Inferring the behavior of distributed energy resources with online learning

In this paper, we apply an emerging method, online learning with dynamics, to deduce properties of distributed energy resources (DERs) from coarse measurements, e.g., measurements taken at distribution substations, rather than household-level measurements. Reduced sensing requirements can lower infrastructure costs associated with reliably incorporating DERs into the distribution network. We specifically investigate whether dynamic mirror descent (DMD), an online learning algorithm, can determine the real-time controllable demand served by a distribution feeder using feeder-level active power demand measurements. In our scenario, DMD incorporates various controllable demand and uncontrollable demand models to generate real-time controllable demand estimates. In a realistic scenario, these estimates have an RMS error of 8.34% of the average controllable demand, which improves to 5.53% by incorporating more accurate models. We propose topics for additional work in modeling, system identification, and the DMD algorithm itself that could improve the RMS errors.