An Application of Machine Learning for a Smart Grid Resource Allocation Problem

The ability to predict aggregator profits is important in the design of an efficient aggregator-based residential demand response (DR) system. In this paper, supervised machine learning models are designed based on historical data to investigate the influence of customer schedulable loads and the forecasted daily electricity price profile on aggregator profits. The k-nearest neighbors (ICNN) and Gaussian process regression (GPR) are chosen because of their consistent performance and high accuracy compared to other machine learning (ML) classification and regression algorithms. Our study demonstrates the classification model is an effective approach to identify the set of schedulable loads that may yield high aggregator profits and the regression model may enable awareness of day-ahead aggregator profits.