Real-E: The Largest Real-World Multivariate Electricity Forecasting Benchmark

Real-E is a comprehensive, high-resolution benchmark dataset for multivariate time series forecasting in energy systems. It spans 10 years across 39 European countries, including over 74 electricity stations and 20+ energy categories. With rich metadata, non-stationary dynamics, and high temporal granularity, Real-E provides a rigorous foundation for developing and evaluating robust forecasting models.

Dataset

We employ two key metrics—\textbf{overlapping rate} and \textbf{valid percentage}—to assess temporal alignment and data completeness across features. Guided by these metrics, we provide three dataset versions: the \textbf{OLD} version retains raw data with minimal filtering for robustness testing; the \textbf{V60} version ensures that each time interval has at least 60% valid data, offering improved reliability without overly restricting coverage; the \textbf{O20} version filters out variables with an overlapping rate below 20%, resulting in a temporally consistent and well-aligned subset. We also provide variable-level visualization plots for each sub-dataset to support intuitive exploration and quality assessment. This multi-version design supports a wide range of forecasting and evaluation scenarios.

🧱 Dataset Overview

Category	Name	Duration	Resolution	Length	EC	Countries
Generation	Actual-ByType	9.5 y	15 min	>330k	20	39
	Actual-ByUnit	9.5 y	1 hour	~8.7k	20	39
	Renewables-Forecast	9.5 y	15 min	>330k	3	39
	Capacity-Annual	9.5 y	1 year	~10	20	39
Load	Actual	9.5 y	15 min	>330k	20	39
	Forecast-WeekAhead	9.5 y	1 day	~3.4k	20	39
Market	Price-QuarterHourly	9.5 y	15 min	>330k	20	39
	Price-Hourly	9.5 y	1 hour	~8.7k	20	39
Transmission	Capacity-Forecast	9.5 y	1 hour	~8.7k	20	39
	Flow-Actual	9.5 y	1 hour	~8.7k	20	39
Balancing	Energy-Activated	9.5 y	15 min	>330k	20	39
	System-Imbalance	9.5 y	1 hour	~8.7k	--	39

Preprocessing

We provide three versions, which keep 100% original time series, less than 60% and 20% missing value respectively.

1. Original: Retains the raw data.
2. V60: Ensures that each time interval has at least 60% valid data.

Dataset versions are organized into the following subdirectories:

Spatial Dimensions

Real-E supports three spatial aggregation levels, offering flexible granularity for various forecasting and analysis tasks:

• BZN – Bidding Zone Level

  • Represents electricity market regions where prices are uniform.
  • Closely tied to market operations and congestion management.
  • Useful for forecasting market-related variables such as generation, consumption, and cross-border flows.
  • Example: BZN|DE-LU (Germany–Luxembourg bidding zone), BZN|FR (France).
  • A single country may have multiple BZNs, or several countries may share one (e.g., DE-LU).

• CTA – Control Area Level

  • Reflects the operational boundaries of Transmission System Operators (TSOs).
  • Ideal for studying grid stability, dispatching, and load balancing.
  • Example: CTA|50Hertz, CTA|Amprion, CTA|RTE.
  • Suitable for tasks involving transmission, balancing, or grid-centric modeling.

• CTY – Country Level

  • National-level aggregation of energy data.
  • Simplifies modeling while maintaining sufficient realism for macro-level analysis.
  • Example: CTY|DE (Germany), CTY|FR (France).
  • Recommended for baseline models, policy simulations, or when comparing nations.

Visualization

• Statistic: Visual statistical analysis Each version includes variable-level visualization plots to support intuitive exploration and quality assessment.

This multi-version design supports a wide range of forecasting and evaluation scenarios.

We also publish the preprocessed dataset in Zenodo: Real-E (OLD O20 V60)

Preprocessing

In this repository we leverage the preprocessed datasets.

🧪 Benchmarking Overview

We benchmark 20 models from different families: We evaluate a comprehensive set of models, including classical methods (ARIMA, S-ARIMA, VAR), MLP-based approaches (DLinear, N-Beats, TimeMixer), RNN/CNN architectures (LSTM, TCN, DeepGLO, SFM), Transformer-based models (Informer, Autoformer, FEDformer, Reformer), and Graph Neural Networks—both spectral (FourierGNN, LSGCN, StemGNN) and spatial (MTGNN, TPGNN, WaveNet).

Results Summary

Till Juni 2025, the top three models for five different tasks are:

We conduct multivariate time series forecasting on the Real-E dataset with a fixed prediction horizon of 12 time steps.

• Top-performing class: spatial-based models outperform other categories on average.

Rank	Model
🥇 1st	[GWaveNet]
🥈 2nd	[MTGNN]
🥉 3rd	[TPGNN]

These models consistently achieved the lowest forecasting errors (MAE & RMSE) across multiple Real-E subsets, demonstrating strong generalization on large-scale, high-variance electricity data.

🚀 Quick Start

1. Clone the Repository

git clone https://github.com/YueW26/Real-E.git

2. Install Dependencies

pip install -r requirements.txt

---

⚙️ Training Examples

🧭 Spatial Graph Neural Network (GNN)

python train.py --data data/FRANCE --gcn_bool --adjtype doubletransition --addaptadj --randomadj --epochs 50

• --gcn_bool: Use GCN layers
• --adjtype: Type of adjacency matrix ("doubletransition" recommended)
• --addaptadj: Learnable adjacency
• --randomadj: Random init of graph structure

---

🔁 Reformer (Transformer-based)

python /EnergyTSF/run.py \
  --task_name long_term_forecast \
  --is_training 1 \
  --model_id Reformer_test \
  --model Reformer \
  --data Opennem \
  --data_path Germany_processed_0.csv \
  --features M \
  --seq_len 12 --label_len 12 --pred_len 12 \
  --enc_in 16 --dec_in 16 --c_out 16 \
  --des 'debug_run' --itr 1

• seq_len, label_len, pred_len: Input-output time window
• enc_in, dec_in, c_out: Number of input/output features
• itr: Repeat experiment for robustness

---

Contact

For questions, suggestions, or contributions, please feel free to open an issue on GitHub or contact the maintainer via email at: Joellawang2013@gmail.com