Wietzke T, Graichen K (2026)
Publication Language: English
Publication Type: Journal article
Publication year: 2026
Book Volume: 373
Article Number: 118362
URI: https://www.sciencedirect.com/science/article/pii/S0378778826014222
DOI: 10.1016/j.enbuild.2026.118362
Reducing the energy cost while maintaining comfort in buildings is a heavily researched area. Using more sophisticated control algorithms like model predictive control (MPC) or reinforcement learning (RL) are promising approaches. Unfortunately MPC and RL are rarely adopted in real applications since their initial configuration is cumbersome and complex. Especially model free RL architectures need long interaction times to learn a sufficiently control policy. Model-based RL (MBRL) architectures instead combine MPC and RL to tremendously speed up the learning process while also inheriting the robustness of MPC. The model consists of Gaussian Processes (GP), which are capable to quantify the uncertainty of the model. This work combines MBRL with online learning of the GPs with a maximum datapoint budget, alleviating the need for episodic learning while also incorporating new unseen datapoints into the controller. We show that our online learned agents need only one day worth of initial training data to outperform a rule-based PI-controller for our two benchmark buildings, where one is the multizone office complex air testcase of BOPTEST.
APA:
Wietzke, T., & Graichen, K. (2026). Online Model-Based Reinforcement Learning for Building Energy Systems. Energy and Buildings, 373. https://doi.org/10.1016/j.enbuild.2026.118362
MLA:
Wietzke, Thore, and Knut Graichen. "Online Model-Based Reinforcement Learning for Building Energy Systems." Energy and Buildings 373 (2026).
BibTeX: Download