Integrating machine learning for energy management: Applications in power system, building, and agriculture

Abstract

Energy management sits at the nexus of global sustainability objectives, striving not only to curtail greenhouse gas emissions but also to balance resource conservation and economic viability. Against this backdrop, machine learning (ML) has become an increasingly vital tool, offering flexible, data-driven methods for forecasting, optimization, and real-time decision-making in complex energy environments. By leveraging high-resolution data—from solar and wind generation to occupant behavior and crop conditions—these approaches uncover patterns and adapt strategies in real-time to enhance both operational efficiency and system resilience. This dissertation explores the application of tailored ML solutions across three critical areas. 1. For island power systems with high renewable penetration, it demonstrates how data-driven frequency nadir constraints can bolster grid stability without imposing excessive costs. 2. For occupant-centric building controls, it illustrates how multi-agent reinforcement learning can considerably reduce energy use while preserving occupant comfort. 3. Within agricultural operations, it highlights how advanced reinforcement learning frameworks can align irrigation, fertilization, and renewable resource utilization to increase yields while limiting environmental impact. Taken together, these studies exemplify the transformative role ML can play in shaping a more sustainable and cost-effective energy future. The first section of this dissertation focuses on island or weakly interconnected power grids, which face unique challenges due to high penetrations of inverter-based renewable energy resources. As intermittent sources like photovoltaics reduce overall system inertia, operators must ensure stable frequency responses following severe disturbances such as generation outages. To this end, a data-driven unit commitment (UC) model is proposed, incorporating frequency nadir constraints derived from comprehensive dynamic simulations and year-long generation data. By capturing the relationship between scheduled generation and post-disturbance frequency levels, these constraints effectively limit frequency deviations to acceptable ranges, thereby improving system reliability. Extensive simulations reveal that this approach yields significantly more robust frequency nadirs than a simpler minimum inertia constraint, with only a marginal increase in generation costs. The result is a more sustainable, renewable-powered grid that maintains stability and reliability without compromising economic feasibility. Next, the dissertation addresses energy efficiency in buildings, where HVAC (Heating, Ventilation, and Air Conditioning) systems account for a considerable share of total electricity usage. Traditional control methodologies often rely on static occupant behavior assumptions, which fail to capture short-term fluctuations in clothing insulation, metabolic rates, and occupancy patterns. To overcome these limitations, a multi-agent deep reinforcement learning (MADRL) framework is introduced for multi-zone HVAC control. In this setup, each zone is managed by an intelligent agent that learns to dynamically adjust heating and cooling setpoints based on real-time occupant behavior and energy cost signals. Simulation studies demonstrate that the proposed occupant-centric approach reduces electricity expenses significantly compared to rule-based methods and by slightly saving relative to single-agent deep reinforcement learning (DRL), all while preserving occupant comfort. Such improvements underscore the role of advanced learning algorithms in achieving energy efficiency and occupant satisfaction simultaneously, thereby contributing to the broader sustainability agenda. The final section explores the Food-Energy-Water (FEW) nexus, where agriculture intersects with renewable energy utilization. A novel framework is developed to integrate solar power, green ammonia production, and deep reinforcement learning–based optimization. By producing ammonia on-site using renewable energy, farms can leverage a valuable agricultural input while also storing excess energy for later use. The intelligent control system, driven by DRL, coordinates the timing of ammonia production, energy storage, and irrigation schedules to maximize farm revenue and optimize resource use. Simulation results confirm that this integrated approach can significantly enhance both economic and environmental outcomes, paving the way for greener, more resilient agricultural practices. Collectively, these three ML-driven frameworks showcase the versatility and impact of data-based decision-making in advancing sustainability and energy efficiency. Whether in stabilizing low-inertia power grids, managing building energy consumption with occupant-centric insights, or aligning agricultural operations with renewable resources, machine learning stands out as a transformative technology. By rigorously addressing each domain’s distinct challenges and synthesizing overarching lessons learned, this dissertation provides a blueprint for policymakers, engineers, and researchers aiming to foster sustainable energy systems. Ultimately, the proposed methodologies demonstrate that targeted applications of ML can lead to robust, cost-effective, and resource-conscious energy management solutions, propelling global efforts toward a more efficient and low-carbon future.