PaperlayerView exact revision on arXiv

Source-linked AI summary

Europe's Climate Ambition Under Scrutiny: Evidence from Deep Learning Emission Projections

Jacopo Ghirri, Carlos Rodriguez-Pardo, Lara Aleluia Reis, Massimo Tavoni

arXiv:2608.18690v1cs.LGcs.AIecon.GN

TL;DR

Whether observed sectoral dynamics keep the EU on course for its 55% 2030 emissions-reduction objective remains unanswered. We use deep learning to project sectoral emissions under current trends, finding a projected 620 Mt CO2 shortfall by 2030.

  • Problem

    Whether observed sectoral dynamics keep the EU on course for its 55% 2030 emissions-reduction objective remains unanswered.

  • Method

    We use deep learning on high-frequency socioeconomic and sectoral indicators to extrapolate empirical sectoral momentum into a current-trends counterfactual through 2030.

  • Results

    Approximately 620 Mt CO2 short of the Fit for 55 target, EU27 emissions are projected at 2.36 Gt in 2030, with only a small minority of countries aligned.

  • Takeaways & Limitations

    Closing Europe’s ambition-implementation gap will require interventions substantially beyond current policy portfolios, particularly in transport.

  • Takeaways & Limitations

    The projections exclude future policy interventions, behavioral responses, and structural shocks, so they are a current-trends counterfactual rather than forecasts of realized outcomes.

Abstract

from arXiv · show

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$_2$ 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$_2$ 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.

1 Introduction

The introduction frames short-term, sectoral emissions forecasting as essential for assessing whether the EU remains on course for its 55% reduction objective in 2030. It motivates empirical current-trends projections that avoid assumptions about future policy effectiveness while explicitly confronting structural and deep uncertainty.

  • EU sectoral emissions trajectories are needed to assess progress toward the 55% reduction objective in 2030 and identify where additional policy action is required.
  • Empirical trend extrapolation complements scenario models by grounding projections in observed behavioral and structural emission dynamics rather than assumptions of rational actors or perfect policy implementation.
  • The central policy question is what emissions trajectory Europe is currently on if observed sectoral dynamics persist.
  • Structural and deep uncertainty from shocks, transitions, and uneven policy changes can make deterministic or scenario-based forecasts overconfident and vulnerable to fragile extrapolation.
  • The study uses high-frequency sectoral indicators to construct a current-trends counterfactual without specifying future policy effectiveness, technology adoption, or demand trajectories.These projections represent Europe’s trajectory absent additional intervention and complement rather than replace scenario-based approaches.

2 Framework Overview

The framework forecasts sectoral CO2 emissions across EU27 countries from high-dimensional socioeconomic and sectoral indicators, using a latent representation to address nonlinear co-movement and uncertainty. It sequentially compresses observations, forecasts latent states through 2030, and targets emissions across six sectors without assuming hypothetical policy outcomes.

  • Data and targets: The study models sector-wise CO2 emissions across EU27 countries using high-dimensional indicators spanning energy, transport, economic activity, land use, agriculture, climate, and trade.CO2 is used as a proxy for total greenhouse gas emissions, while methane and nitrous oxide are excluded because their sectoral dynamics differ.
  • Data and targets: The target is a vector of emissions across s=6 sectors, indexed by country i and year t alongside transition indicators x(t, i) and context variables c(t, i).The formal setup distinguishes transition indicators from context variables for each country-year observation.
  • Core challenges: The framework addresses high dimensionality, nonlinear indicator co-movement, forward projection without assumed policy outcomes, and noise in emissions and transition indicators.Its objective is to extrapolate empirical momentum rather than simulate hypothetical policy interventions.
  • Sequential architecture: The sequential architecture uses a Variational Autoencoder to compress each country-year observation into a low-dimensional latent state representation capturing data structure and natural variability.The latent representation is modeled as z(t, i) ∼ N(μ, σ^2).
  • Sequential architecture: A latent forecaster then iterates each country’s latent state autoregressively from the last observed year through 2030, conditioned on context variables.The framework chains three components sequentially for each EU27 country, as reported in Figure 1.

3 Results

Under current trends, EU27 emissions decline through 2030 but remain substantially above the 55%-reduction pathway. Aggregate progress masks uneven country and sector trajectories, with Power decarbonizing strongly while Mobility remains structurally resistant.

