Will the Carbon Border Adjustment Mechanism Impact European Electricity Prices? A GNN-Based Network Analysis

Jiachen Shen, Jian Shi, Dan Wang, Zhu Han · 2026

Read Will the Carbon Border Adjustment Mechanism Impact European Electricity Prices? A GNN-Based Network Analysis by Jiachen Shen, Jian Shi, Dan Wang, Zhu Han (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-03304.txt.

Introduction

Jiachen Shen, Jian Shi, Dan Wang, and Han Zhu model how the EU carbon border charge could reshape electricity prices across linked national markets. Static studies miss cross-border spillovers, which matter greatly in Europe's interconnected grid.

Their spatio-temporal graph neural network tracks prices and carbon intensity together across eight countries. In simulated scenarios, the charge did not act as a flat tax: low-carbon France and Switzerland gained an edge with possible price drops, while high-carbon Poland faced rising costs on two fronts. A shift in the market merit order drove the split.

Catalog record

CREATORJiachen Shen, Jian Shi, Dan Wang, Zhu Han
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.03304

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Will the Carbon Border Adjustment Mechanism Impact European Electricity Prices? A GNN-Based Network Analysis Jiachen Shen, Jian Shi, Dan Wang, Zhu Han Abstract The European Union’s Carbon Border Adjustment Mechanism (CBAM) creates a complex challenge for the interconnected European electricity market. Traditional static analyses often miss the cross-border spillover effects that are vital for understanding this policy. This paper addresses this gap by developing a spatio-temporal Graph Neural Network (GNN) framework. It quantifies how CBAM affects electricity prices and carbon intensity (CI) at the same time. We modeled a subgraph of eight European countries. Our results suggest that CBAM is not just a uniform tax. Instead, it acts as a tool that transforms the market and creates structural differences. In our simulated scenarios, we observe that low-carbon countries like France and Switzerland can gain a competitive advantage. This suggests a potential decrease in their domestic electricity prices. Meanwhile, high-carbon countries like Poland face a double burden of rising costs. We identify the primary driver as a fundamental shift in the market’s merit order. I Introduction The European Union (EU) has committed to ambitious climate goals under the European Green Deal [ 1 ] . This includes the Carbon Border Adjustment Mechanism (CBAM) to address carbon leakage [ 2 , 3 ] . This policy arrives while people are very sensitive about European energy costs. This follows a recent energy crisis that caused prices to surge. Consequently, there is significant public concern that CBAM will be just another tax that burdens consumers. This paper challenges that assumption. We ask a simple question. Is this blame accurate? Does CBAM add costs uniformly, or could its true impact on electricity prices be more complex and asymmetric? The question is not trivial because the European power grid is a deeply interconnected network. It is not a set of isolated markets [ 4 , 5 ] . Static, single-country analyses often miss the dynamic, cross-border spillover effects that define this system. To address this gap, we develop a modeling framework based on Graph Neural Networks (GNNs). This allows us to capture these dynamic, networked effects better than static analysis. Our analysis suggests that CBAM is not a uniform financial burden. Instead, it acts as a market transformation tool that creates clear differences between countries. We find that low-carbon countries like France and Switzerland gain a significant competitive advantage. In our simulated scenarios, the model suggests that CBAM may lead to a decrease in domestic electricity prices for these nations. Conversely, high-carbon grids like Poland and the Czech Republic face a dual challenge of rising costs and lost competitiveness. We identify the primary driver as a fundamental shift of the market’s merit order. The main contributions of this paper are threefold: • We develop a novel dual-target GNN framework to model the relationship between price and carbon intensity. This extends beyond previous single-task forecasting models [ 6 ] . • We provide quantitative, network-aware evidence that challenges the idea of a uniform tax. We demonstrate asymmetric economic impacts in our simulations, including potential price decreases for specific regions. • We identify the market mechanisms like merit-order shifting that drive this divergence. Because we cannot observe ground truth for a policy that has not fully happened yet, we focus on internal validity. We support our findings through predictive back-testing on observed data, robustness checks, and placebo tests. II Background and Methodology II-A CBAM and the Network-Based Challenge The CBAM regulation requires importers to submit certificates for the embedded emissions of imported electricity [ 3 ] . The European market is highly coupled and managed by ENTSO-E [ 7 ] . In this system, electricity