Machine Learning-Based Infrastructure for Energy Management in Smart Energy Buildings

Abstract

The energy management in smart energy buildings has become the major issue because the buildings account for a significant portion of global energy consumption and the amount is expected to continuously increase. However, managing the energy of the buildings in a smart grid environment is a very challenging task given the informational unknowns about factors that change over time and are difficult to predict. Therefore, many researchers have been interested in the energy management of buildings in order to obtain better solutions under the stochastic environment. In particular, machine learning (ML)-based methods have received significant attention as a way to overcome the difficulties in dealing with the informational unknowns. The significant advantage of application of ML-based methods to the energy management system in the smart energy building is that once a single-point system-based ML infrastructure is built into the building energy management system, any ML-based algorithm with various kinds of goals can be applied to this infrastructure and executed in order to manage the energy of the entire building. Motivated by these advantages of ML, this dissertation proposes ML-based energy management methods for the smart energy buildings including a reinforcement learning-based energy management algorithm for reducing the energy cost of the smart energy building and a kernel lifelong learning-based energy management algorithm for building demand response (DR).

The first part of this dissertation considers an energy management system for a smart energy building connected to an external grid (utility) as well as distributed energy resources including a renewable energy source, energy storage system, and vehicle-to-grid station. First, the energy management system is modeled using a Markov decision process that completely describes the state, action, transition probability, and rewards of the system. Subsequently, a reinforcement-learning-based energy management algorithm is proposed to reduce the operation energy costs of the target smart energy building by managing the energy storage system under unknown future information. The results of numerical simulation based on the data measured in real environments show that the proposed energy management algorithm gradually reduces energy costs via learning processes compared to other existing approaches such as random and non-learning-based algorithms.

The second part of this dissertation considers an energy management system for adjusting the energy consumption of controllable appliances in each room of the buildings participating in DR. DR contributes to improving grid reliability and energy efficiency in the smart grid by shaping the load over time. As buildings account for a significant portion of global energy demand, effective building energy management is a critical component of the overall DR portfolio. A data-driven approach is proposed, which derives the control policies for lighting and chiller systems of individual spaces in the buildings, designed to minimize the user discomfort caused by the reduction in energy consumption while respecting the specified DR requests. The stochasticity of the environmental variables such as solar illuminance and ambient temperature is taken into account. The key contribution is three-fold. First, the control policies are learned over time based on streaming data. Second, the structural similarity of the policies across the rooms is exploited in a multi-task learning framework. To this end, a recently developed kernel-based efficient lifelong learning algorithm is adapted to tackle the problem. Third, a dual decomposition method is exploited to relax the coupled constraint among multiple tasks through splitting the overall problem into several independent subproblems. The numerical evaluation based on semi-real data examines the efficacy of the proposed approach.

Through the studies in this dissertation, it is expected that the proposed ML-based energy management systems can greatly contribute to developing the infrastructure for energy management in the smart energy buildings in real-world environments, where more complicated structures and more types of informational unknowns are to be taken into account. Concluding remarks provide the aspects of contribution and significance of this dissertation including the applicability and extendability to the research field of smart living environments, and suggest future directions with discussion.