Download files
Abstract
This thesis focuses on development of efficient data-driven model-based and model- free solution methodologies as a key feature to address the challenges of building energy management systems (BEMSs). As the major consumers of global elec- tricity, BEMSs have recently attracted significant research interest. Nevertheless, traditional methods suffer from inefficient control strategy with regard to building local and central controllers. Moreover, they rely on assumption of a known model of uncertainties and system characteristics which could be far from reality, and impair the system optimality and security, and consumer’s comfort. Modern technologies have provided a data-rich environment that the traditional methods fail to fully exploit or to adapt. Thus, data-driven methods by use of advanced mathematical- and deep reinforcement learning (DRL)-based methods are required. While virtual energy storage systems have become a viable solution for provision of frequency reg- ulation services (FRS), the optimal capacity of a building for provision of FRS is not modelled. Lastly, available real-time DRL-based methods suffer from computational inefficiency for hyper-parameter tuning, and security issues for constraint violations.
In chapter 2, a novel hierarchically coordinated control methodology is proposed which fully addresses the mutual impact between local and central controllers. In chapter 3, a novel model for optimal capacity of building for provision of FRS is developed, in which safe operation in pre- and post-FRS is considered. In chapter 2 and 3, data-driven distributionally robust optimisation method based on Wasserstein metric and a finite set of data is developed. The proposed data-driven methods tackle various uncertainties by guaranteeing that the out-of-sample performance complies with a pre-defined conservatism level. In chapter 4, an advanced DRL methodology is proposed for day-ahead and real-time BEMSs. Specifically, a multi- agent deep constrained Q-learning algorithm is developed, where safe action spaces of various agents are directly addressed in the Q-update process. The proposed method satisfies the agents’ individual and system-wide constraints, and improves hyper-parameter tuning by avoiding reward shaping technique.
In the comprehensive numerical analysis of all the proposed methods, comparison with various traditional methods is carried out, demonstrating superiority of the proposed methods in terms of computational efficiency and optimal performance.
