Reinforcement learning for residential energy management system control
Lissa, Paulo Ricardo
Lissa, Paulo Ricardo
Loading...
Publication Date
2021-12-07
Type
Thesis
Downloads
Citation
Abstract
Modern solutions for residential energy management systems control are emerging and helping to improve the way energy is consumed, bringing enhanced comfort and savings to end-users through smart and autonomous actions. This has been facilitated by the introduction of new Internet-of-Things devices for home applications and the advances of the communication infrastructure. In addition to that, the increasing necessity of reducing greenhouse emissions has leveraged the use of renewable sources, hence on-site energy generation such as PV is becoming a reality in a number of buildings. However, this creates a more complex environment and extracting the best it can offer is not an easy task for the operator. For instance, users may not know the best time to turn on or off some specific load ac- cording to the cost of energy, or sometimes not be at home to get benefits from the PV generation. One of the possible solutions is the utilisation of machine learning-based adaptative methods, such as reinforcement learn- ing, which can learn how to operate the system in a near-optimal manner by directly interacting with the environment. This thesis aims to investigate and demonstrate how reinforcement learning can improve residential energy management systems control. Several simulated experiments are conducted based on real demand response projects, applying multiple reinforcement learning algorithms, such as Q-learning and deep Q-networks, thus proving the efficacy of such techniques in this domain.
Publisher
NUI Galway