LLM-Assisted Carbon-Aware Energy Optimization for IoT Smart Environments

Abstract

The proliferation of Internet of Things (IoT) devices, artificial intelligence (AI), and digital technologies has increased drastically worldwide. These consume electricity, which is produced by both renewable and non-renewable sources, releasing carbon into the atmosphere. As energy demand escalates, mitigating the environmental impact of computing systems presents a critical challenge. Carbon-aware computing offers a promising strategy, called demand shifting, that aligns energy consumption with periods of lower grid carbon intensity, hence reducing emissions. Although forecasting and optimization methods to manage emissions have been studied in places such as data centers, the application of Large Language Models(LLMs) for carbon-aware energy optimization in IoT environments remains relatively unexplored. This thesis examines the design and implementation of an LLM-assisted carbon-aware energy optimization framework for IoT smart environments. The study aims to determine whether an LLM-based optimization can enhance load-shifting decisions and improve energy management compared to the traditional rule-based methods, thus reducing emissions. The framework integrates IoT energy consumption data, weather information, electricity costs, carbon-intensity forecasts, statistical and machine-learning-based energy forecasting and optimization techniques to generate carbon-aware scheduling recommendations. The methodology includes data preprocessing, appliance-level energy forecasting, carbon emission and cost estimation, and optimization strategies. Energy consumption was forecasted using a statistical time-series model and machine-learning models with weather-related features such as temperature and humidity. Then, carbon emissions and costs are calculated by combining forecasted energy demand, price data, and carbon intensity data. Finally, two optimization methods were evaluated: a deterministic rule-based load-shifting strategy and an LLM-based optimization strategy through structured prompt engineering. In an experiment, both optimization strategies reduced carbon emissions and electricity cost relative to the baseline consumption. However, the magnitude of the improvement varied across different experiments. In general, the rule-based optimization obtained lower emissions, whereas the LLM-assisted optimization obtained larger cost savings. This finding indicates that rule- based optimization may be suitable for consistent carbon-focused scheduling, whereas LLM-assisted optimization can combine multiple constraints, balancing both environmental and economic objectives simultaneously. Overall, the study contributed to the field of carbon-aware computing, sustainable IoT systems, and AI-assisted energy management by demonstrating the potential of LLMs as decision- making tools for energy optimization. This system can support intelligent, sustainable energy management techniques and provide a scalable model for future IoT-enabled smart environments.