Energy in Communication, Information, and Cyber-physical Systems (E6)

Invited keynote by Costas A. Courcoubetis

Athens University of Economics and Business, Greece

Title: A price-based load shifting approach for demand/response

Our DR model is centralized with the planner exerting central control over all the individual users, taking into account the specific demand structure of each user type. The DR signals can be specialized to each user type or even differ at some finer granularity. The basis to obtain such signals is a model about quantified user preference for running each of its electrical appliances (if these can be shifted in time). This essentially creates for each user type a micro-economic model for its household, in which for each device there is a specific utility for it being scheduled at a particular time and power level. In order to optimize electricity consumption for the next time frame (say, which consists of 24 one hour time slots), the central planner solves centrally the corresponding optimization problem (to maximize the social welfare) for the predicted electricity costs and load constraints, and derives an optimal schedule for the user devices of each user type. He then generates appropriate DR signals as personalized prices for each user type (or, equivalently, as suggestions for device load shifting). These prices have the property that they are incentive compatible, i.e., each when user optimizes the scheduling of his devices this also achieves the optimal schedule that is desired by the central planner.

There are many issues that need to be investigated in this context. These include the complexity of the computations, the simplification of more accurate user models, the randomness of some important parameters, such as: the context of the specific user (when this cannot be derived by our sensors), the scalability of DR construction, the accuracy of the user models, and the capability to improve them using feedback from the observed reactions of the users to the DR signals. We discuss these issues in more detail within the context of a simple model which we have developed for this purpose. We also discuss some initial research on the issue of 'optimal probing'. By probing we refer to the exposure of a user to a specific DR signal and using his response to reduce our uncertainty about his type. Optimal probing refers to the construction of the minimum number of such probes which are needed to reveal the exact type of a user, if randomly selected from a set of possible user types.