Generalization in Advanced Building Energy Control through Data-Driven Modeling and Transfer Learning

Buildings are responsible for about 40% of final energy usage and about 36% of greenhouse gas emissions solely in Europe. Reducing the energy consumption of buildings is thus a decisive step to decrease the production of greenhouse gases. Recently, the application of machine learning (ML) methods showed great potential for improving building energy management. Especially advanced control methods such as reinforcement learning (RL) and model predictive control (MPC) captured the attention of the community. However, a widespread deployment of these techniques in the building industry still remains to be seen. 

Data-driven modeling (DDM) could help to accelerate the usage of advanced control in the building industry, yet DDM requires large amounts of data to be accurate. Ongoing research suggest that leveraging transfer learning (TL) can reduce the data intensity of DDM significantly, enabling its widespread usage. 

The PhD thesis aims therefore to improve the methodology of DDM for advanced control in building energy management with emphasis on transfer learning and generalization.  For this, simulations are used to generate synthetic data to pretrain predictive ML-models, which can then be fine-tuned and thus fitted to the target domain (target building). 

Domain randomization (DR) is analyzed as a suitable technique to tackle the sim2real transfer, which describes the transfer problem of ML-models trained on a synthetic domain to the real target domain. With DR, the physical simulations are parameterized and intelligently altered to generate high variance synthetic data for training data-driven building models with higher generalization capabilities. Later, the fine-tuned DDMs may be used as tailored data-driven training environments for RL (surrogate building model).

Further, this work intends to use the parametrized physical simulations to pretrain RL-agents, so that a higher generalization potential is reached. Lastly, the work compares RL-agents pretrained on the physical simulation variations and fine-tuned to the target building with the help of a surrogate building model with other advanced control methods such as MPC to determine their performance, data efficiency and suitable future research paths.