Energy Optimization and Climate Control of a NEST Unit at Empa Dübendorf: A Data Predictive Control Approach

The building sector globally accounts for more than one third of the nal energy consumption. In industrialized countries about 50% of the consumed energy in buildings is used for heating, ventilation, and air conditioning. Consequently, increasing the energy e ciency of buildings has a relevant impact on climate change mitigation. Most residential space heating and cooling systems are currently controlled by a simple rule-based controller, also known as on-o or hysteresis controller. Despite its simplicity rule-based controllers are rather energy ine cient. Model predictive control has demonstrated to be an energy e cient approach for building climate control. However, the modelling process for rst-principle based models is assumed to be too expensive in terms of time, money, and expert knowledge for widespread application of model predictive control in residential buildings. As the digital revolution is also taking place in the building domain the number of installed sensors, smart meters, and monitoring systems increases steadily. Thus, the availability of historical data from buildings enables the usage of machine learning algorithms for black-box modelling. Control-oriented machine learning models could potentially allow to reduce the modelling e ort signi cantly. Data-driven models based on random forests combined with linear models have been shown to be a promising approach in simulation studies. However, there is no demonstration of the applicability of data predictive control based on random forests combined with linear models in real-life so far. During this master thesis a data predictive control algorithm based on random forests combined with linear models was implemented and demonstrated for energy optimization and climate control in a residential apartment of the NEST demonstrator building at Empa D ubendorf. The data predictive controller reduced the energy consumption in various cooling experiment by about 30% compared to a hysteresis controller. Moreover, the data predictive controller provides higher thermal comfort by having up to 80% less cumulative constraint violations. In an experimental sensitivity analysis, the data predictive control algorithm showed relative robust performance for di erent control parameters. Furthermore, the performed cost-bene t analysis revealed that data predictive control might have the potential to become a commercial alternative to standard hysteresis controller in single family houses and apartments