  • EU27 aggregate trajectory: 622 Mt CO2 [m.c.b. 454 – 793 Mt] separates projected 2030 emissions from the 1.74 Gt target.EU27 emissions decline from 2.72 Gt in 2024 to 2.36 [m.c.b. 2.19 – 2.53 Gt] Gt by 2030, a 13.2% [m.c.b. -19.5% – -7%] reduction versus 36% needed.
  • Member-state heterogeneity: 10 countries show projected reductions exceeding 10%, while the other 17 show slower decarbonization or increases.The five largest emitters, collectively accounting for 65.5% of 2024 emissions, achieve a 15% [m.c.b. -24% – -6%] collective decrease.
  • Sectoral trajectories: 67% [m.c.b -74.9% – -59.9%] reduction brings Power emissions from 1.09 Gt in 2010 to 0.35 Gt [m.c.b. 0.27 – 0.44 Gt] in 2030.The trajectory reflects an accelerated structural shift toward renewable generation and is broadly consistent with Fit for 55 reduction requirements.
  • Sectoral trajectories: 31.4% from 2010 [m.c.b. -37.5% – -25.1%] leaves Mobility as the largest emitting sector in 2030.Heating & Cooling, Land Use, and Other sectors together account for 21% aggregate share of 2030 projected emissions [m.c.b. 19.7% – 22.4%], totaling 0.5 Gt [m.c.b. 0.45 – 0.54 Gt].
  • Member-state heterogeneity: 11 countries show projected Mobility increases, with no clear regional pattern, indicating structural rather than geographical resistance to transport decarbonization.By contrast, 11 of 27 countries show projected Power reductions exceeding 50%, reflecting the renewable transition.
  • Prediction drivers: GDP and population strongly influence predictions, while energy-related variables—including fuel pricing, power generation, and electricity trade—shape latent emission sensitivities.Context variables enter the predictor directly and have broad activations across all sectors.

4 Conclusions

Under current trends, EU27 emissions remain substantially off track for the 2030 Fit for 55 target. Sectoral divergence is pronounced: Power follows a credible decarbonization path, while Mobility remains the largest obstacle and additional intervention is required, particularly in transport.

  • Implementation gap: 620 Mt CO2 shortfall: EU27 emissions are projected to reach 2.36 Gt by 2030, a 13% reduction from 2024 levels.The data-driven framework extrapolates empirical sectoral momentum without embedding policy assumptions.
  • Sectoral divergence: 67% reduction from 2010 levels: Power generation is projected to follow a structurally credible decarbonization path by 2030.The projected decline reflects the accelerating renewable transition across the EU.
  • Sectoral divergence: 36.8% of 2030 emissions: Mobility is projected to achieve only a 7.5% reduction since 2010, making it the largest obstacle to European climate targets.Its resistance to change is not geographically concentrated.
  • Actionable determinants: Energy system structure, generation mix, electricity trade balance, and fuel pricing are identified as primary actionable determinants of near-term emission trajectories.These factors are presented as the most plausible levers for shifting current momentum.
  • Governance implications: Substantially beyond current policy portfolios: closing the gap will require additional interventions within the remaining timeframe, particularly in transport.The gap reflects the distance between the policy environment shaping observed emission dynamics and conditions assumed in optimistic scenarios.

4.1 Data Sources

The section identifies the datasets used to train all models and the sectoral decomposition of emission sources reported in Tables 1 and 2.

  • Table 1 reports the data sources used to train all models.
  • Table 2 reports the sectoral decomposition of emission sources.

4.2 Framework and Modeling rationale

The framework projects sectoral CO2 emissions through 2030 by chaining a probabilistic encoder, latent forecaster, and emission predictor trained on annual EU27 panel data. It forecasts historical emission dynamics at their observed pace without embedding policy targets, future technology adoption, or behavioral responses to price signals.

  • Framework: The model represents six-sector CO2 emissions across EU27 member states using high-dimensional transition indicators and context variables observed annually.The learned components are trained sequentially on the historical panel and applied autoregressively through 2030.
  • Latent representation: A probabilistic VAE encoder maps transition indicators to latent distributions, accommodating measurement noise and heterogeneous cross-national data quality while preserving covariance structure.This approach is motivated by the high dimensionality and nonlinear co-movement of the indicators, unlike deterministic alternatives such as PCA.
  • Latent forecasting: Forecasting occurs in latent space because its structured low-dimensional representation produces smoother, more regular year-to-year transitions than raw indicators.The latent forecaster iterates predictions from the last observed year through 2030.
  • Context and outputs: GDP, population, and climate variables enter the forecaster directly because their near-term trajectories are comparatively tractable and their emission effects may not be fully captured latently.The emission predictor maps the projected latent sequence and context variables to sectoral outputs alongside a learned uncertainty vector.
  • Modeling assumptions: The framework extrapolates historically manifested emission dynamics at their observed pace without embedding policy targets, assuming future technology adoption, or modeling behavioral responses to price signals.Uncertainty is represented by a learned sector-level vector detailed in Section 4.3.2.

