Learning-Based Predictive Control via Real-Time Aggregate Flexibility

Aggregators have emerged as crucial tools for the coordination of distributed, controllable loads. To be used effectively, an aggregator must able to communicate available flexibility loads they control, known aggregate a system operator. However, most existing measures often are slow-timescale estimations and much less attention has been paid real-time between In this paper, we consider solving online optimization in closed-loop present design feedback, termed maximum entropy feedback (MEF). addition deriving analytic properties MEF, combining learning show that it can approximated using reinforcement penalty term novel control algorithm -- penalized predictive (PPC), which modifies vanilla model (MPC). The benefits our scheme (1). Efficient Communication. An operator running PPC does not need know exact states constraints loads, but only MEF. (2). Fast Computation. number variables than MPC formulation. (3). Lower Costs. We under certain regularity assumptions, is optimal. illustrate efficacy dataset from adaptive electric vehicle charging network outperforms classical MPC.