Correction: Mapping regional disparities in the global shift toward decarbonized power systems

Following publication of the original article [1], the authors reported that coal-fired capacity data for the Chinese mainland were omitted during the data synchronization process, resulting in several text and numerical corrections. The global coal-fired capacity has been revised from 847,000 MW to 1.9 million MW; Asia’s coal-fired capacity has been revised from 384.8 GW to 1,441 GW; the ranking of gas-fired capacity has been revised to second largest, and the ranking of hydropower capacity to third. Following data completion and recalculation, China’s share of global coal-fired capacity has been corrected from 27.5% to 57.4%. Approximately 0.25% of the records in the dataset became text-formatted entries and were not converted back to numeric format during the original Excel-based analysis. This resulted in several updates to global totals reported in Section 4.1, including nuclear capacity revised from 373,000 MW to 380,000 MW, oil-fired capacity from 273,000 MW to 278,092 MW, bioenergy capacity from 136,000 MW to 142,589 MW, and geothermal capacity from 13,070 MW to 13,341 MW. Accordingly, the global share of gas-fired capacity attributed to Texas and Florida has been corrected from 12% to 7%, and Brazil’s share of global bioenergy capacity has been corrected from 15% to 12%. Based on the above corrections, the continental affiliations of several subnational regions were further adjusted, resulting in slight numerical changes in Section 4.4. Asia’s gas-fired capacity has been revised from 626.9 GW to 631.4 GW; Africa’s gas-fired capacity from 107.9 GW to 111.2 GW; Europe’s gas-fired capacity from 314 GW to 324 GW, coal-fired capacity from 129 GW to 124 GW, and bioenergy capacity from 41.3 GW to 47.2 GW; and North America’s gas-fired capacity from 584.9 GW to 586.1 GW, coal-fired capacity from 242 GW to 230.6 GW, and geothermal capacity from 3.99 GW to 4.3 GW. Based on these corrections, the results presented in Table 2 have been recalculated; Fig. 1, Fig. 3, and Fig. 4 have been updated; and the Supplementary dataset has been revised accordingly. Furthermore, following the above data corrections, the ML-based driving-factor analysis in Section 5.2 has been recalculated, and the corresponding analytical results and figures have been updated, including Fig. 5 and Fig. S1. The main conclusions derived from the dataset remain unchanged, although some numerical values and figures must be adjusted after data correction. The correct Table 2, Figs. 1, 3, 4 and 5 have been provided in this Correction. Supplementary Materials 1 and 2 have been updated. The relevant data have been updated in the manuscript as well. The incorrect Table 2 is: Changes in installed capacity by energy type and continent (2015–2020) Asia 27.17% 20.75% −1.15% 77.46% 20.74% 15.01% 89.92% 381.05% 12.32% Africa 11.74% 30.91% 15.23% 7.16% 30.91% 0.00% 79.19% 193.85% 20.52% Europe −8.24% −7.83% −43.82% −0.25% −7.31% −7.09% 40.21% 65.66% 3.72% South America 12.13% 15.73% 4.30% 16.20% 15.73% 24.77% 163.70% 884.16% 12.04% North America −20.64% 1.72% −24.42% −7.23% 1.72% −0.71% 57.56% 219.83% 1.77% Oceania −10.73% 0.56% 9.62% 9.43% 0.56% 0.00% 79.47% 4086.76% 2.96% The correct Table 2 is: (Table presented.) Changes in installed capacity by energy type and continent (2015–2020) Asia 20.64% 21.30% −1.02% 77.22% 36.08% 15.01% 89.91% 377.82% 12.29% Africa 11.52% 41.43% 16.10% 6.88% 27.91% 0.00% 79.78% 207.45% 20.61% Europe −8.34% −4.81% −42.25% 16.16% 10.88% −1.96% 39.10% 62.46% 3.63% South America 10.80% 25.62% 7.73% 16.23% 6857.14% 24.77% 163.70% 884.16% 12.04% North America −19.86% 2.11% −19.60% −6.94% −1.60% −0.71% 60.07% 248.42% 1.79% Oceania −8.48% 0.56% 12.50% 9.76% 2.22% 0.00% 79.51% 4086.76% 1.89% The incorrect Fig. 1 is: Global patterns of installed power capacity in 2020. Subnational regions with darker colors represent higher total installed power capacity (in megawatts) The correct Fig. 1 is: Global patterns of installed power capacity in 2020. Subnational regions with darker colors represent higher total installed power capacity (in megawatts) The incorrect Fig. 3 is: Subnational-level heterogeneity in power generation capacity structure and dominant energy mix. Each pie chart illustrates the top 15 subnational administrative units where a specific electricity source accounted for the highest share of local power generation in 2020 The correct Fig. 3 is: Subnational-level