4.3 Model Architecture

The framework combines a variational autoencoder, an uncertainty-aware sectoral emission predictor, and a latent forecaster to produce predictions with uncertainty estimates and autoregressive projections. The components use temporal context and probabilistic latent representations to model emission drivers and their evolution.

  • Architecture: The framework uses a variational autoencoder to learn country-year latent representations, followed by an emission predictor and a latent forecaster.The VAE captures meaningful representations; the predictor uses fixed latent states for sectoral estimates, while the forecaster projects latent features into the future.
  • Variational Autoencoder: The VAE maps transition indicators to probabilistic latent states, with a decoder that reconstructs the original input space.The encoder represents inputs as latent distributions, and the decoder reverses the information flow from latent samples to inputs.
  • Variational Autoencoder: 0.95 and 0.05 are the VAE reconstruction-loss and KL-divergence weights, respectively, with L1 reconstruction loss selected for convergence and meaningful representations.The VAE loss combines reconstruction loss with KL regularization toward a standard Gaussian.
  • Emission Predictor: The emission predictor combines projected latent states with current and previous-year context to estimate sectoral emissions and per-sector confidence.It jointly models outputs and uncertainty, allowing higher uncertainty where emissions are more variable or difficult to extrapolate.
  • Latent Forecaster: The latent forecaster uses latent states from the previous two years and current and previous-year climate and macroeconomic context to generate autoregressive latent projections.This preserves the cross-indicator covariance structure encoded by the VAE and respects its probabilistic latent structure.

4.4 Projection Procedure

The projection procedure autoregressively generates 2025–2030 sectoral emissions for each EU27 member state, using a special 2024 initialization. Monte Carlo uncertainty is applied only during emission prediction, not to latent dynamics.

  • Autoregressive projection: 2025–2030 projections are generated autoregressively for each EU27 member state, initialized with observed 2022 and 2023 indicator data.The latent forecaster advances one step per projection year, after which the emission predictor maps the latent state to per-sector emission change.
  • 2024 initialization: 2024 uses a latent representation from the forecaster because the broader indicator dataset is not sufficiently complete for robust forward conditioning.This 2024 representation initializes the 2025–2030 chain, while emission deltas are accumulated from 2025 onward.
  • Uncertainty treatment: Uncertainty is introduced only at the emission prediction step by resampling the encoder’s latent distribution while holding the forecasting chain fixed at the posterior mean.The resulting intervals capture uncertainty in mapping indicators to latent states and emissions but exclude uncertainty in latent dynamics.

4.5 Uncertainty Interval Calibration

The framework distinguishes model-internal predictability from Monte Carlo confidence intervals, which capture only uncertainty in mapping observed indicators to latent states and sectoral emissions. Post-hoc calibration rescales the intervals to achieve 90% historical coverage while preserving means and joint sectoral distributions.

  • Uncertainty representations: Two uncertainty representations are used: the learned output u(t, i) indicates relative sectoral predictability, while Monte Carlo intervals quantify predictive emission uncertainty.Monte Carlo intervals are generated by repeatedly resampling the encoder’s latent distribution during emission prediction.
  • Uncertainty representations: Monte Carlo sampling fixes the latent forecasting chain at the posterior mean and captures uncertainty from observed indicators through latent states to sectoral emissions.The latent distribution is sampled as z_t∼N(μ(x_t), σ^2(x_t)); uncertainty in latent forecasting and the predictor’s learned sectoral structure is not included.
  • Calibration: 90% coverage is targeted for the reported bands over the historical period 2010–2023 through post-hoc rescaling against empirical prediction errors.For each country-sector cell, N=500 emission samples are drawn per historical observation, and standardised residuals compare errors with the model’s estimated spread.
  • Calibration: The calibration leaves the mean unchanged, adjusts spread using T_i,s, and aggregates calibrated sector samples before percentiles to preserve the joint distribution.T_i,s>1 indicates underestimated historical errors, whereas T_i,s<1 indicates over-dispersion; marginal quantiles are not summed independently.

4.6 Limitations

The framework’s projections have several important limitations: they cover CO2 rather than all greenhouse gases, rely on lagged data, extrapolate historical momentum without future interventions, and do not provide formal probabilistic or causal inference. These constraints make comparisons with the FF55 target indicative and limit how projections, uncertainty measures, and attribution results should be interpreted.

