Clustering of Groups of Key Greenhouse Gas-Emitting Countries Based on Economic, Demographic, and Energy Indicators

Introduction. The central issue of this study is that while the global goals for decarbonization have been formulated as general and universal, the actual conditions for achieving them vary significantly between countries. For some countries, the carbon footprint is determined by factors such as the size of their economy and population, while for others it is influenced by a high reliance on coal generation, export-oriented fuel and energy complexes, energy-intensive industries, or insufficient technological modernization. As a result, simply ranking countries based on their GHG emissions does not provide a reliable basis for determining the priorities for climate policy or the direction of decarbonization strategies. The focus of the global discussion on reducing greenhouse gas emissions is shifting from setting common goals to choosing decarbonization trajectories for countries with fundamentally different energy structures. In the long term, the goal of the Paris Agreement is to limit the growth of the global average temperature. The starting point in this case is the pre-industrial level. The goal is formulated as follows: “well below 2°C.” Solution to this problem requires a serious transformation of the global energy system [1].

More than 130 countries have expressed their intention to achieve full carbon neutrality by 2050. These countries have either already committed to ensuring a zero balance of greenhouse gas emissions or are considering achieving zero emissions by the middle of this century [2].

The practical implementation of the adopted strategies depends not only on political commitments, but also on factors such as the energy balance, level of industrialization, structure of final energy consumption, export role of the fuel and energy complex, demographics, and initial carbon intensity of the economy [3].

Modern research shows that the energy sector remains the main source of anthropogenic GHG emissions, and decarbonization scenarios vary significantly in terms of the depth of technological transformation, the cost of transition, and the degree of involvement of low-carbon energy sources [4]. Materials about the problem were published by REA^5,6, RANEPA⁷, IEF RAS⁸, CENEF⁹, Energy Institute^10,11, IEA^12,13.

Scenarios for achieving carbon neutrality, macroeconomic consequences of energy transition, potential of hydrogen energy, the role of nuclear generation, renewable energy sources (RES), energy efficiency, and carbon capture technologies have all been thoroughly discussed in Russian and international literature. Through a comparative analysis of these scenarios based on macroeconomic modelling, we have identified three key pathways.

The first category includes inertial approaches. Experts from the REA, RANEPA and CENEF pointed to unjustified adherence to established models and factors hindering development in the field of decarbonization. The authors noted that infrastructure modernization focuses on fossil fuels. This leads to a moderate change in the structure of primary energy resources, as well as stagnation or an increase in GHG emissions (up to 6.5 gigatons of carbon dioxide equivalent by 2050).

The second one combines moderately optimistic scenarios. The authors of IEF RAS refer to these as a “reasonable” choice [5], and REA as a “rational technological” option. CENEF believes that reaching the so-called “carbon plateau” would be a significant achievement. This refers to the gradual implementation of low-carbon technologies. The potential for a gradual increase in energy efficiency and a partial diversification of the energy mix formed the basis for the goal: by 2050, emissions should be reduced to 1.5–3.5 gigatons of carbon dioxide equivalent, or stabilized at that level.

Finally, the third group proposes radical approaches to solving the problem. They aim to achieve low [6] or zero carbon footprints, acting actively or even aggressively in this direction [5]. The implementation of these solutions requires a deep structural restructuring of the economy, large-scale introduction of renewable energy sources, nuclear generation, hydrogen technologies, and carbon capture and storage systems. By 2050, emissions are expected to be reduced to 0.54–2 gigatons of carbon dioxide equivalent. The total consumption of primary energy resources is forecast to decrease to 12.4 billion tons of oil equivalent. These indicators are included in the “Net Zero” scenario of the REA of the Ministry of Energy of Russia.

In all the scenarios considered, the peak demand for coal, oil, and natural gas is projected to occur in 2030, after which it is expected to decline, especially for coal. Factors critical to the implementation of these scenarios include balanced modernization of energy-intensive industries, the development of regulatory mechanisms, including carbon pricing, and ensuring the economic affordability of low-carbon solutions [7].

The published studies can be broadly divided into three groups. The first group includes work on modeling scenarios for low-carbon development, with estimates of potential trajectories for changes in energy balances, emissions, and consumption patterns for primary energy resources. The second category compares countries and regions based on individual indicators, such as absolute GHG emissions, per capita emissions, the carbon intensity of GDP, the share of fossil fuels, energy intensity of the economy or the pace of renewable energy introduction. The authors of the third group use methods of multidimensional analysis and clustering. However, this work is often limited to specific groups of countries, narrow sets of energy or emissions characteristics, or fails to relate the results to transition risks related to climate change and practical trends towards decarbonization.

The solution in this case is to classify countries according to a set of features reflecting structural changes in the carbon load [8]. Cluster analysis will help overcome the disadvantages of the standard ranking and identify groups of countries that are similar in their carbon footprint, as well as in their economic, demographic, energy, and climate characteristics.

