Greenhouse gas emissions because of energy usage are among the foremost drivers of anthropogenic climate change. Decarbonization, as a mitigation strategy, therefore focuses on both improved energy efficiency and a cleaner supply mix in the form of renewable energy sources. Focusing on the residential sector, this thesis attempts to tackle both these issues for thermostatically controllable loads in a large scale demonstrator in The Netherlands. The primary research question it attempts to answer is whether it is possible to learn a model that describes the behavio​ur of different energy systems in a data-driven manner. It also investigates using this model to improve energy efficiency, or provide demand response, diagnostic analysis and user engagement, with the goal to decarbonize the built environment.


Wanneer
-

Waar

KU Leuven, Campus Arenberg
Jozef Heuts, Landbouwinstituut Kasteelpark Arenberg 20
3001 Heverlee
België

Historically, models for energy-flexible loads have been constructed by human domain experts: this can often be a costly undertaking. With the proliferation of sensors recording data in real time, it is now possible to automate this process using machine learning and data-driven approaches. However, most data-driven methods require an abundance of collected data, which can be impractical in real world settings. This thesis shows that it is possible to vastly accelerate the data-driven modelling approach using transfer learning, an emerging field in machine learning which draws many parallels to how sentient beings learn. The models in this thesis are learned and evaluated using 61 real world Dutch residential buildings which were recently renovated to net zero energy status. These models were then used for optimal control of hot water systems, where they were able to improve energy efficiency and self-consumption of renewable energy sources without causing discomfort to building occupants in a trial that lasted one year. The use of these learned models to provide other services such as engaging and better informing end users was also considered over the course of this thesis. Finally, some of the barriers preventing such demand side technologies from taking off in developing countries were also considered.

Promoter

Prof. dr. ir. J. Driesen

Members of the jury

Prof. dr. ir. J. Vandewalle

Prof. dr. G. Deconinck

Prof. dr. ir. J. Suykens

Prof. dr. ir. L. Helsen

Prof. dr. ir. D. Ernst

Prof. dr. ir. A. Nowe