heterogeneity in power generation capacity structure and dominant energy mix. Each pie chart illustrates the top 15 subnational administrative units where a specific electricity source accounted for the highest share of local power generation in 2020 The incorrect Fig. 4 is: (Figure presented.) Heterogeneity in the structure of installed electricity across continents. Radar charts are employed to visualize the share of nine energy types in each continent The correct Fig. 4 is: (Figure presented.) Heterogeneity in the structure of installed electricity across continents. Radar charts are employed to visualize the share of nine energy types in each continent The incorrect Fig. 5 is: (Figure presented.) Top 15 positively contributing country-level drivers worldwide (SHAP value). (a) shows the global feature importance ranking based on normalized mean absolute SHAP values. To maintain clarity, only the top 15 of 19 variables are shown, accounting for 98.79% of the total contribution. The remaining four variables—OilR, RnD, CO ₂int, and CoalR—are omitted due to their minor combined impact (1.21%). (b) is a SHAP beeswarm plot, where each point represents the signed SHAP value (x-axis) of a sample for a given feature. Color indicates the original feature value from low (blue) to high (red). Points on the right (> 0) indicate a positive contribution to the average installed capacity (across nine energy types), while those on the left (< 0) indicate a negative effect The correct Fig. 5 is: (Figure presented.) Top 15 positively contributing country-level drivers worldwide (SHAP value). (a) shows the global feature importance ranking based on normalized mean absolute SHAP values. To maintain clarity, only the top 15 of 19 variables are shown, (b) is a SHAP beeswarm plot, where each point represents the signed SHAP value (x-axis) of a sample for a given feature. Color indicates the original feature value from low (blue) to high (red). Points on the right (> 0) indicate a positive contribution to the average installed capacity (across nine energy types), while those on the left (< 0) indicate a negative effect The incorrect 5.2 Key drivers of global energy capacity section is: Based on the random forest regression combined with SHAP value decomposition, we quantify the relative contribution of each variable and uncover complex nonlinear interactions. The resulting feature importance ranking is illustrated in Fig. 5(a). The five leading drivers are: energy intensity of output (EnInt, 22.6%), share of renewable electricity (RElec, 12.6%), urban population (UrbPop, 12.3%), population density (PopD, 10.9%), and total population (Pop, 8.3%). These five alone account for 66.6% of the total SHAP importance. When the next five are included (Fossil, EnImp, HHFCE, REcons, and GFCF), the top ten explain 92.1% of the total variation. This concentration suggests that a relatively small number of systemic, structural factors dominate the global patterns of installed capacity, and that changes in these few areas may disproportionately shape the energy landscape. Grouped by category, energy mix and efficiency variables (EnInt, RElec, REcons, Fossil and CO2int) for 47.2%; demand-and-scale variables (UrbPop, PopD, Pop, GDPc15 and HHFCE) account for 38.1%; capital and capability variables (GFCF, GNIpc, HTexp and RnD) for 5.8%; external vulnerability (EnImp) alone contributes 6.1%; and resource-rent indicators (CoalR/GasR/OilR) account for only 1.6%. Innovation-capacity indicators (RnD, HTexp) account for 1.3%; in a cross-technology average specification, their role is primarily enabling or threshold-setting rather than directly expansionary. Notably, although decarbonization is a key objective of power system transformation, our analysis shows that CO ₂ emissions intensity contributed less than 1% to the model. Energy mix and demand factors directly scale or constrain deployment, institutional and innovation related factors act more as long-run enablers. In other words, while capital or R&D capacity may not drive capacity growth uniformly, they are prerequisites in contexts where clean energy transitions are initiated or accelerated. This pattern implies that decarbonization, while crucial in narrative terms, remains an indirect outcome of broader development trajectories rather than a standalone policy target operationalized through capacity shifts. We also report the average marginal-effect relationships for these drivers and delineate their operational