally inefficient or inaccurate. Time series decomposition to its constituent parts is a crucial process that aids in time series comprehension and can improve forecasting. Increasing adoption of sensors and data storage devices has made possible availability of time series with higher sampling rates (minutely, hourly, and daily). Higher-frequency time series data may show more complex characteristics, like numerous seasonal cycles and non-integer seasonality, than typical time series data. Time series data that can display several seasonal trends can be found in many real-world datasets. This study forecasts the electric load for a residential apartment using the Multiple Seasonal-Trend decomposition technique (MSTL). In our evaluation on real-world time series dataset, MSTL demonstrates computationally efficient forecast with competitive accuracy.