---
title: 'Real-E: Foundation Electricity Forecast Benchmark'
url: https://www.emergentmind.com/topics/real-e-dataset
type: topic
---

# Real-E: Foundation Electricity Forecast Benchmark

Real-E is a foundation benchmark for electricity forecasting introduced by Chen Shao, Yue Wang, Zhenyi Zhu, Zhanbo Huang, Sebastian Pütz, Benjamin Schäfer, Tobias Käfer, and Michael Färber. Built from the ENTSO-E Transparency Platform and updated to June 2024, it is designed to support robust and generalizable forecasting at real-world scale by covering the full electricity lifecycle—generation, transmission, balancing, market, and load—across Europe with rich metadata, long temporal coverage, and multiple temporal resolutions. The benchmark couples this data release with standardized evaluation protocols, an extensive baseline suite, and explicit analysis of non-stationary correlation structures that are often underrepresented in earlier benchmarks [2509.05768].

## 1. Origin and design rationale

Real-E was introduced to address limitations that the authors identify in existing electricity forecasting benchmarks: restricted spatial and temporal scope and insufficient multi-energy coverage. The dataset is therefore explicitly organized as a multi-country, multi-energy, and multi-resolution benchmark intended to capture the operational complexity of real-world electricity systems rather than a single-country or single-modality slice.

A central design choice is lifecycle completeness. Instead of restricting the benchmark to demand or generation alone, Real-E includes signals spanning generation, load, market prices, transmission, and balancing. This permits analyses of coupled processes rather than isolated time series. The accompanying paper presents this breadth as necessary for studying robust generalization, especially under evolving inter-series dependencies that cannot be adequately characterized by static correlation assumptions [2509.05768].

The paper further positions Real-E as a benchmark for non-stationary environments. A plausible implication is that the dataset is meant not only to evaluate point forecasting accuracy, but also to expose whether models can remain reliable when correlations among energy types, countries, or grid components shift over time.

## 2. Coverage, modalities, and schema

Real-E covers 39 European countries spanning 10.18 million km². Its current release spans up to 10 years, specifically 2014–2024, with a 9.5-year duration in the released corpus. It includes over 74 power stations and over 20 distinct energy categories, including wind, solar, hydro, thermal, nuclear, pumped storage, biomass, gas, hard coal, and lignite/brown coal. The temporal resolutions are 15-minute and hourly, with annual records for capacity [2509.05768].

The dataset is structured around lifecycle segments and subtypes. Each series is associated with a country or bidding zone, and, where applicable, an energy type. Unit-level generation is available through Actual-ByUnit. Logical grouping by lifecycle segment and subtype supports consistent joins across modalities.

| Lifecycle segment | Representative subtypes | Reported resolution |
|---|---|---|
| Generation | Actual-ByType, Actual-ByUnit, Renewables-Forecast, Capacity-Annual | 15-min, hourly, yearly |
| Load | Actual, Forecast-WeekAhead | 15-min, daily |
| Market | Price-QuarterHourly, Price-Hourly | 15-min, hourly |
| Transmission | Capacity-Forecast, Flow-Actual | hourly |
| Balancing | Energy-Activated, System-Imbalance | 15-min, hourly |

The reported sequence lengths illustrate the scale of the release. Actual-ByType, Renewables-Forecast, Load/Actual, Price-QuarterHourly, and Energy-Activated each have length \(>330k\), corresponding to roughly 9.5 years at 15-minute resolution. Capacity-Forecast, Flow-Actual, Price-Hourly, and System-Imbalance are reported at length \(\sim 8.7k\), while Capacity-Annual has length \(\sim 10\), and Load/Forecast-WeekAhead has length \(\sim 3.4k\).

Real-E also includes substantial contextual metadata. Spatial descriptors include plant coordinates or bidding zones. System-level attributes include transmission distance, voltage level, and grid topology. Energy-type descriptors include generation category or type and annual capacity. These metadata support geographic clustering, cross-zone analysis, and operationally informed forecasting. The paper notes that weather or other exogenous variables are not enumerated in detail, although Renewables-Forecast series are included.