  • Coverage and target comparison: CO2-only modeling excludes methane and nitrous oxide, so comparisons with the all-greenhouse-gas FF55 target are indicative rather than exact.These gases have more erratic sectoral dynamics and distinct governing processes.
  • Data recency: Data observed through 2023 leave 2024 reconstructed rather than directly encoded, preventing the model from reflecting the most recent policy, price, and technology shifts.The training-data publication delay means the latest panel necessarily lags the present, although the framework can be re-estimated as new annual data arrive.
  • Extrapolation assumptions: Historical-momentum extrapolation misses exogenous shocks and abrupt policy changes without precedent in the historical record.The framework does not condition on future policy interventions, behavioral responses, or structural shocks.
  • Uncertainty quantification: Learned confidence scores and calibrated Monte Carlo consistency bands are not formal classical prediction intervals and should not be translated into probabilistic statements.Confidence scores indicate relative sectoral predictability rather than probabilities.
  • Causal interpretation: Gradient-based attribution and sensitivity analyses reveal historical co-movement rather than structural causal relationships.They provide high-level insight into model structure and emission-dynamic drivers but do not constitute causal inference analysis.

Supplementary … Rolling-Origin evaluation of the Forecasting Procedure

The supplementary analyses show that the VAE learns structured country trajectories and that the forecasting pipeline performs well on held-out data and short-term rolling-origin projections. Forecast errors increase with horizon and are concentrated around structural disruptions and sectors exposed to demand shocks.

  • Latent representation inspection: The VAE latent space improves country separation and temporal coherence relative to the dense clustering and scale-driven outliers in the original input space.Countries with similar decarbonisation profiles become more spatially separated and temporally coherent.
  • Model Performance: 0.981 aggregate validation R2 for the emission predictor, with sectoral values from 0.964 (Power) to 0.991 (Land), indicates strong predictive performance.Validation R2 exceeds training R2, 0.981 versus 0.969, which is attributed to dropout-based regularization during training.
  • Model Performance: 0.0062 versus 0.0065 train and validation MSE for the latent forecaster shows minimal generalization gap and predictable year-to-year latent transitions.The random validation split does not constitute a genuinely unseen-time evaluation.
  • Rolling-Origin evaluation of the Forecasting Procedure: Rolling-origin evaluation retrains the VAE, emission predictor, and latent forecaster through cutoff years 2018–2021, then evaluates autoregressive projections over the next three years.This procedure tests projections for genuinely unseen time periods rather than interpolation within the training window.
  • Metric Train Validation: 0.227 to 0.326 mean RMSE from one to three years ahead increases monotonically, with horizon +3 approximately 40% larger than horizon +1.The degradation is consistent with accumulated autoregressive errors while remaining modest for short-term extrapolation.
  • Metric Train Validation: 0.416 RMSE at horizon +2 for the 2018 cutoff reflects projections into the 2020 COVID-19 demand shock, which trend extrapolation could not anticipate.Later cutoffs overlapping post-shock recovery show RMSE values between 0.172 and 0.311 across horizons.
  • Metric Train Validation: 0.121–0.203 for Land and 0.129–0.232 for Heating & Cooling are the lowest sectoral RMSE ranges across horizons, reflecting structural regularity and shock insulation.Elevated errors are concentrated in sectors and years affected by real-world structural disruptions, which the framework treats as counterfactual out-of-trend events.

Model sensitivity · Sectoral emission changes, and model confidence, by country and sector

The model’s sensitivity is concentrated in a small number of information-dense inputs, with fuel prices dominating across nearly all sectors. Sectoral projections show highest confidence for Power, while Mobility generally has medium confidence and uncertainty around near-zero changes across several countries.

  • Model sensitivity: Total-order Sobol indices assess which inputs most strongly shape prediction variance, rather than estimating real-world policy transmission.The analysis uses the full variable set but displays only variables ranking among the top five for at least one sector.
  • Model sensitivity: Predictive variance is attributable to a small number of information-dense inputs after aggressive dimensionality reduction.Most variables carry negligible sensitivity indices.
  • Model sensitivity: Fuel prices dominate the sensitivity structure across nearly all sectors.Their prominence is interpreted as reflecting information density and co-movement with the business cycle, energy markets, and geopolitical shocks.
  • Model sensitivity: GDP shows concentrated sensitivity in Industry and Power.The passage links this pattern to the economically coherent signals captured by the model’s sensitivity structure.
  • Sectoral emission changes, and model confidence, by country and sector: Power carries the highest confidence across EU27 member states.This reflects the structural regularity of the renewable transition and the predictability of its near-term trajectory.
  • Sectoral emission changes, and model confidence, by country and sector: Mobility confidence is generally medium, including in several countries projecting near-zero changes.These cases indicate uncertainty over whether changes tip marginally positive or negative; Figure 9 reports percentage changes from 2024 to 2030 and encodes direction and confidence.
Loading 2608.18690v1…