Data Science
Energy load forecasting with Darts
Rolling-window backtesting pipelines for building energy load forecasting: LightGBM, NHiTS, TSMixer and a robust seasonal baseline.
Python Darts LightGBM NHiTS TSMixer pandas MLflow
What it’s about
As part of my work at Fraunhofer ISE I built load-forecasting pipelines for buildings. The project shows how to use Darts to reproducibly evaluate different model classes against a strong seasonal baseline (GlobalNaiveSeasonal).
Architecture
- Rolling-window backtesting: each model is evaluated across multiple forecast horizons and windows
- Models: LightGBM (gradient boosting), NHiTS (Neural Hierarchical Interpolation), TSMixer, against
GlobalNaiveSeasonalas a non-negotiable baseline - Pipeline: deterministic preprocessing, feature engineering (lag features, calendar features), hyperparameter tuning
- MAE / RMSE / sMAPE per horizon in a unified results table
What I learned
The seasonal baseline is surprisingly strong on energy loads. Without it, you’ll miss how much of a complex model’s “performance” actually comes from the weekly cycle. In forecasting, reproducibility (same seed, same splits, same metric) matters more than a fancy model.