Synergy Between Control Theory and Machine Learning for Building Energy Management

According to the International Energy Agency (IEA), energy use accounted 9384 Mtoe in 2015. The building sector is the largest energy-demanding sector, with over one-third of final energy use globally. In addition, Building Energy Management (BEM) allows to unlock flexibility for the grid(s) by Demand Response (DR) strategies. This facilitates the integration of more Renewable Energy Sources (RES), thus achieving lower overall emissions of CO2 in the global energy system. However, most of the existing Building Management Systems (BMS) are based on Rule Based Controllers (RBC). Advanced optimal control techniques like Model Predictive Control (MPC) or Approximate Dynamic Programming (ADP) have proven to outperform RBC in both simulation and real implementation.

This PhD research seeks to develop novel algorithms to increase the share of RES and harness the available flexibility aiming for the broad use of optimal control in buildings. To achieve this main objective, the developed algorithms are innovative and effective in the optimal control sense and, at the same time, prove that they can be applied in reality with minimal amount of engineering effort.

In a first step, a common terminology is unambiguously formulated in a set of definitions. As such inputs and outputs are clearly defined which enhances the reproducibility of results. At the same time the requirements for a benchmark setup to assess optimal control of energy systems in buildings are being formalized. Then, the different ‘modelling’ approaches are compared with respect to control performance for MPC and ADP separately. This comparative study leads to insights in when one method is preferred over the other which allows developing a decision tree to select the most appropriate approach.

An innovative hybrid approach is developed to enrich the data-based methods by models in order to learn better and/or faster, which results in a more performant regression. Moreover, variations within the MPC/ADP formulations are investigated. The impact of these variations on the generic character, scalability, engineering cost and computation cost is investigated. Finally, the algorithms developed in this PhD are validated by simulations as well as against measurement data.