Analysing drivers and interdependencies in European electricity markets using XAI

Technical Summary: Analysing Drivers and Interdependencies in European Electricity Markets using XAI

Problem Statement
European electricity markets are characterized by strong nonlinearities, high-dimensional interactions, and increasing regional interdependence. While Deep Neural Networks (DNNs) have demonstrated superior predictive accuracy for electricity prices compared to traditional econometric models, their "black-box" nature limits their utility for economic interpretation and policy analysis. Existing explainable AI (XAI) applications in Energy Price Forecasting (EPF) often treat explainability as an intermediary step for stacking models rather than a tool to uncover the fundamental drivers of market dynamics. Furthermore, standard feature importance methods struggle in high-dimensional settings where features (such as generation mixes across 39 bidding zones) are numerous and interdependent, potentially leading to misleading aggregations of importance.

Methodology
The authors analyze 39 European bidding zones across 27 countries using a dataset spanning January 1, 2023, to December 31, 2025. The data includes day-ahead electricity prices, load, power generation mixes (categorized into solar, wind, hydro, nuclear, other renewables, fossil gas, and other fossil fuels), and TTF gas prices.

1. Modeling: The study employs a Multi-Layer Perceptron (MLP) with two hidden layers (128 and 32 neurons) trained with the Adam optimizer and Mean Absolute Error (MAE) loss. The models are non-temporal; inputs and outputs correspond to the same hourly period, meaning the models are designed to explain market mechanisms rather than forecast future prices.
2. Feature Engineering: For each bidding zone, the input features include local load and generation mix, gas prices, and the load/generation mix of all electrically connected neighboring zones. This results in up to 364 potential features per model.
3. Explainable AI (XAI) Framework:
   • SHAP (SHapley Additive exPlanations): Used to quantify feature contributions. The authors utilize conditional expectation for Shapley value estimation via Monte-Carlo sampling to handle dependencies.
   • SSHAP (Super-SHAP): To address the limitations of standard SHAP in high-dimensional settings (specifically, the risk of overestimating importance when features within a subset offset each other), the authors apply and extend the SSHAP framework. This method aggregates features into "super-features" (e.g., grouping all solar features across zones or grouping all features belonging to a specific country) and calculates netting within these subsets before determining global importance.
4. Counterfactual Scenario: A synthetic "EU-wide" market model is constructed as a linear combination of the 39 individual zone models, weighted by mean load, to simulate a scenario with a single common clearing price.

Key Contributions

• Comparative Analysis: The paper provides a comprehensive analysis and comparison of 39 distinct European electricity markets, moving beyond single-market studies.
• Dual-Perspective Drivers: It identifies key price drivers from two distinct angles: the power generation mix and the influence of interconnected neighbors.
• Methodological Extension: The study demonstrates the application of SSHAP to handle high-dimensional models in electricity markets, offering a more robust aggregation of feature importance than standard SHAP when dealing with correlated variables.
• Counterfactual Exploration: It constructs a synthetic EU market to explore the "What if" scenario of a fully integrated European market with a single price.

Results

• The Disproportionate Role of Solar: Despite solar power having the lowest share of total generation (approx. 10%) among the main sources, SSHAP analysis reveals it as the most influential super-feature in price formation. Solar generation exhibits high variability and strong directional impact (long negative tails in SHAP distributions), capable of driving prices down significantly. In the EU-wide model, solar is the top super-feature, followed by gas price.
• Gas Price Dominance: Gas prices remain a dominant and consistent driver across markets, acting as the primary marginal cost setter in the merit order. In the EU model, the gas price feature is the single most important individual feature, capable of pushing prices up by €40/MWh.
• Load vs. Generation: While load is traditionally viewed as a primary driver, SSHAP analysis suggests its explanatory power is less significant than initially thought when feature offsets are accounted for. Load features often exhibit mixed directional impacts (positive and negative correlations) that cancel each other out in global importance metrics.
• Interdependence and Neighbors: Neighboring markets exert a massive influence on domestic prices. On average, neighbor super-features account for 61% of the price formation drivers across the 39 zones. For Germany, 6 out of 11 neighbors have German features as their dominant super-feature. Conversely, Spain and the Italian islands show the lowest neighbor dependence (34%), correlating with lower interconnector capacity relative to load.
• Regional Specifics:
  • Poland: Despite having 72% fossil fuel generation, solar and wind are the most important sources for explaining price movements.
  • Italy: Gas is the dominant generation source and the dominant price driver across all six Italian bidding zones, consistent with high average prices and low volatility.
  • Nordic Regions: Hydro and wind are dominant, though exceptions exist where cross-border influences (e.g., German hydro affecting Denmark) dictate the dominant super-feature.

Significance and Claims
The paper claims that combining DNNs with XAI, specifically the SSHAP framework, allows researchers to move beyond predictive performance to deliver interpretable insights into the mechanisms of electricity price formation. The study highlights that traditional linear models fail to capture the nonlinear relationships and high-dimensional dependencies of modern markets.

The authors assert that their findings challenge conventional wisdom regarding price drivers:

1. Renewables as Price Setters: Renewable sources, particularly solar, play a disproportionately large role in setting prices relative to their generation volume, primarily due to their variability.
2. Systemic Interdependence: Electricity prices are not determined solely by domestic conditions; cross-border flows and neighbor markets are often the primary determinants of price dynamics.
3. Policy Implications: The insights provided can assist in assessing market design choices, such as increasing interconnection capacity. The synthetic EU market model serves as a tool to understand the potential impacts of a fully integrated European market.

The paper concludes modestly, noting that while the current approach uses neighboring features as proxies for interconnection, future work could explicitly model transmission constraints and flows to better capture congestion effects. Additionally, the authors note that current XAI methods may not fully explain extreme high-price events, as the long negative tails of renewable generation (price suppression) are more pronounced than their positive impacts.