Structure-Aware Machine Learning for National CO Emission Composition: A Classification Framework

CO structure. The sectoral distribution of national CO emissions is a necessary basis for structural mitigation options and transformation pathways. This paper introduces a structure-based machine learning architecture designed to work with patterns of emission composition rather than the absolute emission trend. We construct normalized sectoral vectors that capture the relative weight of power, industry, transport, residential, and aviation, using data from Carbon Monitor (20192025). We perform a PCA to reduce dimensionality and explore the variability of the structure among countries and years. K-Means clustering is subsequently employed to identify archetypal structural types across country-year pairs, and a Support Vector Machine (SVM) classifier is trained to learn the separation boundaries for these types, enabling classification and visual interpretation of the data. We also investigated K-Nearest Neighbors (KNN) for measuring the structural nearest neighbor. Still, KNN was practically useless due to the invariance of the normalized vectors, which often resulted in multiple self-nearest neighbor matches. Yet, the global framework successfully discovers structure-based emission categories and traces structural changes over time. The method can be applied to different scales in space and time and used as a diagnostic tool for monitoring the emission structure, analyzing the type, and designing policy. The combination of unsupervised and supervised learning schemes brings in analytic power as well as the interpretability of a policy.