is pooled. This makes it nearly impossible to trace the exact origin and carbon content of an imported megawatt-hour [ 4 ] . This complexity requires a network-based model. A tax applied at one border like Germany will not stop there. Its cost will factor into market bids. This influences trade and alters supply and demand for all neighbors like France and Poland. It eventually creates spillover effects [ 8 ] . These effects vary over time. This requires a graph analysis that considers both space and time. II-B Spatiotemporal GNN Framework Our approach extends the carbon intensity forecasting model (CFCG) presented in [ 6 ] . The original CFCG model is a single-task framework. It was designed only to predict carbon intensity (CI). Our main innovation is a dual-target, multi-task learning architecture. This allows us to model and predict two coupled market variables together: electricity Price ( Y_{Price} ) and Carbon Intensity ( Y_{CI} ). This section details the theoretical foundation of our framework. II-B 1 Dual-Target Problem Formulation Price and CI are not independent. Under CBAM, they are explicitly linked. The policy adds a carbon cost to the base price: P_{t}^{CBAM}=P_{t}^{base}+\tau\cdot CI_{t}+\epsilon_{t} . This creates a non-zero covariance ( \text{Cov}(CI_{t},P_{t})\neq 0 ) that a single-task model would miss. Therefore, we formulate the problem as a multi-task learning challenge. The original CFCG model optimized a single loss \mathcal{L}_{CFCG}=\min_{\theta}\mathbb{E}[\ell(f_{\theta}(x),y_{CI})] . In contrast, our dual-target framework optimizes a composite loss function: \small\begin{split}\mathcal{L}_{dual}=\min_{\theta,\phi}&\mathbb{E}_{(x,y)\sim\mathcal{D}}\Big[\lambda_{1}\cdot\text{MSE}(Y_{CI},\hat{Y}_{CI})\\ &+\lambda_{2}\cdot\text{MSE}(Y_{Price},\hat{Y}_{Price})+\lambda_{3}\cdot\mathcal{L}_{Corr}\Big]\end{split} (1) Here, \lambda_{1},\lambda_{2},\lambda_{3} are weighting hyperparameters. The first two terms are standard Mean Squared Error (MSE) losses for each task. The third term, \mathcal{L}_{Corr} , is a novel Correlation Loss. It ensures the model preserves the underlying correlation structure between the two targets: \small\mathcal{L}_{Corr}=\left|\rho(Y_{CI},Y_{Price})-\rho(\hat{Y}_{CI},\hat{Y}_{Price})\right| (2) where \rho(\cdot,\cdot) is the Pearson correlation coefficient. This term acts as a coupling regularization. This approach is based on multi-task learning theory [ 9 , 10 ] , which provides two key benefits: Inductive Bias Transfer , where the model is forced to find a joint hypothesis space \mathcal{H}_{joint}=\mathcal{H}_{CI}\cap\mathcal{H}_{Price} leading to more generalizable features; and Statistical Efficiency , as joint estimation reduces parameter variance compared to independent training. II-B 2 Model Architecture and Optimization Fig. 1: Our dual-target GNN architecture. It shows the shared encoder and parallel task-specific (CI and Price) MLP heads. We design a parallel architecture to learn this representation efficiently (Fig. 1 ). Instead of using the complex GNN-LSTM stacks found in the original CFCG, our model employs a shared GNN encoder ( \text{GNN}_{\theta} ) to extract features from the input graph G=(V,E) , utilizing node features X and adjacency matrix A . This shared encoder learns a latent representation Z that captures the network-wide state: \small Z=\text{GNN}_{\theta}(X,A)\in\mathbb{R}^{|V|\times d}\vskip-5.69054pt (3) Subsequently, this shared representation Z is processed by two parallel, independent Task-Specific Heads implemented as Multi-Layer Perceptrons (MLPs). These simpler heads correspond to our dual targets and generate the final predictions: \small\hat{Y}_{CI}=\text{MLP}_{CI}(Z),\hskip 9.24994pt\hat{Y}_{Price}=\text{MLP}_{Price}(Z)\vskip-5.69054pt (4) Algorithm 1 Dual-Target GNN Training Framework 1: Input: Graph G=(V,E) , Features X (incl. policy), Adjacency A , Ground Truth Y_{CI},Y_{Price} , Loss weights \lambda_{1},\lambda_{2},\lambda_{3} 2: Output: Trained models \text{GNN}_{\theta} , \text{MLP}_{CI} , \text{MLP}_{Price} 3: Initialize parameters \theta (for \text{GNN}_{\theta} ), \phi_{CI} (for \text{MLP}_{CI} ), \phi_{Price} (for \text{MLP}_{Price} ) 4: for each training epoch do 5: \triangleright Shared GNN Encoder (Eq. 3 ) 6: Z\leftarrow\text{GNN}_{\theta}(X,A) \triangleright Task-specific attention applied 7: \triangleright Parallel Task-Specific Heads (Eq. 4 ) 8: \hat{Y}_{CI}\leftarrow\text{MLP}_{CI}(Z;\phi_{CI}) 9: \hat{Y}_{Price}\leftarrow\text{MLP}_{Price}(Z;\phi_{Price}) 10: \triangleright Calculate Multi-Task Loss (Eq. 1 ) 11: \mathcal{L}_{MSE-CI}\leftarrow\text{MSE}(Y_{CI},\hat{Y}_{CI}) 12: \mathcal{L}_{MSE-Price}\leftarrow\text{MSE}(Y_{Price},\hat{Y}_{Price}) 13: \mathcal{L}_{Corr}\leftarrow\left|\rho(Y_{CI},Y_{Price})-\rho(\hat{Y}_{CI},\hat{Y}_{Price})\right| 14: \mathcal{L}_{dual}\leftarrow\lambda_{1}\mathcal{L}_{MSE-CI}+\lambda_{2}\mathcal{L}_{MSE-Price}+\lambda_{3}\mathcal{L}_{Corr} 15: \triangleright Backpropagation 16: Update \theta,\phi_{CI},\phi_{Price} using \nabla\mathcal{L}_{dual} 17: end for 18:

Read the complete text