Non-Intrusive Performance Analysis of Grid Serving Behind-the-Meter DERs via Generative Adversarial Imitation Learning

Distributed Energy Resources (DERs) that are Behind-the-Meter (BTM) have the potential to provide grid-services but suffer from poor visibility. It is crucial for the DER aggregators that facilitate the participation of BTM-DERs to assess the flexibility for delivering committed grid-services and for incentivizing or penalizing them based on their performance. In this work, we explore the possibility of learning the behavior of the BTM-DERs by the DER aggregator non-intrusively. We demonstrate that given sufficient information to a learning agent at the DER aggregator end on the environment that influence the decision-making of the BTM-DER controller and the actions guided by its preferences, the learning agent can learn its control policies and predict its response prior to sending it the dispatch commands for grid-services. Specifically, we employ Generative Adversarial Imitation Learning (GAIL) algorithm for a continuous state space representing learning environment to learn using a set of demonstrations and predict the BTM-DER response to grid-service request in real-time.