Enhancing web sustainability: Carbon emission evaluation and LLM-powered reduction techniques
Loading...
Supplementary material
Other Title
Authors
Zhu, Ziying
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2025
Supervisors
Liu, William
Song, Lei
Song, Lei
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
websites
emission reduction pathways
greenhouse gas (GHG) emissions
large language models
environmental impact analysis
emission reduction pathways
greenhouse gas (GHG) emissions
large language models
environmental impact analysis
ANZSRC Field of Research Code (2020)
Citation
Zhu, Z. (2025). Enhancing web sustainability: Carbon emission evaluation and LLM-powered reduction techniques (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
https://hdl.handle.net/10652/6954
Abstract
RESEARCH OBJECTIVES
• To propose and validate a new carbon emission model (NWCE) and validate its core assumptions through experiments.
• To evaluate the effectiveness of neural image compression in reducing website data size and carbon emissions.
• To investigate the feasibility of fine-tuning a lightweight Large Language Model for generating targeted web sustainability suggestions.
• To assess the accuracy and safety of an instruction-tuned LLM for automated code refactoring, such as removing redundant CSS.
• To compare the impact of optimising different web assets (code versus images) on both carbon reduction and web performance scores.
• To design and evaluate an integrated, automated pipeline that combines multiple LLM optimisation techniques to measure their cumulative impact.
ABSTRACT
As digitalisation progresses, the ICT sector has become a growing source of carbon emissions, with websites playing a critical role in this trend due to their function as gateways to digital content and services. This study proposes a new website carbon emission evaluation model (NWCE), which incorporates user interaction time and device power consumption parameters to address the limitations of traditional models that overly rely on data volume. The study further investigates the feasibility and effectiveness of AI-based webpage optimisation techniques. Based on experiments conducted on 30 static websites, the CompressAI neural image compression method significantly reduced carbon emissions, achieving a maximum reduction of 49.86% on image-intensive websites. In applying lightweight LLMs, the TinyLLaMA model is first examined for its ability to generate carbon reduction suggestions. The results indicate no statistically significant difference in the quality of the suggestions before and after fine-tuning. Additionally, the DeepSeek-Coder 6.7B model is assessed for its effectiveness in locally identifying and removing redundant code. Findings suggest that the model performs reliably when code redundancy is minimal. Finally, an integrated optimisation framework based on LLM API is also developed and applied to the website dataset. The results demonstrate an average carbon emission reduction of 7.25%, with a maximum of 51.48%, along with consistent improvements in performance scores across all tested websites. These findings underscore the potential of AI-driven solutions in promoting scalable web sustainability and offer a practical pathway toward reducing the carbon footprint of digital platforms.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved