cs.LGAug 19, 2026
SaveEurope's Climate Ambition Under Scrutiny: Evidence from Deep Learning Emission Projections
Organizations: Politecnico di Milano, Piazza Leonardo da Vinci 32, Milan, 20133, Italy · CMCC Foundation - Euro-Mediterranean Center on Climate Change, Via Marco Biagi 5, Lecce, 73100, Italy · RFF-CMCC European Institute on Economics and the Environment, Via Bergognone 34, Milan, 20144, Italy
Abstract
The European Union has committed to reducing greenhouse gas emissions 55% below 1990 levels by 2030, but whether current trends are compatible with this ambition remains uncertain. We apply deep learning to high-resolution socioeconomic and sectoral data across EU27 member states till 2023 to project sectoral CO trajectories under current trends, extrapolating observed sectoral momentum without assuming changes in the pace or effectiveness of the policy environment beyond what is already reflected in historical data. We project that EU27 emissions will exceed the 2030 target by 35% (620 Mt CO shortfall), with only a small minority of countries on trajectories consistent with the bloc's commitments. While the Power sector achieves target-consistent reductions driven by the renewable transition, Mobility shows minimal progress and accounts for over a third of total emissions by 2030, reflecting a structural inertia across member states rather than geographically concentrated lag. Our findings indicate that substantial additional intervention is required to close Europe's ambition-implementation gap, and call for establishing up-to-date energy information in Europe.
Figures & tables
Figure 1: Overview of the modeling framework. The framework consists of three components: a VAE that maps high-dimensional transition indicators to a probabilistic latent representation ; a latent-space forecasting model that projects these representations forward using context variables ; and an emission predictor that maps the projected states and context variables to sector-wise emission estimates and uncertainty outputs . Here denotes the year and the EU27 member state. For each country, observed indicators are encoded, iteratively projected to 2030, and decoded into sectoral CO 2 emission projections
Figure 2: Projected EU27, and all member states, CO 2 emission change from 2024 to 2030. Panels (a) and (b) show the projected percentage and absolute change in EU27 aggregate emissions relative to the 2024 baseline, respectively, for this study and six reference projections (OECD BAU, OECD Energy Transition, EEA With Existing Measures, EEA With Additional Measures, PyPSA Baseline, PyPSA Fit for 55). The dashed line indicates the change needed to meet a 55% reduction from 1990 levels. Panel (c) shows the distribution of projected emission changes by member state on a symmetric logarithmic scale, with each source plotted separately per country. Countries are ordered by their projected change under this study.
Figure 3: EU27 CO 2 emission trajectories from 2010 to 2030, decomposed by country and sector, with projected 2030 uncertainty distributions. Panels (a) and (b) show historical (2010–2024) and projected (2025–2030) stacked area charts of total CO 2 emissions, decomposed by country group and sector, respectively. The dashed vertical line marks the boundary between the observed and projected periods. Panels (c) and (d) show box-and-whisker plots of the 2030 projected emissions across 10,000 Monte Carlo samples, by country group and sector. Boxes show the interquartile range and whiskers the 90% model-consistency band, rescaled to match model residuals ditribution (see Methods Section 4.5 ).
Figure 4: Projected total and sectoral CO 2 emission changes from 2024 to 2030 across EU27 member states. Panel a shows the aggregate change across all sectors. Panels b to g show changes for Heating & Cooling, Industry, Land Use, Mobility, Other, and Power, respectively. Circle color encodes the direction and magnitude of change (blue: decreasing; red: increasing), with each panel using an independent color scale to reflect sector-specific ranges. Circle size encodes 2024 per capita emissions, with reference values shown below each panel.
Figure 5: Gradient-based attribution flows across the emission prediction pipeline. Input variables (left) connect through VAE latent dimensions (center) to sectoral emission outputs (right). Context variables enter the predictor directly, bypassing the encoder. For each emission output, the five most influential variables are shown; for each selected latent dimension, the five most influential input variables are shown. Node and edge widths are proportional to mean absolute gradient, averaged across all country-year observations in the training sample.
| Category | Variables | Source |
|---|---|---|
| Emissions | CO 2 emissions by NACE activity and households (total and per capita) | Eurostat |
| Agriculture | Production of fruits, meats (total and poultry), vegetables, wheat | FAO |
| Energy Consumption | GJ per capita energy use (industry, transport, final consumption) | Eurostat |
| Economic Indicators | Energy taxes (€M), GDP (quarterly in €M), Population | Eurostat |
| Transportation | Modal split of freight (air, rail, road, sea), electric vehicle sales and stock, trains (diesel/electric) | Eurostat, IEA |
| Renewable Energy | Heat pumps (GWh), solar thermal collectors’ surface | Eurostat |
Table 1: Input variables used in our study, and their sources.