threshold intervals. For detailed visualizations, please refer to Fig. S1 in Supplementary Material 1. We further visualize nonlinear relationships and quantile-specific heterogeneity among the principal drivers through SHAP beeswarm plots and dependence curves in Fig. 5(b). Demand-scale variables (UrbPop, Pop, PopD, GDPc15 and HHFCE) exhibit acceleration above the median, evidencing a marginal effect of load agglomeration and scale amplification. Within structural variables, observations for RElec and REcons are concentrated on the positive axis and strengthen at higher quantiles, indicating a path dependence whereby a larger green share facilitates expansion. By contrast, Fossil shows a stronger negative contribution at higher quantiles, reflecting structural crowding-out of clean-capacity expansion. EnImp is typically positive in countries with high import dependence (security and price exposure spur domestic installation), but exhibits diminishing marginal effects as the “high-dependence” range is approached. ElecA reaches a plateau among high-coverage observations, suggesting that once universal access is achieved, incremental capacity shifts toward structural optimization and substitution-led upgrades. Together, these patterns support a tripartite mechanism of necessity–availability–accessibility: demand determines the necessity of additions; the energy mix and efficiency determine availability and the scale ceiling; and capital, industrial and institutional capacity determine the speed and difficulty of implementation. The correct 5.2 Key drivers of global energy capacity section is: Based on the random forest regression combined with SHAP value decomposition, we quantify the relative contribution of each variable and uncover complex nonlinear interactions. The resulting feature importance ranking is illustrated in Fig. 5(a). The two leading drivers are GDP per capita (GDPc15, 42.52%) and Household & NPISH final consumption expenditure (HHFCE, 33.38%). Collectively, these two factors alone account for nearly three-quarters of the total SHAP importance. Following these, the next most influential variables are urban population, GNI per capita, CO ₂ emissions, and total population. All remaining variables contribute less than 1% individually. This highly concentrated distribution indicates that structural socioeconomic conditions, particularly those categorized under "Economic Development" and "Urbanization", exert a paramount influence on national-level changes in installed capacity. Countries and regions with higher levels of economic development, stronger consumer demand, and larger or rapidly urbanizing populations tend to experience the most significant expansion in generation infrastructure. In contrast, variables representing Technology and Innovation Capabilities, Energy Structure and Efficiency, and Environmental and Emission-Reduction Pressures exhibit negligible contributions. We also report the average marginal-effect relationships for these drivers and delineate their operational threshold intervals. For detailed visualizations, please refer to Fig.S1 in Supplementary Material 1. We further visualize nonlinear relationships and quantile-specific heterogeneity among the principal drivers through SHAP beeswarm plots and dependence curves in Fig. 5(b). Most variables exhibit stronger positive contributions at the higher end of their value distributions. This pattern is particularly evident for indicators that reflect economic demand, such as GDP per capital, household consumption expenditure, and other measures of economic activity as well as demographic factors including population size and urban population. Countries and regions with higher levels of income, consumption, and population generally experience greater expansion in installed capacity, leading these variables to display larger SHAP values in their upper quantiles. In contrast, the SHAP distribution of gross fixed capital formation (GFCF) shows an atypical pattern. Because GFCF is highly correlated with major economic indicators that already dominate the model's explanatory power, and because much of GFCF reflects investment categories unrelated to power-sector expansion, its conditional contribution appears weaker and sometimes negative. This reflects statistical overlap rather than a substantive negative relationship, and is consistent with the presence of shared variance among correlated development indicators. The original article [1] has been corrected.