Deep reinforcement learning for coordinated air-conditioner control in groups of buildings using smart meter data
Résumé
Coordinated control of residential air-conditioning systems is a promising demand response scheme to reduce peak loads and lower energy bills at a district level. Existing schemes have achieved only limited real-world success due to infrastructural requirements and low consumer adoption. In this paper, a scalable method for the coordinated control of air-conditioners in a group of hundreds of buildings is proposed in which consumers are allowed to override control signals based on their local thermal comfort. The control is based on a novel Split-Input Actor-Critic Reinforcement Learning architecture with a neural network that suggests temperature setpoints for each consumer. It is trained on a grey-box model of the system developed using historical smart meter data. The approach notably does not require behind-the-meter inputs during deployment. In a test system, the controller is able to reduce peak loads by up to 20% without increasing the total energy consumption compared to the baseline case with local control only. The robustness of the controller to different reinforcement learning architecture, inputs, model accuracy and stochasticity is studied. Additionally, the learnt policies are visualized, providing insights about the decision-making of the controller.
| Origine | Publication financée par une institution |
|---|---|
| Licence |