Autonomous Household Energy Management Using Deep Reinforcement Learning

摘要

With the massive growth of smart meters and availability of data, there is a unique opportunity to offer novel services to utility customers. The emergence of new technologies can be used to learn a customer's electrical consumption patterns and to offer customized solutions for minimizing electricity cost. One such technology is deep reinforcement learning (DRL), which is explored in this paper. The proposed approach for autonomous household energy management utilizes deep Q-learning with a novel method to deal with potential scalability issues that involves grouping dependent electrical devices. The defined reward function addresses the problems of previous research by adding more scenarios for the penalty and reward. Transfer learning using an adviser agent is also utilized to improve training time. By grouping dependent devices, we are able to resolve the problem of scalability. The results of these experiments demonstrate that DRL can be utilized to autonomously control household energy consumption.