At the schema level, common fields include timestamps aligned to 15-minute or hourly cadences, values following ENTSO-E definitions, and identifiers such as country code, bidding zone, unit identifier, and energy type. Measurements for generation, load, flow, and capacity are typically in MW or MWh, while prices are in EUR/MWh. Full directory layouts and schemas are provided in the benchmark repository.

## 3. Forecasting tasks and experimental protocol

Real-E defines fixed-horizon multi-step forecasting tasks with a standard temporal split of 70% training, 20% validation, and 10% test across all benchmarked models. The forecasting horizon is 12 time steps. In the 15-minute setting this corresponds to 3 hours; in the hourly setting it corresponds to 12 hours. The benchmark explores rolling input windows of size \(12 \times (1\text{–}3)\), while the baseline suite commonly uses 12-step input windows matched to the 12-step horizon [2509.05768].

The benchmark emphasizes multivariate forecasting across energy categories, such as generation by type, and supports multi-series training where appropriate. It also supports cross-country and cross-station analyses. The paper’s own experiments focus especially on Germany and France as representative cases and on cross-dataset generalization between classic benchmarks and Real-E.

Preprocessing is guided by data quality and temporal consistency. Time series with substantial missing values were excluded in order to maximize coverage and alignment across energy types and countries. Native ENTSO-E cadences are preserved, and any harmonization is handled within the repository’s loaders. Residual gaps, if present, are treated per modality by those loaders. The implementation also addresses alignment and split strategies so as to avoid DST-induced sampling errors. The paper does not describe rolling-origin evaluation, although it states that the codebase supports standard time-series evaluation practices.

The benchmark repository provides Python loaders to access, align, and split series consistently across modalities and countries. This infrastructure is a substantive component of the benchmark because it standardizes task construction across heterogeneous resolutions and lifecycle segments.

## 4. Correlation-shift analysis

A distinctive contribution of Real-E is its explicit treatment of time-evolving correlation structure. To quantify correlation shifts, the benchmark introduces Temporal Graph Volatility (TGV) and Graph Spectral Divergence (GSD), both computed from correlation-derived graphs constructed over sliding windows [2509.05768].

TGV measures structural variation in adjacency matrices across adjacent windows:

$$
\mathrm{TGV} = \sum_{t=0}^{T-1} \left\| A_{t+1} - A_t \right\|_F
= \sum_{t=1}^{T} \sqrt{ \sum_{i=1}^{m} \sum_{j=1}^{n} \left| a^{t+1}_{ij} - a^{t}_{ij} \right|^2 }.
$$

Here, \(A_t \in \mathbb{R}^{m \times n}\) is an adjacency matrix derived from inter-series correlations, such as Pearson correlations computed in sliding windows, and \(a^t_{ij}\) denotes its entries.

GSD measures spectral topology shift using Laplacian eigenvalues:

$$
\mathrm{GSD} = \left\| \lambda(L_t) - \lambda(L_{t-1}) \right\|_2, \quad L_t = D_t - A_t,
$$

where \(D_t = \mathrm{diag}(A_t \mathbf{1})\) is the degree matrix, \(L_t\) is the graph Laplacian, and \(\lambda(\cdot)\) is the vector of sorted Laplacian eigenvalues.

The reported values indicate substantially greater correlation volatility in Real-E than in classic datasets. Existing datasets average approximately \(0.5822\) in TGV, with maximum approximately \(0.6245\), whereas Real-E averages approximately \(1.03335\), with maximum approximately \(1.0969\). For GSD, existing datasets average approximately \(1.6298\), with maximum approximately \(1.8027\), whereas Real-E averages approximately \(2.3237\), with maximum approximately \(2.7289\).

The empirical analysis links these shifts to concrete energy-system phenomena. In Germany, solar generation peaks in summer while brown coal peaks in autumn and winter, and system demand exhibits bimodal annual patterns supported by different sources. The paper also reports evolving blocks of strong association among biomass, brown coal, and gas over certain periods. This suggests that the benchmark is not merely large; it is structured to reveal dynamic dependency regimes that challenge models relying on static global relationships.

