Data-driven energy management system for smart buildings

• Main Author: Ding, Hongyuan
• Other Authors: Xu Yan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/182526
• Institution: Nanyang Technological University

This thesis explores innovative data-driven approaches to enhance energy management in smart buildings, addressing challenges through advanced deep reinforcement learning (DRL) methods. The thesis follows a logical progression, starting from the smallest unit of a smart building, a multi-energy smart home, and then moving to the residential building level, focusing on a typical scenario of an apartment building often found in city areas. Finally, the thesis addresses a general case of building energy management, proposing an improved approach that integrates advanced forecasting algorithms to enhance the performance of building energy management systems (BEMSs). Chapter 2 models a multi-energy smart home equipped with various appliances, battery energy storage systems (BESS), thermal energy storage systems (TES), micro combined heat and power systems (mCHP), electrical heat pumps (EHP), rooftop photovoltaic (PV), and electric vehicles (EV). The home energy management (HEM) problem is formulated as a cost minimization task with strict constraints to manage energy generation, storage, and consumption efficiently. The chapter introduces a safe DRL approach, primal-dual deep deterministic policy gradient (PD-DDPG) algorithm. Unlike existing DRL methods, this approach learns to minimize costs from accumulated cost functions and automatically tunes the cost function coefficients. Additionally, a convolutional neural network-long short-term memory (CNN-LSTM)-based dynamic electricity price forecasting model addresses future price uncertainties. Simulations with Singapore wholesale electricity price data demonstrate the proposed method's effectiveness in minimizing energy costs and avoiding constraint violations. This chapter addresses the research gap of efficiently managing time-coupled operational constraints in smart homes by proposing a novel approach that ensures constraint satisfaction and convergence, surpassing the limitations of empirical penalty functions used in previous studies. Chapter 3 explores the concept of smart sustainable apartment buildings (SSABs), focusing on installing renewable energy systems in common areas, integrating smart meters for precise energy consumption tracking, and coordinating energy use across multiple residential units. The internal energy coordination scheme is formulated as a hierarchical partially observable Markov decision process (HPOMDP), emphasizing efficient RES utilization and introducing an equitable internal electricity settlement mechanism to ensure fair benefits for all residents. The self-attention prioritized deep deterministic policy gradient (SAP-DDPG) algorithm addresses complex challenges related to RES integration, privacy considerations, and energy consumption patterns within SSABs. Real-world data simulations show that SAP-DDPG significantly enhances energy efficiency, provides cost savings for residents, increases revenue for building operators, and boosts local RES generation utilization. This chapter addresses the unique challenges of energy management in SSABs, including hierarchical coordination, privacy preservation, and efficient internal electricity settlement schemes. The chapter also proposes the use of multi-agent DRL to effectively manage interactions and energy distribution among multiple agents within the SSAB framework. Chapter 4 introduces the multi-step interval forecast soft actor-critic (MSIF-SAC) algorithm. This approach incorporates multi-step interval forecasts of PV generation and electricity prices into the soft actor-critic (SAC) algorithm's state representation. By redesigning the actor and critic network architectures, MSIF-SAC effectively extracts vital information from complex input data, enabling BEMSs to make informed decisions. Real-world dataset evaluations show that MSIF-SAC outperforms existing DRL-based strategies, demonstrating superior learning efficiency, stability, and cost-reduction capabilities. This chapter addresses the research gap of integrating multi-step interval forecasts into DRL-based BEMSs to improve decision-making under uncertainty, especially with the challenges posed by uncertainties in PV generation and fluctuating electricity prices. In conclusion, this thesis presents advanced data-driven energy management strategies for smart buildings that aim to enhance efficiency, reduce costs, and maximize renewable energy utilization. By pushing the boundaries of deep reinforcement learning applications in smart building energy management, this thesis makes significant contributions to the development of more sustainable, cost-effective, and intelligent energy management systems.