A Transformer-Driven Framework for Accurate CO₂ Emission Modelling and Forecasting

Musunuru, Lokesh (2025) A Transformer-Driven Framework for Accurate CO₂ Emission Modelling and Forecasting. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (894kB) | Preview

Abstract

Accurately forecasting CO₂ emissions has become essential as global climate challenges intensify and transportation-related emissions continue to rise. Traditional prediction approaches struggle to model the complex, nonlinear relationships among vehicle characteristics, fuel consumption patterns, and emission behaviour, motivating the development of more robust data-driven methods. Although machine learning, deep learning, and hybrid AI techniques have shown substantial progress, many existing models face key limitations, including weak generalizability, difficulty handling high-dimensional tabular data, and limited capability in capturing long-range feature interactions. This study introduces a transformer-based framework using the FT-Transformer architecture to address these shortcomings and advance CO₂ emission forecasting. First, we construct a comprehensive modelling pipeline integrating data preprocessing, feature embedding, and attention-based learning optimized for mixed-type vehicle datasets. Second, we implement and evaluate the FT-Transformer alongside several baseline models including machine learning and deep learning to rigorously assess performance improvements in predictive accuracy, scalability, and robustness. The architecture leverages numerical and categorical embeddings, multi-head self-attention, and layer normalization to learn complex inter-feature dependencies with minimal manual feature engineering. Experimental results demonstrate that the FT-Transformer achieves the strongest performance among all models tested, delivering an MAE of 0.0219, MSE of 0.0010, RMSE of 0.0312, and an R² of 0.9512. These findings confirm its superior capability to capture nonlinear feature relationships and emission determinants more effectively than tree-based, linear, and neural baselines. The model’s attention mechanisms also provide enhanced, enabling clearer identification of influential variables such as engine size, fuel type, and vehicle class.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Tomer, Vikas UNSPECIFIED
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences G Geography. Anthropology. Recreation > GE Environmental Sciences > Environmental protection Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning G Geography. Anthropology. Recreation > GF Human ecology. Anthropogeography > Sustainability
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 10:55
Last Modified:	08 Sep 2026 10:55
URI:	https://norma.ncirl.ie/id/eprint/9889

Actions (login required)

View Item