The analysis of scientific literature has shown that the comprehensive classification of GHG-emitting countries is insufficiently developed, and the presented scientific work is intended to fill this gap. The proposed approach will make it possible to determine which decarbonization measures can be replicated within a group of countries, and which require targeted adaptation, taking into account the national structure of the energy balance and transitional climate risks. This is the difference between this concept and others, which are limited to the analysis of individual energy or emission indicators.

Thus, this research aims to form and interpret the typology of key greenhouse gas-emitting countries based on a comprehensive analysis of economic, demographic, energy, and climatic indicators. The resulting classification can be used to identify differences in the prerequisites for decarbonization and transitional climate risks for countries with similar profiles.

To achieve the goal, the following tasks were set:

− formation, preprocessing, and normalization of economic, demographic, energy, and climate indicators for key GHG-emitting countries to ensure cross-country comparability;

− establishing the most informative feature groups and performing clusterization of countries with an assessment of the sustainability of the groups;

− interpretation of the profiles of the formed clusters and identification of priority areas for decarbonization strategies adaptation.

Materials and Methods. The research project consisted of 10 stages.

1. Collection of statistical data for more than 40 key greenhouse gas emitting countries between 2010 and 2022 [9].
2. Selection of countries based on their contribution to global emissions and representativeness of the energy model.
3. Creation of a system of economic, demographic, energy, resource, and climate characteristics.
4. Unification of country names, measurement units, and data time periods.
5. Clearing the data set and excluding indicators that do not provide cross-country comparability.
6. Standardization of quantitative features using z-transformation method.
7. Preliminary visual verification of the data structure with dimensionality reduction using the principal component analysis (PCA) method.
8. Clustering of data using Ward’s method and the combined t-SNE + k-means approach.
9. Selection of the optimal number of clusters using the elbow method and the silhouette coefficient. Evaluation of the clustering quality using the Davies-Bouldin index.
10. Interpretation of clusters taking into account the energy profile, carbon load and transitional climate risks.

The study group included countries that meet one of two or two criteria. The first was the value for global GHG emissions. The second was representativeness in terms of the energy model. In other words, not only large absolute emissions were taken into account, but also a high proportion of fossil fuels, the export orientation of the fuel and energy complex, advanced low-carbon generation, or a significant transformation of the energy balance.

The figures for 2022 were used to ensure comparability. This approach provided the most complete cross-section by country. The time series for 2010–2022 was used to interpret the dynamics and verify the stability of the identified differences.

The assessment was based on a system of indicators for two groups, or blocks, of data.

The first (economic and demographic) group included GDP, population, and the share of urban and rural populations.

The second (energy and resources) group included:

− installed generation capacity by type of sources;

− production and import of primary energy;

− production of heat and electricity by type of generation;

− energy consumption by economic sectors;

− recoverable coal, natural gas and oil;

− total reserves of energy resources.

The third (climate) block combined GHG emissions — general and across sectors of the economy (energy, industrial processes, agriculture, waste).

The final matrix was created based only on the quantitative indicators that ensured comparability between countries. For each feature, units of measurement, ranges of values, and omissions were verified. Indicators for which there was no comparable data for a significant number of countries were only used in the descriptive analysis phase and were not included in the final clustering matrix. Standardization was performed using the z-transformation formula: the average value of a set of countries was subtracted from each attribute value, after which the result was divided by the standard deviation. This allowed us to eliminate the influence of features with large absolute scales, such as GDP, population, and installed generation capacity.

In order to analyze the production and consumption of energy resources, assess the effectiveness of their use and the impact of the emitters on the environment, the formation of cluster groups of key emitting countries by economic, demographic and energy indicators was considered. More than 10 methods of statistical data processing for cluster formation were considered [10]. Table 1 provides information on the most suitable methods.

Cluster analysis involved classifying observational objects (cases) into groups (clusters) based on similarities of their characteristics across a variety of variables [11]. Hierarchical clustering was performed using Ward’s method on a standardized feature matrix using the Euclidean distance. Ward’s method was chosen because it minimized intracluster variance and allowed obtaining an interpretable dendrogram of country similarity. Additionally, the combined t-SNE + k-means approach was used. For the nonlinear mapping of a high-dimensional data structure into a two-dimensional space, t-SNE was used, followed by clustering using the k-means method on the resulting representation.

To ensure the reproducibility of the calculations, the parameters of the algorithms were fixed in the software implementation. The number of clusters was checked in the range k = 2–10 using the elbow method and the silhouette coefficient. Additionally, we assessed the quality of the final solutions using the Davies-Bouldin index. The initial state of the random number generator (random_state = 42) was recorded in the k-means procedures. When selecting the number of clusters, initialization of k-means++ centroids was used. The n_init and max_iter parameters were not set separately and were assumed to be equal to the default values of the used version of the scikit-learn library. For t-SNE, n_components = 2, perplexity = 30, random_state = 42 were set. The parameters learning_rate, max_iter / n_iter, init and metric were not set separately. After constructing the two-dimensional t-SNE representation, clustering was performed using the k-means method at k = 4.