XAI in European electricity markets: SHAP analysis across 39 zones

Antoine Pesenti and Aidan O'Sullivan on 17 June 2026 submitted a 12‑page paper (arXiv:2606.19118) that combines deep neural networks with explainable AI to analyse electricity price formation across 39 European bidding zones. The authors pair DNN prediction models with SHAP explanations and extend the SSHAP aggregation framework to improve interpretability in high‑dimensional settings.

What did the authors build and test?

The paper trains deep neural networks to predict electricity prices for 39 European bidding zones and then applies SHAP to quantify each feature's contribution, while extending SSHAP to aggregate explanations across many features and zones. The modelling pipeline produces localized SHAP attributions for the DNNs and an aggregated view via SSHAP, and the authors also construct a synthetic EU‑wide electricity market to explore the counterfactual of a single integrated price.

The submission lists renewable generation inputs, gas prices and interconnection flows among the features examined. The authors emphasise solar generation as having a disproportionate influence on price formation despite its lower share in total generation, and they flag gas prices as a dominant, consistent driver across markets.

How do drivers and interdependencies emerge in the results?

Across the 39 bidding zones the analysis finds renewable energy sources and gas exert different patterns of influence: solar shows an outsized role in price formation relative to its generation share, while gas prices remain a steady primary driver. Interconnections between zones are shown to significantly shape price dynamics, underscoring the interdependence of European electricity systems.

The paper highlights two contrasts: first, renewable inputs, particularly solar, punch above their weight in explaining price swings; second, cross‑border links alter local price responses because flows change exposure to external drivers. To make those contrasts interpretable at scale the authors rely on SSHAP to aggregate SHAP contributions across zones and features.

Why it matters

The combination of DNNs with SHAP and an aggregation layer tackles a practical tradeoff: neural networks capture nonlinear, high‑dimensional relations that traditional econometric approaches may miss, but they are hard to interpret. By quantifying feature contributions and aggregating them across 39 zones, the paper turns predictive power into diagnostic insight about what actually drives prices: solar variability, gas markets and interconnections. That insight matters for market participants, system operators and policymakers who need to understand not just forecasts but the drivers behind them.

What did the paper add methodologically?

The core methodological contribution is the application of SHAP to DNN forecasts across many zones and the extension of SSHAP to improve interpretability in high‑dimensional settings. The authors present an aggregated explainability workflow that scales local SHAP attributions to a pan‑European analysis and use a synthetic EU‑wide market to probe counterfactuals where a single price prevails.

What to watch

Watch for a peer‑reviewed version or accompanying release of code and data tied to arXiv:2606.19118, which would let other researchers reproduce the SSHAP aggregations and the synthetic EU‑wide market experiments. Also look for follow‑on work that applies the SSHAP aggregation to alternative feature sets or to operational decision problems that rely on interpretable price drivers.

Details and identifiers: the paper is arXiv:2606.19118, submitted on 17 June 2026, and runs 12 pages; authors are Antoine Pesenti and Aidan O'Sullivan.