Machine Learning for Carbon Emissions Prediction and Anomaly Detection: A Comparative Study of Traditional and Deep Learning Approaches
Loading...
Date
2026
Journal Title
Journal ISSN
Volume Title
Publisher
KWASU Journal of Information, Communication and Technology (KJICT.)
Abstract
Accurate greenhouse gas emissions monitoring is critical for evaluating climate policies and ensuring regulatory
compliance. Although machine learning offers promising capabilities, systematic comparisons between tradition
al and deep learning approaches for structured environmental time-series data remain limited. Here, we evaluated
seven prediction models and three anomaly detection approaches on monthly CO₂ emissions data from 20 major
emitting countries between 2000 and 2024. Random Forest achieved superior performance (R² = 0.902) com
pared to all deep learning architectures (R² ≈ 0), while requiring 15–60 times less training time. SHapley Additive
exPlanations revealed that the 12-month lag feature dominated the predictions (75.6% importance), indicating
strong year-over-year persistence with direct policy implications. Deep learning failure stems from insufficient
training data (3,863 samples), structured tabular characteristics with domain-engineered features, and unnecessary
architectural complexity, thereby aligning with emerging evidence that traditional methods often outperform deep
learning on tabular datasets. Anomaly detection proved challenging, with all unsupervised models achieving F1
scores below 0.05, suggesting that supervised approaches with labelled examples are necessary for operational
deployment. We recommend prioritising Random Forest for prediction, investing in domain-informed feature en
gineering, integrating explainability tools for transparency, and deploying anomaly detection with human review.
We identified eight research frontiers, including satellite data integration, federated learning, causal inference, and
hybrid physics-machine learning models. This study demonstrates that traditional machine learning, combined
with domain expertise and interpretability tools, can provide accurate, efficient, and transparent emission predic
tions for operational climate monitoring systems.