Towards Fast, Nonintrusive Data-Driven Modeling and Control for Demand Response Applications of Residential Heating Systems

With an increasing amount of renewable energy capacity coming online both at industrial and residential scales, the characteristics of the electricity system are changing. The intermittency of such plants increases the stress on the electricity grid and therefore an increased amount of system level services is required to guarantee safe and cost-effective operation, such as demand response. Demand response aims to alleviate imbalances by aligning the demand for electricity with the supply. It counters the variability of renewable energy sources by managing demand for electricity in a responsive manner. This thesis focuses on demand response in the context of residential heating systems, mainly electric heat pumps. There is a large body of work on modeling and controlling such systems to unlock the flexibility potential in their operation. This work aims to deliver fast, nonintrusive and cost-effective algorithms through data-driven modeling and control. The first part of this thesis addresses the demand response problem through an approximate dynamic programming algorithm. It is shown that utilizing different data features can have a significant impact on the performance of such algorithms. Data complexity can be high, with the algorithm requiring 20-25 days of data on the system to obtain energy shifting behavior. The results show that this can be reduced by utilizing a model predictive controller to do the initial operation during the first days to obtain training data. The main issue is shown to be the tractability, with training time being around 30 minutes to obtain a single feedback controller. With the goal of reducing computation time, the second part of this thesis addresses tractable linear models of residential heating systems and how to utilize them within a control context for demand response objectives. Two linear data-driven modeling approaches are proposed to model heating system dynamics, both based on the Koopman modeling framework. They are found to have comparable prediction accuracy to linear greybox models. In particular kernel extended dynamic mode decomposition and its application to heating systems is found to be capable of consistent prediction accuracy, even over longer horizons. The results show mean prediction errors of 0.4-0.5 degrees on zone temperatures. To perform optimal control within the demand response context, the model predictive control framework was used to optimize costs while maintaining comfort levels. The contro