Enhancing household-level load forecasts using daily load profile clustering

Forecasting the electricity demand for individual households is important for both consumers and utilities due to the increasing decentralized nature of the electricity system. Particularly, utilities often have very little information about their consumers except for aggregate building level loads, without knowledge of interior details about the household appliance sets or occupants. In this paper, we explore the possibility of enhancing the day-ahead load forecasts for hundreds of individual households by clustering their daily load profile history to obtain each consumer's specific typical consumption patterns. The clustering method is based on load profile shape using the Earth Mover's Distance metric to calculate similarity between load profiles. The forecasting methods then predict the next day shape from the empirical probability of previous cluster transitions in the consumer's load history and estimate the magnitude either by using historical load relationships with temperature and forecast temperatures or previous day consumption levels. The generated forecasts are compared to a benchmark Multiple Linear Regression (MLR) day-ahead forecast and persistence forecasts for all individuals. While at the aggregate level the MLR method represents a significant improvement over persistence forecasts, on an individual level we find that the best forecasting model is specific to the individual. In particular, we find that the MLR model produces lower errors when consumers have a consistent daily temperature response and the cluster model with previous day magnitude produces lower errors for consumers whose consumption changes abruptly in magnitude for several days at a time. Our work adds to the state of knowledge surrounding individual household load forecasting and demonstrates the potential for cluster-based methodologies to enhance short term load forecasts.