Demand Response, Model Predictive Control (MPC), Randomized Control, Virtual Energy Storage
We study the problem of coordination of a collection of on/off thermostatically controlled loads (TCLs) to act as a “virtual battery”. Virtual Energy Storage (VES) is provided by the collection by either consuming more (charging) or less (discharging) power than the baseline. VES can be an inexpensive alternative to batteries when a large share of the electricity comes from volatile sources such as solar and wind. The level of charging or discharging is dictated by an exogenous power reference signal provided by a balancing authority. In practice, there are many challenges for successful implementation of a virtual battery from a TCL collection. At the grid level, a complex coordination problem needs to be solved to decide which TCLs to turn on or off at each sample point so to meet the appropriate level of charging or discharging. In addition, exogenous disturbances that affect the system, weather and occupant behavior, may not be accurately known. Grid level challenges are coupled to the local TCL level since each load has associated with it a quality of service (QoS) that can not be violated. A number of solutions have been proposed in the literature to address these challenges. Many of these solutions model the collection as a Markov chain and then try to control the state of the Markov chain by deciding which TCLs to turn on or off. These methods do not guarantee how well consumers’s QoS will be satisfied, or how accurately the grid level reference signal will be tracked. In addition, almost all prior work has assumed that the outside weather - which significantly effects a TCLs behavior - is constant. In this paper we propose an architecture for solving the coordination problem that is based on recent work on randomized control. The architecture consists of a local controller at each TCL, that replaces the traditional thermostat control, and a grid-level controller. This controller is designed to behave identically to a thermostat when a TCLs temperature is outside the deadband, which ensures that the consumer’s QoS is satisfied. However, there is a non-zero probability for a TCL to switch modes when the temperature is within the deadband. This probability is influenced by a scalar variable ζ: positive values increase the chance for TCLs to turn on and vice versa. When ζ is zero this corresponds to a scenario where no discharging or charging is required from the collection. At the grid level MPC (model predictive control) is used to compute the control command ζ. MPC requires a model from input ζ to the output, fraction of loads on, and predictions of disturbances over a planning horizon. As well as measurement of the output at time k. Additionally, irrespective of the choice of control architecture, there is a fundamental limit to the power and energy capacity of the collection of TCLs. We partially address this issue by scaling the reference signal by a function of the outside air temperature.