Paper Conference

Proceedings of Building Simulation 2021: 17th Conference of IBPSA

     

Comparing Machine Learning based Methods to standard Regression Methods for MPC on a virtual Testbed

Felix Bünning 1,2, Corentin Pfister 1,2, Ahmed Aboudonia 2, Philipp Heer 1, John Lygeros 2
1 Urban Energy Systems Laboratory, Empa, Switzerland
2 Automatic Control Laboratory, ETH Zürich, Switzerland


DOI: https://doi.org/10.26868/25222708.2021.30346
Abstract: Data Predictive Control has emerged as a promising way to control buildings optimally with the help of data-driven models. Besides conventional system identification methods, also Machine Learning based methods can be the basis of such models. While these methods have been validated in individual simulations or experiments, there is a lack of comparability due to changing experimental conditions or mismatch between simulation cases. Here, we present a comparison of three different data enabled building control methods one of the test cases of the virtual building controller testbed BOPTEST: a conventional ARX model with a one-hot encoded solar model, a Random Forest model with linear control inputs, and an Input Convex Neural Network model. Our results suggest that the ARX model outperforms the other models in most of the relevant criteria in the one-zone hydronic test case.
Keywords: Model Predictive Control, Machine Learning, Identification, Data Predictive Control
Pages: 127 - 134
Paper:
bs2021_30346