Abstract:

A lack of resources, the increased focus on CO2 emissions, and the resulting climate changes have increased the demand for energy-saving technologies. This demand has driven this study of Reinforcement Learning (RL) based control for Heating Ventilation and Air-Conditioning (HVAC) systems.

Studies of the world’s energy consumption show that 40% of the world’s energy consumption goes into HVAC systems. Studies also show that cost can be reduced by up to 20% of the current industrial practice by using optimal control on HVAC systems. However, to do optimal control of HVAC systems is difficult. Model Predictive Controllers (MPC) require a unique model for a given building, which may or may not be economically feasible, hence the reason why it is not normal practice. On the other hand, there are model-free methods that typically are data-driven. Data-driven methods have previously been shown possible but data expensive.

This dissertation focused on data-driven methods for HVAC, and how to make these methods more robust and data efficient.

In this defense are three methods presented for doing robust and/or faster Reinforcement Learning. These three methods can all be integrated together. The result of these three methods is shown in a field test where the oscillation is reduced by roughly 50 percent when compared to a benchmarking control method. Additionally, is it argued that the cost of heating is reduced by a minimum of 10 percent again when compared to a benchmarking control method.

Assessment Committee:       

Associate professor Peter Nielsen (chair)
Department of Mathematical Sciences
Aalborg University
Denmark

Associate Professor Roel Pieters
Tampere University
Finland

Lars Finn Sloth Larsen, Senior Manager R&D
Danfoss A/S
Denmark

Supervisor:
Associate professor Simon Bøgh
Department of Materials and Production
Aalborg University

Co-supervisors:
Carsten Skovmose Kallesøe
Grundfos A/S

Søren Emil Sørensen
Grundfos A/S

The PhD defense will be hosted by Moderator Peter Nielsen. The lecture constitutes a 45 minutes presentation by Christian Blad followed by a short break and a discussion session with questions from the opponents and the auditorium.

After the defense the department will host a reception.

Online participation via Teams

The PhD defence will be hosted as a hybrid event (physical + online attendance). Thus, you may join the PhD defence online by using the MS Teams link provided below. Note, that only the 45 minutes presentation by Christian Blad will be broadcasted online while the discussion session is reserved for the physical attendees.  

Please click here