Managing load contract restrictions with online learning

Abstract

Demand Response (DR) is an effective means of providing flexibility in power systems facing increased variability from renewables. Aggregators must dispatch loads for demand response which provide the most useful services while respecting each load's constraints. In this work, we propose an online learning model where a DR aggregator has to manage a portfolio of curtailable loads subject to several types of restrictions, such as the number of times each load may be curtailed and the total budget. We address this problem with the recent bandits with knapsacks framework. We test the algorithm on numerical examples and discuss the resulting behavior of the algorithm.