## 5. Baseline suite and benchmark findings

The benchmark evaluates over 20 baselines spanning statistical, neural, Transformer, and graph-based families. Statistical baselines include VAR, ARIMA, and Seasonal ARIMA. MLP-based models include DLinear, N-Beats, and TimeMixer. RNN/CNN-based models include LSTNet, TCN, DeepGLO, and SFM. Transformer-based models include Informer, Autoformer, FEDformer, and Reformer. Spectral GNN baselines include StemGNN, FourierGNN, and TGGC. Spatial GNN baselines include GWaveNet, MTGNN, and TPGNN [2509.05768].

Hyperparameters are tuned over common global ranges: learning rates from \(10^{-2}\) to \(10^{-4}\), hidden embedding sizes from \(2^8\) to \(2^{10}\), batch sizes from \(2^4\) to \(2^9\), encoder or decoder layers from 1 to 3, attention heads from 2 to 8, convolutional kernel sizes from 3 to 7, and rolling window sizes of \(12 \times (1\text{ to }3)\). All results are reported with early stopping and averaged over five random seeds.

On the Germany and France generation-by-type subsets, spatial GNNs are reported to outperform other model families under dynamic correlation structures. The paper gives the following representative 12-step results:

| Subset | Model(s) highlighted | Reported MAE/RMSE |
|---|---|---|
| Germany | GWaveNet | 0.013 / 0.028 |
| Germany | MTGNN | 0.016 / 0.034 |
| Germany | Autoformer | 0.204 / 0.376 |
| France | MTGNN | 0.012 / 0.023 |
| France | GWaveNet | 0.012 / 0.025 |

The broader generalization analysis is equally central. Transformer-based models show an average MAE increase from 0.117 on classic datasets to 0.215 on Real-E, which the paper reports as an 85.4% rise. Spectral GNNs degrade by approximately 16.20% relative to classic benchmarks. The stated interpretation is that models relying on global attention or static spectral structure struggle when inter-series correlations shift rapidly, whereas spatial GNNs that dynamically adapt graph structure remain more robust.

The paper argues that this performance gap is less visible on smaller, single-country datasets such as Electricity, Solar, and ETT, but becomes pronounced on Real-E. A plausible implication is that Real-E functions as a stress test for correlation-aware modeling rather than merely another accuracy leaderboard.

## 6. Accessibility, reproducibility, and limitations

Real-E is openly released. The dataset is hosted on Zenodo at `https://zenodo.org/records/15685930`, and the benchmark code is hosted at `https://github.com/YueW26/Real-E`. Both are released under the Creative Commons Attribution 4.0 license. The recommended bibliographic reference is the CIKM ’25 paper, with DOI `https://doi.org/10.1145/3746252.3761637` [2509.05768].

The benchmark is reproducible by design. The repository provides scripts for data loading, splitting, training, and evaluation, together with hyperparameter ranges and configurations. The paper also documents practical guidance for users: prefer series with low missingness, maintain strict temporal ordering, avoid cross-window leakage in multivariate settings, use multiple seeds with early stopping, and match horizon and resolution to the operational context. For highly dynamic multi-energy settings, the authors recommend dynamic spatial GNNs or hybrid models that can re-estimate adjacency from recent windows. They also recommend monitoring TGV and GSD for regime changes and considering adaptive retraining or graph updates when these metrics exceed thresholds.

Several limitations are explicit. Details of weather or other exogenous variables are not enumerated in the paper. The experiments focus on selected settings—especially Germany and France subsets and comparisons against classic benchmarks—although the dataset structure supports broader cross-country and transfer evaluations. More substantively, the benchmark exposes limitations of existing model families rather than resolving them: Transformer variants and some Spectral GNNs struggle under pronounced correlation dynamics and regime changes, while single-country and single-energy benchmarks may understate real-world complexity.

In this sense, Real-E serves both as a dataset and as an argument about evaluation methodology. It proposes that robust electricity forecasting should be assessed under long-horizon temporal coverage, lifecycle coupling, rich operational metadata, and explicitly measured correlation drift, rather than only on narrowly scoped stationary benchmarks.

Source: https://www.emergentmind.com/topics/real-e-dataset