Physics-Enhanced Machine Learning for Energy Systems

Abstract: Building operations account for a large amount of energy usage and the HVAC (Heating, Ventilation and Air Conditioning) systems are the largest consumer of energy in this sector. To reduce this demand, more energy-efficient control algorithms are implemented and a popular choice for a controller is the model predictive control. However, this demands a precise model. Modeling the indoor climate in buildings is a difficult task since a lot of disturbances affect the process. Some of these disturbances are also unmeasured such as emitted heat from computers and various human patterns. This thesis aims to use data-driven methods to find suitable procedures to model the indoor climate in buildings. This is done in two steps. First, a gray box model is created and its parameters fitted using different data-driven methods. Then, more complex learning-based models are applied and added to the gray box part to catch some of the unmeasured disturbances. Feed-forward neural networks, LSTM networks and ARX models are methods used for this unmeasured disturbance modelling part of the project. The results showed that a gray box model can capture most of the dynamics of the heat flow in buildings, although the obtained parameters and the performance depended a lot of the method used for parameter estimation. Adding a more complexdisturbance part to the gray box model improved the results significantly as it allowed for unmeasured disturbances to be taken into consideration.