Electricity Load Forecasting

Machine-learning based 24-hour ahead German electricity load forecasting using historical ENTSO-E load data, calendar features, weather variables and LightGBM.

The project was developed for the DDMO SoSe 2026 electricity-load forecasting challenge and focuses on a robust, leakage-aware forecasting workflow with recent backtesting and benchmark comparison.

ENTSO-E Actual Load
 +
Calendar Features
 +
Weather Features
 ↓
Data Cleaning / Outlier Handling
 ↓
Lagged Load Features
 ↓
LightGBM Recursive Forecaster
 ↓
24-Hour German Load Forecast
 ↓
MAE / RMSE / MAPE Evaluation
 ↓
Persistence Benchmark Comparison

The final live-style backtest used a 7-day recent evaluation window.

For an additional forecast-vs-actual test day, the model achieved:

• MAE: 895.3 MW
• RMSE: 1,117.6 MW
• MAPE: 1.70%

These metrics are from the project's recorded June 2026 backtesting workflow. They should be interpreted as a recent validation result rather than a guarantee of future forecasting performance.

The main forecasting model uses:

• LightGBM
• Recursive multi-step forecasting
• 168-hour weekly lag
• Calendar features
• Weather features
• Explicit outlier annotation / weighting
• Deterministic training configuration

Core estimator configuration:

n_estimators = 400
learning_rate = 0.05
num_leaves = 63
min_child_samples = 20
random_state = 2026
deterministic = True
force_col_wise = True

Historical German electricity load is obtained from ENTSO-E Actual Load data.

The pipeline converts the source series to hourly resolution and trains only on actually published historical values.

The forecasting workflow uses time-based information such as:

• hour of day
• day of week
• month / seasonal information
• weekday/weekend structure
• other calendar-derived features used by the forecasting pipeline

Weather variables are incorporated as exogenous predictors for the German load forecast.

A 168-hour weekly lag captures the strong weekly structure in electricity demand. Additional historical information is handled through the forecasting framework and feature pipeline.

The project contains explicit checks for:

• ENTSO-E data freshness
• frame coverage
• missing hourly values
• training/prediction feature completeness
• stale Actual Load data
• forecast horizon length
• NaN predictions

Outlier rows are annotated and handled through a weighting mechanism rather than silently replacing the underlying historical observations.

The evaluation is designed to avoid using future information.

For each backtest day:

1. Build the training set only from information available before the target day.
2. Prepare calendar and weather exogenous variables.
3. Fit the LightGBM recursive forecaster.
4. Predict the next 24 hours.
5. Compare predictions with actual ENTSO-E load.
6. Calculate MAE, RMSE and MAPE.
7. Compare against weekly persistence.

The recent backtest also records forecast bias and per-day performance.

electricity-load-forecasting/
│
├── README.md
├── LICENSE
├── SECURITY.md
├── requirements.txt
├── pyproject.toml
├── .gitignore
│
├── notebooks/
│ └── live_preprocessing_v2_original_restored.ipynb
│
├── src/
│ ├── forecast_pipeline.py
│ ├── metrics.py
│ ├── preprocessing.py
│ └── ...
│
├── data/
│ └── README.md
│
├── submissions/
│ └── neura/
│ └── forecast CSV submissions
│
├── docs/
│ └── project documentation
│
└── .github/
 └── workflows/
 └── scorecard.yml

Python 3.11+ is recommended.

python -m venv .venv

.venv\Scripts\Activate.ps1
pip install -r requirements.txt

source .venv/bin/activate
pip install -r requirements.txt

The live notebook reads the ENTSO-E API key from an environment variable:

ENTSOE_API_KEY

Do not commit API keys or api.env files.

Example Windows PowerShell configuration:

$env:ENTSOE_API_KEY="YOUR_API_KEY"

The notebook checks whether the variable exists before attempting a live download.

Start Jupyter:

jupyter notebook

Open:

notebooks/live_preprocessing_v2_original_restored.ipynb

The notebook contains the end-to-end live-style workflow, including:

• data acquisition
• preprocessing
• feature construction
• model training
• backtesting
• forecast-vs-actual validation
• persistence comparison
• provenance logging

The submissions/neura/ directory contains forecast CSV files generated for the challenge workflow.

Raw training data, cached datasets and trained model artifacts are intentionally not committed.

The repository includes SECURITY.md and automated security-analysis configuration.

Never commit:

• ENTSO-E API keys
• .env files
• api.env
• downloaded raw datasets containing credentials or private information
• local cache directories
• trained model binaries unless explicitly intended for publication

• The recent validation window is relatively short.
• Electricity demand is affected by weather, holidays, market conditions and unusual events that are difficult to model perfectly.
• Backtest performance does not guarantee leaderboard or future live performance.
• The public repository intentionally excludes the large raw datasets and local cache used during development.

Python · LightGBM · Pandas · NumPy · Scikit-learn · Jupyter · ENTSO-E · Time-Series Forecasting · Feature Engineering · Weather Data · Machine Learning

Team Neura — DDMO SoSe 2026

Project focused on data-driven electricity-load forecasting and machine-learning based forecasting evaluation.