| Sector | Activity according to Eurostat definitions |
|---|---|
| HeatingCooling | Heating and cooling by households |
| Industry | Manufacturing, mining, construction |
| Land | Agriculture, forestry, fishing, water supply, waste management |
| Mobility | Transportation, storage, trade, vehicle repair, transport by households |
| Power | Electricity, gas, steam, air conditioning |
| Other | Health, science, technical and professional services, finance, insurance, information, communication, public admin., arts, entertainment, defense, administrative services, real estate, accommodation, food service, education, other services, and household activities |
Table 2: Target emission variables , gathered from Eurostat.
Figure 6: t-SNE projections of (a) the original high-dimensional input space and (b) the VAE latent space across all EU27 countries and years. Each point represents a country-year observation; connected points trace the temporal trajectory of each country. The latent space shows markedly improved separation between countries compared to the dense clustering observed in the original input space, indicating that the VAE learns a more structured and disentangled representation of emission-relevant characteristics.
Figure 7: UMAP projections of (a) the original high-dimensional input space and (b) the VAE latent space across all EU27 countries and years. Each point represents a country-year observation; connected points trace the temporal trajectory of each country. Compared to the original space, the latent representation exhibits greater country separation and more continuous temporal trajectories, reflecting the VAE’s ability to organise countries along meaningful structural dimensions while preserving global neighbourhood relationships.
| Metric | Train | Validation |
|---|---|---|
| Reconstruction loss (L1) | 0.124 | 0.179 |
| KL divergence | 0.819 | 0.818 |
| Mean latent mean | 0.006 | 0.013 |
| Mean latent std | 0.580 | 0.578 |
Table 3: Model performance on held-out validation sets (15% random split, seed 0). Predictor metrics are reported in the z-score normalised emission space used during training (normalised kg CO 2 per capita). Latent forecaster metrics are in the normalised latent space.
| Training cutoff | Horizon +1 | Horizon +2 | Horizon +3 |
|---|---|---|---|
| 2018 | 0.282 | 0.416 | 0.369 |
| 2019 | 0.261 | 0.225 | 0.263 |
| 2020 | 0.172 | 0.260 | 0.311 |
| 2021 | 0.191 | 0.279 | 0.363 |
| Mean | 0.227 | 0.295 | 0.326 |
Table 4: Rolling-origin evaluation: aggregate RMSE (normalised emission space) for autoregressive projections of 1, 2, and 3 years ahead. Each row corresponds to models retrained from scratch on data up to the indicated cutoff year.
| Sector | Horizon +1 | Horizon +2 | Horizon +3 |
|---|---|---|---|
| Land | 0.121 | 0.168 | 0.203 |
| Heating & Cooling | 0.129 | 0.190 | 0.232 |
| Industry | 0.164 | 0.226 | 0.267 |
| Mobility | 0.253 | 0.366 | 0.471 |
| Power | 0.305 | 0.414 | 0.345 |
| Other | 0.325 | 0.363 | 0.379 |
Table 5: Rolling-origin evaluation: RMSE by sector and forecast horizon, averaged across training cutoffs . Metrics are in the z-score normalised emission space used during training. Sectors are ordered by mean RMSE across horizons.
Figure 8: Sensitivity of sectoral CO 2 emission predictions to input variables, measured by total-order Sobol indices (ST). Each cell reports the ST index for a given variable-sector pair. Variables are grouped thematically: fuel prices (left), macroeconomic indicators (centre), and electricity generation (right). The analysis is conducted on the full variable set, but only variables appearing in the top five for at least one sector are shown.
Figure 9: Projected sectoral CO 2 emission changes from 2024 to 2030 across EU27 member states. Each cell reports the percentage change in sector-level emissions between 2024 and 2030 under current trends. The bivariate color scheme encodes both the direction of change (blue: decreasing; red: increasing; grey: near zero, defined as within ±5%) and model confidence (light to dark, derived from the learned uncertainty of the emission predictor). The "Overall" row reports the aggregate change across all sectors. The EU27 column reports the emission-weighted aggregate across all member states. Bold separators distinguish the aggregate row and column from individual entries.
More in this topic
Policy Evaluation
Policy evaluation estimates the outcomes, value, risks, and fairness of decision rules before or after deployment.
See all 55 papers