Utility learning model predictive control for personal electric loads

Abstract

A personalized control framework that tightly combines online learning of the energy consumer's utility function and the control of the consumer's electric loads according to real-time updates of the utility is proposed. This framework is particularly useful to automatically customize the controller of electric loads that directly affect the consumer's comfort. Because the utility function is identified and predicted online using Gaussian process regression, the controller is capable of immediately setting its objective function to the learned utility function and of adjusting its control action to maximize the new objective. Furthermore, no separate training period to learn the consumer's utility is needed. The performance of the proposed method is demonstrated by the application to a personalized thermostat controlling indoor temperature.