Multi-objective building system control optimization using machine-learning-based techniques

Abstract

The operation of buildings is responsible for 30% of the world's total energy use and 26% of global energy-related emissions. Proper control of buildings is thus economically, socially, and environmentally crucial for reducing energy consumption. However, most Building Management Systems (BMS) still operate on traditional principles, lack optimization, and function in isolation. This highlights a critical need for more advanced, adaptable, and interconnected solutions to collectively control buildings. The data-driven methods in building control show promise for scalability and transferability, offering the potential to eliminate the time and effort needed to create traditional physics-based models.

This dissertation introduces a data-driven building control framework that adapts to changing environments, coordinates multiple building systems, and balances various optimization objectives utilizing both model-based predictive control and model-free reinforcement learning (RL) control methods. The research first investigates the multi-objective smart control of nonlinear dynamic systems: specifically focusing on natural ventilation. The Ensembled Multi-time scale deep-learning-based Adaptive Model Predictive Control (EMA-MPC) system is proposed. This innovative algorithm aims to optimize thermal comfort, better indoor air quality and energy efficiency by controlling automated windows in a naturally ventilated room during winter. The EMA-MPC system demonstrates better performance as compared to basic-MPC, enhanced-MPC and baseline rule-based control. Additionally, the proposed EMA-MPC system reduces modeling efforts and provides an effective approach towards reliable use of machine learning models in smart building control. Building on the model predictive control, the research further explores model-free approach. A multi-agent RL control algorithm is proposed to tackle the challenges of coordinating control systems with diverse response times. Specifically, the research examines the coordinated optimal control for delayed/slow response radiant floor cooling and fast-response window systems in summer period. The proposed RL algorithm illustrates also better performance compared to the rule-based control in ensuring thermal comfort, maintaining indoor air quality, and minimizing cooling energy consumption. Throughout the dissertation, both the EMA-MPC and RL control algorithms are comprehensively designed, constructed, and assessed in virtual testbed and applied in real building for physical experiment, demonstrating their effectiveness and significant promise for future autonomous building applications.