FAQ#

This page answers common questions from new users about OpenSTEF — what it does, how to set it up, and how to get the most out of it. Click any question to expand the answer.

General#

What is OpenSTEF?

OpenSTEF (Open Short-Term Energy Forecasting) is an open-source Python library that provides a complete machine learning framework for short-term energy load forecasting. It covers the full pipeline: data preprocessing, feature engineering, model training, probabilistic forecasting, evaluation, and post-processing.

OpenSTEF is not a single model — it is a model-agnostic framework that lets you plug in different ML algorithms while handling all the domain-specific logic for energy forecasting.

What does “short-term” mean in this context?

Short-term means predicting energy load hours to approximately 7 days ahead. The practical limit is around 7 days because:

  • Weather forecasts beyond 7 days lack the 15-minute resolution needed for accurate predictions

  • Solar and wind peaks become unpredictable (cloudy vs sunny days)

  • Forecast quality degrades significantly beyond this horizon

What makes OpenSTEF different from a general ML library?

OpenSTEF includes domain knowledge specific to energy forecasting that you would otherwise need to build yourself:

  • Energy-specific feature engineering — e.g., converting solar radiation into PV generation estimates

  • Probabilistic forecasts — generates prediction intervals (uncertainty bandwidths), not just point predictions

  • Complete pipelines — handles the full workflow from raw data to deployable forecasts

  • Built-in evaluation — metrics and benchmarking tools designed for energy forecasting use cases

A general ML library gives you models; OpenSTEF gives you a forecasting system.

What are the primary use cases?

OpenSTEF was originally developed at Alliander (a Dutch grid operator) for congestion management — forecasting load at grid points to identify when equipment limits will be exceeded. Other use cases include:

  • Transport capacity forecasting

  • EV charging capacity estimation

  • Grid loss prediction

  • Any scenario where you need accurate short-term load predictions

Installation & Setup#

What are the system requirements?

  • Python: >=3.12, <4.0

  • OS: Linux, macOS, or Windows

Install the complete framework:

pip install openstef

This meta-package installs all sub-packages: openstef-beam, openstef-core, openstef-meta, and openstef-models.

Can I install only the parts I need?

Yes. OpenSTEF is modular — install individual packages for specific functionality:

# Core data processing and pipelines
pip install openstef-core

# Models (XGBoost, LightGBM, etc.)
pip install openstef-models

# Backtesting, Evaluation, Analysis and Metrics
pip install openstef-beam

# Meta-learning models
pip install openstef-meta

Each package also has optional extras. For example:

# Install models with LightGBM support
pip install openstef-models[lgbm]

# Install models with hyperparameter tuning
pip install openstef-models[tuning]

# Install beam with baseline models
pip install openstef-beam[baselines]
What are the key dependencies?

The core dependencies include:

  • numpy, pandas — data manipulation

  • pydantic — configuration and validation

  • pyarrow — efficient data serialization

  • joblib — parallel processing

Model-specific dependencies:

  • xgboost — XGBoost gradient boosting (via openstef-models[xgb-cpu] or openstef-models[xgb-gpu])

  • lightgbm — LightGBM gradient boosting (via openstef-models[lgbm])

  • optuna — hyperparameter tuning (via openstef-models[tuning])

  • mlflow-skinny — experiment tracking

  • pvlib — solar energy calculations

Models & Forecasting#

Which ML models are supported?

OpenSTEF currently supports:

  • XGBoost — gradient boosting trees, the default choice for most use cases

  • LightGBM — an alternative gradient boosting implementation

Both models support probabilistic forecasting via quantile regression. The framework is model-agnostic, so additional model types (including foundation models) are being developed.

See user_guides/models for details on model configuration.

What input data does OpenSTEF need?

At minimum, you need:

  • Load measurements — historical energy consumption time series

  • Weather forecasts — temperature, wind speed, solar radiation, humidity, pressure

Optional but beneficial:

  • Electricity market prices (e.g., EPEX day-ahead prices)

  • Load profiles — typical daily/weekly consumption patterns

Here’s how to configure weather columns in a workflow:

from openstef_models.presets import ForecastingWorkflowConfig

config = ForecastingWorkflowConfig(
    model_id="my_forecast",
    model="xgboost",
    target_column="load",
    temperature_column="temperature_2m",
    relative_humidity_column="relative_humidity_2m",
    wind_speed_column="wind_speed_10m",
    radiation_column="shortwave_radiation",
    pressure_column="surface_pressure",
)
What are probabilistic forecasts and why do they matter?

Instead of predicting a single value (“load will be 500 kW”), OpenSTEF produces prediction intervals — e.g., “load will be between 450 and 550 kW with 80% confidence.”

This is critical for grid operations because decisions (like curtailing customers) have real costs. Knowing the uncertainty helps operators make better risk-adjusted decisions.

You configure quantiles when setting up a workflow:

from openstef_core.types import LeadTime, Q

config = ForecastingWorkflowConfig(
    # ...
    horizons=[LeadTime.from_string("PT36H")],
    quantiles=[Q(0.5), Q(0.1), Q(0.9)],  # Median + 80% prediction interval
)
Can I tune hyperparameters automatically?

Yes. OpenSTEF integrates with Optuna for hyperparameter tuning. You specify parameter ranges and the tuner searches for optimal values:

from openstef_core.mixins.param_ranges import FloatRange, IntRange
from openstef_models.models.forecasting.xgboost_forecaster import XGBoostHyperParams

hyperparams = XGBoostHyperParams(
    learning_rate=FloatRange(0.01, 0.3, log=True, tune=True),
    n_estimators=IntRange(50, 500, tune=True),
    max_depth=IntRange(3, 10, tune=True),
)

Install the tuning extra to enable this: pip install openstef-models[tuning].

See user_guides/hyperparameter_tuning for a complete guide.

Getting Started#

Where should I start as a new user?

Follow this path:

  1. Install OpenSTEF: pip install openstef

  2. Work through the tutorials/index — start with the introductory tutorial

  3. Try the built-in benchmark dataset to see a full workflow in action

The quickest way to see OpenSTEF in action:

from openstef_core.testing import load_liander_dataset

dataset = load_liander_dataset()
print(f"Dataset shape: {dataset.data.shape}")
print(f"Date range: {dataset.data.index.min()} to {dataset.data.index.max()}")
Is there a benchmark dataset I can use to test things?

Yes. The Liander 2024 Energy Forecasting Benchmark dataset is available from HuggingFace Hub and can be loaded directly:

from openstef_core.testing import load_liander_dataset, prepare_tutorial_datasets

dataset = load_liander_dataset()

This dataset contains load measurements from MV feeders and transformers, versioned weather forecasts, EPEX electricity prices, and typical load profiles. Install the benchmark extra for access: pip install openstef-core[benchmark].

How do I evaluate forecast accuracy?

Accuracy depends on your use case — congestion management cares about peak detection, while other applications may focus on overall error metrics like RMSE.

OpenSTEF provides evaluation tools in the openstef-beam package:

from openstef_beam.evaluation.metric_providers import (
    ObservedProbabilityProvider,
    RMAEProvider,
)

The recommendation is to run the Alliander benchmark with openstef-beam to see performance metrics and plots for your specific setup.

See user_guides/evaluation for details on available metrics.